COMPILED BY TANUJIT CHAKRABORTY

Class Notes
CS229 Course
Machine Learning
Standford University

Topics Covered:

1. Supervised Learning: Linear Regression & Logistic Regression

2. Generative Learning algorithms & Discriminant Analysis
3. Kernel Methods and SVM
4. Basics of Statistical Learning Theory
5. Regularization and model selection
6. Backpropagation & Deep learning
7. Unsupervised Learning & k-means clustering
8. Mixtures of Gaussians
9. EM algorithm
10. Factor analysis
11. Principal Components Analysis
12. Independent Components Analysis
13. Reinforcement Learning
14. Boosting algorithms and weak learning

Reference:
[1] http://cs229.stanford.edu/syllabus.html
[2] https://see.stanford.edu/Course/CS229

Mail: tanujitisi@gmail.com Webpage: www.ctanujit.org

 COMPILED BY TANUJIT CHAKRABORTY

Mail: tanujitisi@gmail.com Webpage: www.ctanujit.org

 CS229 Lecture Notes
Andrew Ng
slightly updated by TM on April 3, 2019

Supervised learning
Let’s start by talking about a few examples of supervised learning problems.
Suppose we have a dataset giving the living areas and prices of 47 houses
from Portland, Oregon:
Living area (feet2 ) Price (1000$s)
2104 400
1600 330
2400 369
1416 232
3000 540
.. ..
. .
We can plot this data:
housing prices

1000

900

800

700

600
price (in $1000)

500

400

300

200

100

500 1000 1500 2000 2500 3000 3500 4000 4500 5000
square feet

1
CS229 Fall 2018 2

Given data like this, how can we learn to predict the prices of other houses
in Portland, as a function of the size of their living areas?
To establish notation for future use, we’ll use x(i) to denote the “input”
variables (living area in this example), also called input features, and y (i)
to denote the “output” or target variable that we are trying to predict
(price). A pair (x(i) , y (i) ) is called a training example, and the dataset
that we’ll be using to learn—a list of n training examples {(x(i) , y (i) ); i =
1, . . . , n}—is called a training set. Note that the superscript “(i)” in the
notation is simply an index into the training set, and has nothing to do with
exponentiation. We will also use X denote the space of input values, and Y
the space of output values. In this example, X = Y = R.
To describe the supervised learning problem slightly more formally, our
goal is, given a training set, to learn a function h : X 7→ Y so that h(x) is a
“good” predictor for the corresponding value of y. For historical reasons, this
function h is called a hypothesis. Seen pictorially, the process is therefore
like this:

Training
set

Learning
algorithm

x h predicted y
(living area of (predicted price)
house.) of house)

When the target variable that we’re trying to predict is continuous, such
as in our housing example, we call the learning problem a regression prob-
lem. When y can take on only a small number of discrete values (such as
if, given the living area, we wanted to predict if a dwelling is a house or an
apartment, say), we call it a classification problem.
 3

Part I
Linear Regression
To make our housing example more interesting, let’s consider a slightly richer
dataset in which we also know the number of bedrooms in each house:
Living area (feet2 ) #bedrooms Price (1000$s)
2104 3 400
1600 3 330
2400 3 369
1416 2 232
3000 4 540
.. .. ..
. . .

(i)
Here, the x’s are two-dimensional vectors in R2 . For instance, x1 is the
(i)
living area of the i-th house in the training set, and x2 is its number of
bedrooms. (In general, when designing a learning problem, it will be up to
you to decide what features to choose, so if you are out in Portland gathering
housing data, you might also decide to include other features such as whether
each house has a fireplace, the number of bathrooms, and so on. We’ll say
more about feature selection later, but for now let’s take the features as
given.)
To perform supervised learning, we must decide how we’re going to rep-
resent functions/hypotheses h in a computer. As an initial choice, let’s say
we decide to approximate y as a linear function of x:

hθ (x) = θ0 + θ1 x1 + θ2 x2

Here, the θi ’s are the parameters (also called weights) parameterizing the
space of linear functions mapping from X to Y. When there is no risk of
confusion, we will drop the θ subscript in hθ (x), and write it more simply as
h(x). To simplify our notation, we also introduce the convention of letting
x0 = 1 (this is the intercept term), so that
d
X
h(x) = θi xi = θT x,
i=0

where on the right-hand side above we are viewing θ and x both as vectors,
and here d is the number of input variables (not counting x0 ).
 4

Now, given a training set, how do we pick, or learn, the parameters θ?

One reasonable method seems to be to make h(x) close to y, at least for
the training examples we have. To formalize this, we will define a function
that measures, for each value of the θ’s, how close the h(x(i) )’s are to the
corresponding y (i) ’s. We define the cost function:
n
1X
J(θ) = (hθ (x(i) ) − y (i) )2 .
2 i=1
If you’ve seen linear regression before, you may recognize this as the familiar
least-squares cost function that gives rise to the ordinary least squares
regression model. Whether or not you have seen it previously, let’s keep
going, and we’ll eventually show this to be a special case of a much broader
family of algorithms.

1 LMS algorithm
We want to choose θ so as to minimize J(θ). To do so, let’s use a search
algorithm that starts with some “initial guess” for θ, and that repeatedly
changes θ to make J(θ) smaller, until hopefully we converge to a value of
θ that minimizes J(θ). Specifically, let’s consider the gradient descent
algorithm, which starts with some initial θ, and repeatedly performs the
update:
∂
θj := θj − α J(θ).
∂θj
(This update is simultaneously performed for all values of j = 0, . . . , d.)
Here, α is called the learning rate. This is a very natural algorithm that
repeatedly takes a step in the direction of steepest decrease of J.
In order to implement this algorithm, we have to work out what is the
partial derivative term on the right hand side. Let’s first work it out for the
case of if we have only one training example (x, y), so that we can neglect
the sum in the definition of J. We have:
∂ ∂ 1
J(θ) = (hθ (x) − y)2
∂θj ∂θj 2
1 ∂
= 2 · (hθ (x) − y) · (hθ (x) − y)
2 ∂θj
d
!
∂ X
= (hθ (x) − y) · θi x i − y
∂θj i=0
= (hθ (x) − y) xj
 5

For a single training example, this gives the update rule:1

(i)
θj := θj + α y (i) − hθ (x(i) ) xj .

The rule is called the LMS update rule (LMS stands for “least mean squares”),
and is also known as the Widrow-Hoff learning rule. This rule has several
properties that seem natural and intuitive. For instance, the magnitude of
the update is proportional to the error term (y (i) − hθ (x(i) )); thus, for in-
stance, if we are encountering a training example on which our prediction
nearly matches the actual value of y (i) , then we find that there is little need
to change the parameters; in contrast, a larger change to the parameters will
be made if our prediction hθ (x(i) ) has a large error (i.e., if it is very far from
y (i) ).
We’d derived the LMS rule for when there was only a single training
example. There are two ways to modify this method for a training set of
more than one example. The first is replace it with the following algorithm:

Repeat until convergence {

(i)
θj := θj + α ni=1 y (i) − hθ (x(i) ) xj
P
(for every j).

The reader can easily verify that the quantity in the summation in the update
rule above is just ∂J(θ)/∂θj (for the original definition of J). So, this is
simply gradient descent on the original cost function J. This method looks
at every example in the entire training set on every step, and is called batch
gradient descent. Note that, while gradient descent can be susceptible
to local minima in general, the optimization problem we have posed here
for linear regression has only one global, and no other local, optima; thus
gradient descent always converges (assuming the learning rate α is not too
large) to the global minimum. Indeed, J is a convex quadratic function.
Here is an example of gradient descent as it is run to minimize a quadratic
function.
1
We use the notation “a := b” to denote an operation (in a computer program) in
which we set the value of a variable a to be equal to the value of b. In other words, this
operation overwrites a with the value of b. In contrast, we will write “a = b” when we are
asserting a statement of fact, that the value of a is equal to the value of b.
 6

50

45

40

35

30

25

20

15

10

5 10 15 20 25 30 35 40 45 50

The ellipses shown above are the contours of a quadratic function. Also
shown is the trajectory taken by gradient descent, which was initialized at
(48,30). The x’s in the figure (joined by straight lines) mark the successive
values of θ that gradient descent went through.
When we run batch gradient descent to fit θ on our previous dataset,
to learn to predict housing price as a function of living area, we obtain
θ0 = 71.27, θ1 = 0.1345. If we plot hθ (x) as a function of x (area), along
with the training data, we obtain the following figure:
housing prices

1000

900

800

700

600
price (in $1000)

500

400

300

200

100

500 1000 1500 2000 2500 3000 3500 4000 4500 5000
square feet

If the number of bedrooms were included as one of the input features as well,
we get θ0 = 89.60, θ1 = 0.1392, θ2 = −8.738.
The above results were obtained with batch gradient descent. There is
an alternative to batch gradient descent that also works very well. Consider
the following algorithm:
 7

Loop {

for i = 1 to n, {

(i)
θj := θj + α y (i) − hθ (x(i) ) xj (for every j).
}

In this algorithm, we repeatedly run through the training set, and each time
we encounter a training example, we update the parameters according to
the gradient of the error with respect to that single training example only.
This algorithm is called stochastic gradient descent (also incremental
gradient descent). Whereas batch gradient descent has to scan through
the entire training set before taking a single step—a costly operation if n is
large—stochastic gradient descent can start making progress right away, and
continues to make progress with each example it looks at. Often, stochastic
gradient descent gets θ “close” to the minimum much faster than batch gra-
dient descent. (Note however that it may never “converge” to the minimum,
and the parameters θ will keep oscillating around the minimum of J(θ); but
in practice most of the values near the minimum will be reasonably good
approximations to the true minimum.2 ) For these reasons, particularly when
the training set is large, stochastic gradient descent is often preferred over
batch gradient descent.

2 The normal equations

Gradient descent gives one way of minimizing J. Let’s discuss a second way
of doing so, this time performing the minimization explicitly and without
resorting to an iterative algorithm. In this method, we will minimize J by
explicitly taking its derivatives with respect to the θj ’s, and setting them to
zero. To enable us to do this without having to write reams of algebra and
pages full of matrices of derivatives, let’s introduce some notation for doing
calculus with matrices.
2
By slowly letting the learning rate α decrease to zero as the algorithm runs, it is also
possible to ensure that the parameters will converge to the global minimum rather than
merely oscillate around the minimum.
 8

2.1 Matrix derivatives

For a function f : Rn×d 7→ R mapping from n-by-d matrices to the real
numbers, we define the derivative of f with respect to A to be:
 ∂f ∂f

∂A11
· · · ∂A 1d

∇A f (A) =  ... ... .. 


. 
∂f ∂f
∂An1
··· ∂And

Thus, the gradient ∇A f (A) is itself an n-by-d matrix,

 whose (i, j)-element is
A11 A12
∂f /∂Aij . For example, suppose A = is a 2-by-2 matrix, and
A21 A22
the function f : R2×2 7→ R is given by
3
f (A) = A11 + 5A212 + A21 A22 .
2
Here, Aij denotes the (i, j) entry of the matrix A. We then have
 3 
10A12
∇A f (A) = 2 .
A22 A21
We also introduce the trace operator, written “tr.” For an d-by-d (square)
matrix A, the trace of A is defined to be the sum of its diagonal entries:
d
X
trA = Aii
i=1

If a is a real number (i.e., a 1-by-1 matrix), then tr a = a. (If you haven’t

seen this “operator notation” before, you should think of the trace of A as
tr(A), or as application of the “trace” function to the matrix A. It’s more
commonly written without the parentheses, however.)
The trace operator has the property that for two matrices A and B such
that AB is square, we have that trAB = trBA. (Check this yourself!) As
corollaries of this, we also have, e.g.,
trABC = trCAB = trBCA,
trABCD = trDABC = trCDAB = trBCDA.
The following properties of the trace operator are also easily verified. Here,
A and B are square matrices, and a is a real number:
trA = trAT
tr(A + B) = trA + trB
tr aA = atrA
 9

We now state without proof some facts of matrix derivatives (we won’t
need some of these until later this quarter). Equation (4) applies only to
non-singular square matrices A, where |A| denotes the determinant of A. We
have:

∇A trAB = BT (1)
∇AT f (A) = (∇A f (A))T (2)
∇A trABAT C = CAB + C T AB T (3)
∇A |A| = |A|(A−1 )T . (4)

To make our matrix notation more concrete, let us now explain in detail
the meaning of the first of these equations. Suppose we have some fixed
matrix B ∈ Rd×n . We can then define a function f : Rn×d 7→ R according to
f (A) = trAB. Note that this definition makes sense, because if A ∈ Rn×d ,
then AB is a square matrix, and we can apply the trace operator to it; thus,
f does indeed map from Rn×d to R. We can then apply our definition of
matrix derivatives to find ∇A f (A), which will itself by an n-by-d matrix.
Equation (1) above states that the (i, j) entry of this matrix will be given by
the (i, j)-entry of B T , or equivalently, by Bji .
The proofs of Equations (1-3) are reasonably simple, and are left as an
exercise to the reader. Equations (4) can be derived using the adjoint repre-
sentation of the inverse of a matrix.3

2.2 Least squares revisited

Armed with the tools of matrix derivatives, let us now proceed to find in
closed-form the value of θ that minimizes J(θ). We begin by re-writing J in
matrix-vectorial notation.
Given a training set, define the design matrix X to be the n-by-d matrix
(actually n-by-d + 1, if we include the intercept term) that contains the
3
If we define A′ to be the matrix whose (i, j) element is (−1)i+j times the determinant
of the square matrix resulting from deleting row i and column j from A, then it can be
proved that A−1 = (A′ )T /|A|. (You can check that this is consistent with the standard
way of finding A−1 when A is a 2-by-2 matrix. If you want to see a proof of this more
general result, see an intermediate or advanced linear algebra text, such as Charles Curtis,
1991, Linear Algebra, Springer.) This
P shows that A′ = |A|(A−1 )T . Also, the determinant
of a matrix can be written |A| = j Aij Aij . Since (A′ )ij does not depend on Aij (as can
′

be seen from its definition), this implies that (∂/∂Aij )|A| = A′ij . Putting all this together
shows the result.
 10

training examples’ input values in its rows:

 
— (x(1) )T —
 — (x(2) )T — 
X= .. .
 
 . 
(n) T
— (x ) —

Also, let ~y be the n-dimensional vector containing all the target values from
the training set:  
y (1)
 y (2) 
~y =  ..  .
 
 . 
y (n)
Now, since hθ (x(i) ) = (x(i) )T θ, we can easily verify that
   
(x(1) )T θ y (1)
Xθ − ~y =  ..   .. 
− . 

.
(x(n) )T θ y (n)
 
hθ (x(1) ) − y (1)
=  ..
.
 
.
(n) (n)
hθ (x ) − y
2
Thus, using the fact that for a vector z, we have that z T z =
P
i zi :

n
1 1X
(Xθ − ~y )T (Xθ − ~y ) = (hθ (x(i) ) − y (i) )2
2 2 i=1
= J(θ)

Finally, to minimize J, let’s find its derivatives with respect to θ. Combining

Equations (2) and (3), we find that

∇AT trABAT C = B T AT C T + BAT C (5)

 11

Hence,
1
∇θ J(θ) = ∇θ (Xθ − ~y )T (Xθ − ~y )
2
1
∇θ θT X T Xθ − θT X T ~y − ~y T Xθ + ~y T ~y


=
2
1
∇θ tr θT X T Xθ − θT X T ~y − ~y T Xθ + ~y T ~y


=
2
1
∇θ tr θT X T Xθ − 2tr ~y T Xθ


=
2
1
X T Xθ + X T Xθ − 2X T ~y


=
2
= X T Xθ − X T ~y

In the third step, we used the fact that the trace of a real number is just the
real number; the fourth step used the fact that trA = trAT , and the fifth
step used Equation (5) with AT = θ, B = B T = X T X, and C = I, and
Equation (1). To minimize J, we set its derivatives to zero, and obtain the
normal equations:
X T Xθ = X T ~y
Thus, the value of θ that minimizes J(θ) is given in closed form by the
equation
θ = (X T X)−1 X T ~y .

3 Probabilistic interpretation
When faced with a regression problem, why might linear regression, and
specifically why might the least-squares cost function J, be a reasonable
choice? In this section, we will give a set of probabilistic assumptions, under
which least-squares regression is derived as a very natural algorithm.
Let us assume that the target variables and the inputs are related via the
equation
y (i) = θT x(i) + ǫ(i) ,
where ǫ(i) is an error term that captures either unmodeled effects (such as
if there are some features very pertinent to predicting housing price, but
that we’d left out of the regression), or random noise. Let us further assume
that the ǫ(i) are distributed IID (independently and identically distributed)
according to a Gaussian distribution (also called a Normal distribution) with
 12

mean zero and some variance σ 2 . We can write this assumption as “ǫ(i) ∼
N (0, σ 2 ).” I.e., the density of ǫ(i) is given by

(ǫ(i) )2
 
(i) 1
p(ǫ ) = √ exp − .
2πσ 2σ 2

This implies that

(y (i) − θT x(i) )2
 
(i) (i) 1
p(y |x ; θ) = √ exp − .
2πσ 2σ 2

The notation “p(y (i) |x(i) ; θ)” indicates that this is the distribution of y (i)
given x(i) and parameterized by θ. Note that we should not condition on θ
(“p(y (i) |x(i) , θ)”), since θ is not a random variable. We can also write the
distribution of y (i) as y (i) | x(i) ; θ ∼ N (θT x(i) , σ 2 ).
Given X (the design matrix, which contains all the x(i) ’s) and θ, what
is the distribution of the y (i) ’s? The probability of the data is given by
p(~y |X; θ). This quantity is typically viewed a function of ~y (and perhaps X),
for a fixed value of θ. When we wish to explicitly view this as a function of
θ, we will instead call it the likelihood function:

L(θ) = L(θ; X, ~y ) = p(~y |X; θ).

Note that by the independence assumption on the ǫ(i) ’s (and hence also the
y (i) ’s given the x(i) ’s), this can also be written
n
Y
L(θ) = p(y (i) | x(i) ; θ)
i=1
n
(y (i) − θT x(i) )2
 
Y 1
= √ exp − .
i=1
2πσ 2σ 2

Now, given this probabilistic model relating the y (i) ’s and the x(i) ’s, what
is a reasonable way of choosing our best guess of the parameters θ? The
principal of maximum likelihood says that we should choose θ so as to
make the data as high probability as possible. I.e., we should choose θ to
maximize L(θ).
Instead of maximizing L(θ), we can also maximize any strictly increasing
function of L(θ). In particular, the derivations will be a bit simpler if we
 13

instead maximize the log likelihood ℓ(θ):

ℓ(θ) = log L(θ)

n
(y (i) − θT x(i) )2
 
Y 1
= log √ exp −
i=1
2πσ 2σ 2
n
(y (i) − θT x(i) )2
 
X 1
= log √ exp −
i=1
2πσ 2σ 2
n
1 1 1 X (i)
= n log √ − 2· (y − θT x(i) )2 .
2πσ σ 2 i=1

Hence, maximizing ℓ(θ) gives the same answer as minimizing

n
1 X (i)
(y − θT x(i) )2 ,
2 i=1

which we recognize to be J(θ), our original least-squares cost function.

To summarize: Under the previous probabilistic assumptions on the data,
least-squares regression corresponds to finding the maximum likelihood esti-
mate of θ. This is thus one set of assumptions under which least-squares re-
gression can be justified as a very natural method that’s just doing maximum
likelihood estimation. (Note however that the probabilistic assumptions are
by no means necessary for least-squares to be a perfectly good and rational
procedure, and there may—and indeed there are—other natural assumptions
that can also be used to justify it.)
Note also that, in our previous discussion, our final choice of θ did not
depend on what was σ 2 , and indeed we’d have arrived at the same result
even if σ 2 were unknown. We will use this fact again later, when we talk
about the exponential family and generalized linear models.

4 Locally weighted linear regression

Consider the problem of predicting y from x ∈ R. The leftmost figure below
shows the result of fitting a y = θ0 + θ1 x to a dataset. We see that the data
doesn’t really lie on straight line, and so the fit is not very good.
 14

4.5 4.5 4.5

4 4 4

3.5 3.5 3.5

3 3 3

2.5 2.5 2.5

y

y
2 2 2

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
x x x

Instead, if we had added an extra feature x2 , and fit y = θ0 + θ1 x + θ2 x2 ,

then we obtain a slightly better fit to the data. (See middle figure) Naively, it
might seem that the more features we add, the better. However, there is also
a danger in adding too many features: P The rightmost figure is the result of
fitting a 5-th order polynomial y = 5j=0 θj xj . We see that even though the
fitted curve passes through the data perfectly, we would not expect this to
be a very good predictor of, say, housing prices (y) for different living areas
(x). Without formally defining what these terms mean, we’ll say the figure
on the left shows an instance of underfitting—in which the data clearly
shows structure not captured by the model—and the figure on the right is
an example of overfitting. (Later in this class, when we talk about learning
theory we’ll formalize some of these notions, and also define more carefully
just what it means for a hypothesis to be good or bad.)
As discussed previously, and as shown in the example above, the choice of
features is important to ensuring good performance of a learning algorithm.
(When we talk about model selection, we’ll also see algorithms for automat-
ically choosing a good set of features.) In this section, let us talk briefly talk
about the locally weighted linear regression (LWR) algorithm which, assum-
ing there is sufficient training data, makes the choice of features less critical.
This treatment will be brief, since you’ll get a chance to explore some of the
properties of the LWR algorithm yourself in the homework.
In the original linear regression algorithm, to make a prediction at a query
point x (i.e., to evaluate h(x)), we would:
1. Fit θ to minimize i (y (i) − θT x(i) )2 .
P

2. Output θT x.
In contrast, the locally weighted linear regression algorithm does the fol-
lowing:
1. Fit θ to minimize i w(i) (y (i) − θT x(i) )2 .
P

2. Output θT x.
 15

Here, the w(i) ’s are non-negative valued weights. Intuitively, if w(i) is large
for a particular value of i, then in picking θ, we’ll try hard to make (y (i) −
θT x(i) )2 small. If w(i) is small, then the (y (i) − θT x(i) )2 error term will be
pretty much ignored in the fit.
A fairly standard choice for the weights is4

(x(i) − x)2
 
(i)
w = exp −
2τ 2

Note that the weights depend on the particular point x at which we’re trying
to evaluate x. Moreover, if |x(i) − x| is small, then w(i) is close to 1; and
if |x(i) − x| is large, then w(i) is small. Hence, θ is chosen giving a much
higher “weight” to the (errors on) training examples close to the query point
x. (Note also that while the formula for the weights takes a form that is
cosmetically similar to the density of a Gaussian distribution, the w(i) ’s do
not directly have anything to do with Gaussians, and in particular the w(i)
are not random variables, normally distributed or otherwise.) The parameter
τ controls how quickly the weight of a training example falls off with distance
of its x(i) from the query point x; τ is called the bandwidth parameter, and
is also something that you’ll get to experiment with in your homework.
Locally weighted linear regression is the first example we’re seeing of a
non-parametric algorithm. The (unweighted) linear regression algorithm
that we saw earlier is known as a parametric learning algorithm, because
it has a fixed, finite number of parameters (the θi ’s), which are fit to the
data. Once we’ve fit the θi ’s and stored them away, we no longer need to
keep the training data around to make future predictions. In contrast, to
make predictions using locally weighted linear regression, we need to keep
the entire training set around. The term “non-parametric” (roughly) refers
to the fact that the amount of stuff we need to keep in order to represent the
hypothesis h grows linearly with the size of the training set.

4
If x is vector-valued, this is generalized to be w(i) = exp(−(x(i) − x)T (x(i) − x)/(2τ 2 )),
or w(i) = exp(−(x(i) − x)T Σ−1 (x(i) − x)/2), for an appropriate choice of τ or Σ.
 16

Part II
Classification and logistic
regression
Let’s now talk about the classification problem. This is just like the regression
problem, except that the values y we now want to predict take on only
a small number of discrete values. For now, we will focus on the binary
classification problem in which y can take on only two values, 0 and 1.
(Most of what we say here will also generalize to the multiple-class case.)
For instance, if we are trying to build a spam classifier for email, then x(i)
may be some features of a piece of email, and y may be 1 if it is a piece
of spam mail, and 0 otherwise. 0 is also called the negative class, and 1
the positive class, and they are sometimes also denoted by the symbols “-”
and “+.” Given x(i) , the corresponding y (i) is also called the label for the
training example.

5 Logistic regression
We could approach the classification problem ignoring the fact that y is
discrete-valued, and use our old linear regression algorithm to try to predict
y given x. However, it is easy to construct examples where this method
performs very poorly. Intuitively, it also doesn’t make sense for hθ (x) to take
values larger than 1 or smaller than 0 when we know that y ∈ {0, 1}.
To fix this, let’s change the form for our hypotheses hθ (x). We will choose
1
hθ (x) = g(θT x) = ,
1 + e−θT x
where
1
g(z) =
1 + e−z
is called the logistic function or the sigmoid function. Here is a plot
showing g(z):
 17

0.9

0.8

0.7

0.6
g(z)
0.5

0.4

0.3

0.2

0.1

0
−5 −4 −3 −2 −1 0 1 2 3 4 5
z

Notice that g(z) tends towards 1 as z → ∞, and g(z) tends towards 0 as

z → −∞. Moreover, g(z), and hence also h(x), is always bounded between
0 and 1. AsP before, we are keeping the convention of letting x0 = 1, so that
θT x = θ0 + dj=1 θj xj .
For now, let’s take the choice of g as given. Other functions that smoothly
increase from 0 to 1 can also be used, but for a couple of reasons that we’ll see
later (when we talk about GLMs, and when we talk about generative learning
algorithms), the choice of the logistic function is a fairly natural one. Before
moving on, here’s a useful property of the derivative of the sigmoid function,
which we write as g ′ :
d 1
g ′ (z) =
dz 1 + e−z
1
e−z


= −z 2
(1 + e )
 
1 1
= · 1−
(1 + e−z ) (1 + e−z )
= g(z)(1 − g(z)).

So, given the logistic regression model, how do we fit θ for it? Following
how we saw least squares regression could be derived as the maximum like-
lihood estimator under a set of assumptions, let’s endow our classification
model with a set of probabilistic assumptions, and then fit the parameters
via maximum likelihood.
 18

Let us assume that

P (y = 1 | x; θ) = hθ (x)
P (y = 0 | x; θ) = 1 − hθ (x)

Note that this can be written more compactly as

p(y | x; θ) = (hθ (x))y (1 − hθ (x))1−y

Assuming that the n training examples were generated independently, we

can then write down the likelihood of the parameters as

L(θ) = p(~y | X; θ)
Yn
= p(y (i) | x(i) ; θ)
i=1
n
Y
y(i)
1−y(i)
= hθ (x(i) ) 1 − hθ (x(i) )
i=1

As before, it will be easier to maximize the log likelihood:

ℓ(θ) = log L(θ)

Xn
= y (i) log h(x(i) ) + (1 − y (i) ) log(1 − h(x(i) ))
i=1

How do we maximize the likelihood? Similar to our derivation in the case

of linear regression, we can use gradient ascent. Written in vectorial notation,
our updates will therefore be given by θ := θ + α∇θ ℓ(θ). (Note the positive
rather than negative sign in the update formula, since we’re maximizing,
rather than minimizing, a function now.) Let’s start by working with just
one training example (x, y), and take derivatives to derive the stochastic
gradient ascent rule:
 
∂ 1 1 ∂
ℓ(θ) = y − (1 − y) g(θT x)
∂θj T
g(θ x) 1 − g(θ x) ∂θj
T
 
1 1 ∂ T
= y − (1 − y) g(θT x)(1 − g(θT x)) θ x
T
g(θ x) 1 − g(θ x)
T ∂θj
= y(1 − g(θT x)) − (1 − y)g(θT x) xj

= (y − hθ (x)) xj
 19

Above, we used the fact that g ′ (z) = g(z)(1 − g(z)). This therefore gives us
the stochastic gradient ascent rule

(i)
θj := θj + α y (i) − hθ (x(i) ) xj

If we compare this to the LMS update rule, we see that it looks identical; but
this is not the same algorithm, because hθ (x(i) ) is now defined as a non-linear
function of θT x(i) . Nonetheless, it’s a little surprising that we end up with
the same update rule for a rather different algorithm and learning problem.
Is this coincidence, or is there a deeper reason behind this? We’ll answer this
when we get to GLM models. (See also the extra credit problem on Q3 of
problem set 1.)

6 Digression: The perceptron learning algo-

rithn
We now digress to talk briefly about an algorithm that’s of some historical
interest, and that we will also return to later when we talk about learning
theory. Consider modifying the logistic regression method to “force” it to
output values that are either 0 or 1 or exactly. To do so, it seems natural to
change the definition of g to be the threshold function:

1 if z ≥ 0
g(z) =
0 if z < 0

If we then let hθ (x) = g(θT x) as before but using this modified definition of
g, and if we use the update rule

(i)
θj := θj + α y (i) − hθ (x(i) ) xj .

then we have the perceptron learning algorithn.

In the 1960s, this “perceptron” was argued to be a rough model for how
individual neurons in the brain work. Given how simple the algorithm is, it
will also provide a starting point for our analysis when we talk about learning
theory later in this class. Note however that even though the perceptron may
be cosmetically similar to the other algorithms we talked about, it is actually
a very different type of algorithm than logistic regression and least squares
linear regression; in particular, it is difficult to endow the perceptron’s predic-
tions with meaningful probabilistic interpretations, or derive the perceptron
as a maximum likelihood estimation algorithm.
 20

7 Another algorithm for maximizing ℓ(θ)

Returning to logistic regression with g(z) being the sigmoid function, let’s
now talk about a different algorithm for maximizing ℓ(θ).
To get us started, let’s consider Newton’s method for finding a zero of a
function. Specifically, suppose we have some function f : R 7→ R, and we
wish to find a value of θ so that f (θ) = 0. Here, θ ∈ R is a real number.
Newton’s method performs the following update:
f (θ)
θ := θ − ′ .
f (θ)
This method has a natural interpretation in which we can think of it as
approximating the function f via a linear function that is tangent to f at
the current guess θ, solving for where that linear function equals to zero, and
letting the next guess for θ be where that linear function is zero.
Here’s a picture of the Newton’s method in action:
60 60 60

50 50 50

40 40 40

30 30 30
f(x)

f(x)

f(x)

20 20 20

10 10 10

0 0 0

−10 −10 −10

1 1.5 2 2.5 3 3.5 4 4.5 5 1 1.5 2 2.5 3 3.5 4 4.5 5 1 1.5 2 2.5 3 3.5 4 4.5 5
x x x

In the leftmost figure, we see the function f plotted along with the line
y = 0. We’re trying to find θ so that f (θ) = 0; the value of θ that achieves this
is about 1.3. Suppose we initialized the algorithm with θ = 4.5. Newton’s
method then fits a straight line tangent to f at θ = 4.5, and solves for the
where that line evaluates to 0. (Middle figure.) This give us the next guess
for θ, which is about 2.8. The rightmost figure shows the result of running
one more iteration, which the updates θ to about 1.8. After a few more
iterations, we rapidly approach θ = 1.3.
Newton’s method gives a way of getting to f (θ) = 0. What if we want to
use it to maximize some function ℓ? The maxima of ℓ correspond to points
where its first derivative ℓ′ (θ) is zero. So, by letting f (θ) = ℓ′ (θ), we can use
the same algorithm to maximize ℓ, and we obtain update rule:
ℓ′ (θ)
θ := θ − ′′ .
ℓ (θ)
(Something to think about: How would this change if we wanted to use
Newton’s method to minimize rather than maximize a function?)
 21

Lastly, in our logistic regression setting, θ is vector-valued, so we need to

generalize Newton’s method to this setting. The generalization of Newton’s
method to this multidimensional setting (also called the Newton-Raphson
method) is given by
θ := θ − H −1 ∇θ ℓ(θ).
Here, ∇θ ℓ(θ) is, as usual, the vector of partial derivatives of ℓ(θ) with respect
to the θi ’s; and H is an d-by-d matrix (actually, d+1−by−d+1, assuming that
we include the intercept term) called the Hessian, whose entries are given
by
∂ 2 ℓ(θ)
Hij = .
∂θi ∂θj
Newton’s method typically enjoys faster convergence than (batch) gra-
dient descent, and requires many fewer iterations to get very close to the
minimum. One iteration of Newton’s can, however, be more expensive than
one iteration of gradient descent, since it requires finding and inverting an
d-by-d Hessian; but so long as d is not too large, it is usually much faster
overall. When Newton’s method is applied to maximize the logistic regres-
sion log likelihood function ℓ(θ), the resulting method is also called Fisher
scoring.
 22

Part III
Generalized Linear Models5
So far, we’ve seen a regression example, and a classification example. In the
regression example, we had y|x; θ ∼ N (µ, σ 2 ), and in the classification one,
y|x; θ ∼ Bernoulli(φ), for some appropriate definitions of µ and φ as functions
of x and θ. In this section, we will show that both of these methods are
special cases of a broader family of models, called Generalized Linear Models
(GLMs). We will also show how other models in the GLM family can be
derived and applied to other classification and regression problems.

8 The exponential family

To work our way up to GLMs, we will begin by defining exponential family
distributions. We say that a class of distributions is in the exponential family
if it can be written in the form

p(y; η) = b(y) exp(η T T (y) − a(η)) (6)

Here, η is called the natural parameter (also called the canonical param-
eter) of the distribution; T (y) is the sufficient statistic (for the distribu-
tions we consider, it will often be the case that T (y) = y); and a(η) is the log
partition function. The quantity e−a(η) essentially plays the role of a nor-
malization constant, that makes sure the distribution p(y; η) sums/integrates
over y to 1.
A fixed choice of T , a and b defines a family (or set) of distributions that
is parameterized by η; as we vary η, we then get different distributions within
this family.
We now show that the Bernoulli and the Gaussian distributions are ex-
amples of exponential family distributions. The Bernoulli distribution with
mean φ, written Bernoulli(φ), specifies a distribution over y ∈ {0, 1}, so that
p(y = 1; φ) = φ; p(y = 0; φ) = 1 − φ. As we vary φ, we obtain Bernoulli
distributions with different means. We now show that this class of Bernoulli
distributions, ones obtained by varying φ, is in the exponential family; i.e.,
that there is a choice of T , a and b so that Equation (6) becomes exactly the
class of Bernoulli distributions.
5
The presentation of the material in this section takes inspiration from Michael I.
Jordan, Learning in graphical models (unpublished book draft), and also McCullagh and
Nelder, Generalized Linear Models (2nd ed.).
 23

We write the Bernoulli distribution as:

p(y; φ) = φy (1 − φ)1−y
= exp(y log φ + (1 − y) log(1 − φ))
   
φ
= exp log y + log(1 − φ) .
1−φ

Thus, the natural parameter is given by η = log(φ/(1 − φ)). Interestingly, if

we invert this definition for η by solving for φ in terms of η, we obtain φ =
1/(1 + e−η ). This is the familiar sigmoid function! This will come up again
when we derive logistic regression as a GLM. To complete the formulation
of the Bernoulli distribution as an exponential family distribution, we also
have

T (y) = y
a(η) = − log(1 − φ)
= log(1 + eη )
b(y) = 1

This shows that the Bernoulli distribution can be written in the form of
Equation (6), using an appropriate choice of T , a and b.
Let’s now move on to consider the Gaussian distribution. Recall that,
when deriving linear regression, the value of σ 2 had no effect on our final
choice of θ and hθ (x). Thus, we can choose an arbitrary value for σ 2 without
changing anything. To simplify the derivation below, let’s set σ 2 = 1.6 We
then have:
 
1 1 2
p(y; µ) = √ exp − (y − µ)
2π 2
   
1 1 2 1 2
= √ exp − y · exp µy − µ
2π 2 2
6
If we leave σ 2 as a variable, the Gaussian distribution can also be shown to be in the
exponential family, where η ∈ R2 is now a 2-dimension vector that depends on both µ and
σ. For the purposes of GLMs, however, the σ 2 parameter can also be treated by considering
a more general definition of the exponential family: p(y; η, τ ) = b(a, τ ) exp((η T T (y) −
a(η))/c(τ )). Here, τ is called the dispersion parameter, and for the Gaussian, c(τ ) = σ 2 ;
but given our simplification above, we won’t need the more general definition for the
examples we will consider here.
 24

Thus, we see that the Gaussian is in the exponential family, with

η = µ
T (y) = y
a(η) = µ2 /2
= η 2 /2
√
b(y) = (1/ 2π) exp(−y 2 /2).

There’re many other distributions that are members of the exponen-

tial family: The multinomial (which we’ll see later), the Poisson (for mod-
elling count-data; also see the problem set); the gamma and the exponen-
tial (for modelling continuous, non-negative random variables, such as time-
intervals); the beta and the Dirichlet (for distributions over probabilities);
and many more. In the next section, we will describe a general “recipe”
for constructing models in which y (given x and θ) comes from any of these
distributions.

9 Constructing GLMs
Suppose you would like to build a model to estimate the number y of cus-
tomers arriving in your store (or number of page-views on your website) in
any given hour, based on certain features x such as store promotions, recent
advertising, weather, day-of-week, etc. We know that the Poisson distribu-
tion usually gives a good model for numbers of visitors. Knowing this, how
can we come up with a model for our problem? Fortunately, the Poisson is an
exponential family distribution, so we can apply a Generalized Linear Model
(GLM). In this section, we will we will describe a method for constructing
GLM models for problems such as these.
More generally, consider a classification or regression problem where we
would like to predict the value of some random variable y as a function of
x. To derive a GLM for this problem, we will make the following three
assumptions about the conditional distribution of y given x and about our
model:
1. y | x; θ ∼ ExponentialFamily(η). I.e., given x and θ, the distribution of
y follows some exponential family distribution, with parameter η.
2. Given x, our goal is to predict the expected value of T (y) given x.
In most of our examples, we will have T (y) = y, so this means we
would like the prediction h(x) output by our learned hypothesis h to
 25

satisfy h(x) = E[y|x]. (Note that this assumption is satisfied in the

choices for hθ (x) for both logistic regression and linear regression. For
instance, in logistic regression, we had hθ (x) = p(y = 1|x; θ) = 0 · p(y =
0|x; θ) + 1 · p(y = 1|x; θ) = E[y|x; θ].)

3. The natural parameter η and the inputs x are related linearly: η = θT x.

(Or, if η is vector-valued, then ηi = θiT x.)

The third of these assumptions might seem the least well justified of
the above, and it might be better thought of as a “design choice” in our
recipe for designing GLMs, rather than as an assumption per se. These
three assumptions/design choices will allow us to derive a very elegant class
of learning algorithms, namely GLMs, that have many desirable properties
such as ease of learning. Furthermore, the resulting models are often very
effective for modelling different types of distributions over y; for example, we
will shortly show that both logistic regression and ordinary least squares can
both be derived as GLMs.

9.1 Ordinary Least Squares

To show that ordinary least squares is a special case of the GLM family
of models, consider the setting where the target variable y (also called the
response variable in GLM terminology) is continuous, and we model the
conditional distribution of y given x as a Gaussian N (µ, σ 2 ). (Here, µ may
depend x.) So, we let the ExponentialF amily(η) distribution above be
the Gaussian distribution. As we saw previously, in the formulation of the
Gaussian as an exponential family distribution, we had µ = η. So, we have

hθ (x) = E[y|x; θ]
= µ
= η
= θT x.

The first equality follows from Assumption 2, above; the second equality
follows from the fact that y|x; θ ∼ N (µ, σ 2 ), and so its expected value is given
by µ; the third equality follows from Assumption 1 (and our earlier derivation
showing that µ = η in the formulation of the Gaussian as an exponential
family distribution); and the last equality follows from Assumption 3.
 26

9.2 Logistic Regression

We now consider logistic regression. Here we are interested in binary classifi-
cation, so y ∈ {0, 1}. Given that y is binary-valued, it therefore seems natural
to choose the Bernoulli family of distributions to model the conditional dis-
tribution of y given x. In our formulation of the Bernoulli distribution as
an exponential family distribution, we had φ = 1/(1 + e−η ). Furthermore,
note that if y|x; θ ∼ Bernoulli(φ), then E[y|x; θ] = φ. So, following a similar
derivation as the one for ordinary least squares, we get:

hθ (x) = E[y|x; θ]
= φ
= 1/(1 + e−η )
T
= 1/(1 + e−θ x )
T
So, this gives us hypothesis functions of the form hθ (x) = 1/(1 + e−θ x ). If
you are previously wondering how we came up with the form of the logistic
function 1/(1 + e−z ), this gives one answer: Once we assume that y condi-
tioned on x is Bernoulli, it arises as a consequence of the definition of GLMs
and exponential family distributions.
To introduce a little more terminology, the function g giving the distri-
bution’s mean as a function of the natural parameter (g(η) = E[T (y); η])
is called the canonical response function. Its inverse, g −1 , is called the
canonical link function. Thus, the canonical response function for the
Gaussian family is just the identify function; and the canonical response
function for the Bernoulli is the logistic function.7

9.3 Softmax Regression

Let’s look at one more example of a GLM. Consider a classification problem
in which the response variable y can take on any one of k values, so y ∈
{1, 2, . . . , k}. For example, rather than classifying email into the two classes
spam or not-spam—which would have been a binary classification problem—
we might want to classify it into three classes, such as spam, personal mail,
and work-related mail. The response variable is still discrete, but can now
take on more than two values. We will thus model it as distributed according
to a multinomial distribution.
7
Many texts use g to denote the link function, and g −1 to denote the response function;
but the notation we’re using here, inherited from the early machine learning literature,
will be more consistent with the notation used in the rest of the class.
 27

Let’s derive a GLM for modelling this type of multinomial data. To do

so, we will begin by expressing the multinomial as an exponential family
distribution.
To parameterize a multinomial over k possible outcomes, one could use
k parameters φ1 , . . . , φk specifying the probability of each of the outcomes.
However, these parameters would be redundant, or more formally, they would
not be independent (since knowing any Pk k − 1 of the φi ’s uniquely determines
the last one, as they must satisfy i=1 φi = 1). So, we will instead pa-
rameterize the multinomial with only kP − 1 parameters, φ1 , . . . , φk−1 , where
k−1
φi = p(y = i; φ), and p(y = k; φ) = 1 − i=1 φi . For notational convenience,
Pk−1
we will also let φk = 1 − i=1 φi , but we should keep in mind that this is
not a parameter, and that it is fully specified by φ1 , . . . , φk−1 .
To express the multinomial as an exponential family distribution, we will
define T (y) ∈ Rk−1 as follows:
         
1 0 0 0 0

 0 


 1 


 0 


 0 


 0 

T (1) = 
 0  , T (2) = 
  0  , T (3) = 
  1  , · · · , T (k−1) = 
  0  , T (k) = 
  0 ,

 ..   ..   ..   ..   .. 
 .   .   .   .   . 
0 0 0 1 0

Unlike our previous examples, here we do not have T (y) = y; also, T (y) is
now a k − 1 dimensional vector, rather than a real number. We will write
(T (y))i to denote the i-th element of the vector T (y).
We introduce one more very useful piece of notation. An indicator func-
tion 1{·} takes on a value of 1 if its argument is true, and 0 otherwise
(1{True} = 1, 1{False} = 0). For example, 1{2 = 3} = 0, and 1{3 =
5 − 2} = 1. So, we can also write the relationship between T (y) and y as
(T (y))i = 1{y = i}. (Before you continue reading, please make sure you un-
derstand why this is true!) Further, we have that E[(T (y))i ] = P (y = i) = φi .
We are now ready to show that the multinomial is a member of the
 28

exponential family. We have:

1{y=1} 1{y=2} 1{y=k}
p(y; φ) = φ1 φ2 · · · φk
P
1{y=1} 1{y=2} 1− k−1 1{y=i}
= φ1 φ2 · · · φk i=1
P
(T (y))1 (T (y))2 1− k−1 (T (y))i
= φ1 φ2 · · · φk i=1
= exp((T (y))1 log(φ1 ) + (T (y))2 log(φ2 ) +
 Pk−1 
· · · + 1 − i=1 (T (y))i log(φk ))
= exp((T (y))1 log(φ1 /φk ) + (T (y))2 log(φ2 /φk ) +
· · · + (T (y))k−1 log(φk−1 /φk ) + log(φk ))
= b(y) exp(η T T (y) − a(η))
where
 
log(φ1 /φk )
 log(φ2 /φk ) 
η =  .. ,
 
 . 
log(φk−1 /φk )
a(η) = − log(φk )
b(y) = 1.
This completes our formulation of the multinomial as an exponential family
distribution.
The link function is given (for i = 1, . . . , k) by
φi
ηi = log .
φk
For convenience, we have also defined ηk = log(φk /φk ) = 0. To invert the
link function and derive the response function, we therefore have that
φi
eηi =
φk
φk eηi = φi (7)
k
X Xk
φk eηi = φi = 1
i=1 i=1
Pk
This implies that φk = 1/ i=1 eηi , which can be substituted back into Equa-
tion (7) to give the response function
eηi
φ i = Pk
j=1 eηj
 29

This function mapping from the η’s to the φ’s is called the softmax function.
To complete our model, we use Assumption 3, given earlier, that the ηi ’s
are linearly related to the x’s. So, have ηi = θiT x (for i = 1, . . . , k − 1),
where θ1 , . . . , θk−1 ∈ Rd+1 are the parameters of our model. For notational
convenience, we can also define θk = 0, so that ηk = θkT x = 0, as given
previously. Hence, our model assumes that the conditional distribution of y
given x is given by

p(y = i|x; θ) = φi
eηi
= Pk
j=1 eηj
T
e θi x
= Pk T (8)
j=1 e θj x

This model, which applies to classification problems where y ∈ {1, . . . , k}, is

called softmax regression. It is a generalization of logistic regression.
Our hypothesis will output

hθ (x) = E[T (y)|x; θ]

 
1{y = 1}
 1{y = 2} 
= E .. x; θ
 
 . 
1{y = k − 1}
 
φ1
 φ2 
=  .. 
 
 . 
φk−1
exp(θ T x)
 
1
Pk
 j=1exp(θjT x) 
exp(θ2T x)
Pk
 
T
j=1 exp(θj x)
 
= 
 ..
.

 . 
 T
exp(θk−1 x)

Pk T
j=1 exp(θj x)

In other words, our hypothesis will output the estimated probability that
p(y = i|x; θ), for every value of i = 1, . . . , k. (Even though hθ (x) as defined
above is only k − 1 dimensional, clearly p(y = k|x; θ) can be obtained as
Pk−1
1 − i=1 φi .)
 30

Lastly, let’s discuss parameter fitting. Similar to our original derivation

of ordinary least squares and logistic regression, if we have a training set of
n examples {(x(i) , y (i) ); i = 1, . . . , n} and would like to learn the parameters
θi of this model, we would begin by writing down the log-likelihood
n
X
ℓ(θ) = log p(y (i) |x(i) ; θ)
i=1
n k T x(i)
!1{y(i) =l}
X Y e θl
= log Pk T (i)
i=1 l=1 j=1 e θj x

To obtain the second line above, we used the definition for p(y|x; θ) given
in Equation (8). We can now obtain the maximum likelihood estimate of
the parameters by maximizing ℓ(θ) in terms of θ, using a method such as
gradient ascent or Newton’s method.
 CS229 Supplemental Lecture notes
John Duchi

1 Binary classification
In binary classification problems, the target y can take on at only two
values. In this set of notes, we show how to model this problem by letting
y ∈ {−1, +1}, where we say that y is a 1 if the example is a member of the
positive class and y = −1 if the example is a member of the negative class.
We assume, as usual, that we have input features x ∈ Rn .
As in our standard approach to supervised learning problems, we first
pick a representation for our hypothesis class (what we are trying to learn),
and after that we pick a loss function that we will minimize. In binary
classification problems, it is often convenient to use a hypothesis class of the
form hθ (x) = θT x, and, when presented with a new example x, we classify it
as positive or negative depending on the sign of θT x, that is, our predicted
label is

1
 if t > 0
T
sign(hθ (x)) = sign(θ x) where sign(t) = 0 if t = 0

−1 if t < 0.


In a binary classification problem, then, the hypothesis hθ with parameter

vector θ classifies a particular example (x, y) correctly if
sign(θT x) = y or equivalently yθT x > 0. (1)
The quantity yθT x in expression (1) is a very important quantity in binary
classification, important enough that we call the value
yxT θ
the margin for the example (x, y). Often, though not always, one interprets
the value hθ (x) = xT θ as a measure of the confidence that the parameter

1
vector θ assigns to labels for the point x: if xT θ is very negative (or very
positive), then we more strongly believe the label y is negative (or positive).
Now that we have chosen a representation for our data, we must choose a
loss function. Intuitively, we would like to choose some loss function so that
for our training data {(x(i) , y (i) )}m (i) T (i)
i=1 , the θ chosen makes the margin y θ x
very large for each training example. Let us fix a hypothetical example (x, y),
let z = yxT θ denote the margin, and let ϕ : R → R be the loss function—that
is, the loss for the example (x, y) with margin z = yxT θ is ϕ(z) = ϕ(yxT θ).
For any particular loss function, the empirical risk that we minimize is then
m
1 X
J(θ) = ϕ(y (i) θT x(i) ). (2)
m i=1

Consider our desired behavior: we wish to have y (i) θT x(i) positive for each
training example i = 1, . . . , m, and we should penalize those θ for which
y (i) θT x(i) < 0 frequently in the training data. Thus, an intuitive choice for
our loss would be one with ϕ(z) small if z > 0 (the margin is positive), while
ϕ(z) is large if z < 0 (the margin is negative). Perhaps the most natural
such loss is the zero-one loss, given by
(
1 if z ≤ 0
ϕzo (z) =
0 if z > 0.

In this case, the risk J(θ) is simply the average number of mistakes—misclassifications—
the parameter θ makes on the training data. Unfortunately, the loss ϕzo is
discontinuous, non-convex (why this matters is a bit beyond the scope of
the course), and perhaps even more vexingly, NP-hard to minimize. So we
prefer to choose losses that have the shape given in Figure 1. That is, we
will essentially always use losses that satisfy

ϕ(z) → 0 as z → ∞, while ϕ(z) → ∞ as z → −∞.

As a few different examples, here are three loss functions that we will see
either now or later in the class, all of which are commonly used in machine
learning.

(i) The logistic loss uses

ϕlogistic (z) = log(1 + e−z )

2
 ϕ

z = yxT θ

Figure 1: The rough shape of loss we desire: the loss is convex and continuous,
and tends to zero as the margin z = yxT θ → ∞.

(ii) The hinge loss uses

ϕhinge (z) = [1 − z]+ = max{1 − z, 0}

(iii) The exponential loss uses

ϕexp (z) = e−z .

In Figure 2, we plot each of these losses against the margin z = yxT θ,

noting that each goes to zero as the margin grows, and each tends to +∞ as
the margin becomes negative. The different loss functions lead to different
machine learning procedures; in particular, the logistic loss ϕlogistic is logistic
regression, the hinge loss ϕhinge gives rise to so-called support vector machines,
and the exponential loss gives rise to the classical version of boosting, both
of which we will explore in more depth later in the class.

2 Logistic regression
With this general background in place, we now we give a complementary
view of logistic regression to that in Andrew Ng’s lecture notes. When we

3
 ϕlogistic
ϕhinge
ϕexp

z = yxT θ

Figure 2: The three margin-based loss functions logistic loss, hinge loss, and
exponential loss.

use binary labels y ∈ {−1, 1}, it is possible to write logistic regression more
compactly. In particular, we use the logistic loss

ϕlogistic (yxT θ) = log 1 + exp(−yxT θ) ,

and the logistic regression algorithm corresponds to choosing θ that mini-

mizes
m m
1 X 1 X
ϕlogistic (y (i) θT x(i) ) = log 1 + exp(−y (i) θT x(i) ) .


J(θ) = (3)
m i=1 m i=1

Roughly, we hope that choosing θ to minimize the average logistic loss will
yield a θ for which y (i) θT x(i) > 0 for most (or even all!) of the training
examples.

2.1 Probabilistic intrepretation

Similar to the linear regression (least-squares) case, it is possible to give a
probabilistic interpretation of logistic regression. To do this, we define the

4
 1

0
-5 5

Figure 3: Sigmoid function

sigmoid function (also often called the logistic function)

1
g(z) = ,
1 + e−z
which is plotted in Fig. 3. In particular, the sigmoid function satisfies
1 1 ez 1
g(z) + g(−z) = + z
= z
+ = 1,
1+e −z 1+e 1+e 1 + ez
so we can use it to define a probability model for binary classification. In
particular, for y ∈ {−1, 1}, we define the logistic model for classification as
1
p(Y = y | x; θ) = g(yxT θ) = . (4)
1 + e−yxT θ
For intepretation, we see that if the margin yxT θ is large—bigger than, say,
5 or so—then p(Y = y | x; θ) = g(yxT θ) ≈ 1, that is, we assign nearly
probability 1 to the event that the label is y. Conversely, if yxT θ is quite
negative, then p(Y = y | x; θ) ≈ 0.
By redefining our hypothesis class as
1
hθ (x) = g(θT x) = ,
1 + e−θT x

5
then we see that the likelihood of the training data is
m
Y m
Y
(i) (i)
L(θ) = p(Y = y | x ; θ) = hθ (y (i) x(i) ),
i=1 i=1

and the log-likelihood is precisely

m m  
(i) T (i)
X X
(i) (i)
ℓ(θ) = log hθ (y x ) = − log 1 + e−y θ x = −mJ(θ),
i=1 i=1

where J(θ) is exactly the logistic regression risk from Eq. (3). That is,
maximum likelihood in the logistic model (4) is the same as minimizing the
average logistic loss, and we arrive at logistic regression again.

2.2 Gradient descent methods

The final part of logistic regression is to actually fit the model. As is usually
the case, we consider gradient-descent-based procedures for performing this
minimization. With that in mind, we now show how to take derivatives of
the logistic loss. For ϕlogistic (z) = log(1 + e−z ), we have the one-dimensional
derivative
d 1 d −z e−z 1
ϕlogistic (z) = ϕ′logistic (z) = · e = − =− = −g(−z),
dz 1 + e dz
−z 1+e −z 1 + ez
where g is the sigmoid function. Then we apply the chain rule to find that
for a single training example (x, y), we have
∂ ∂
ϕlogistic (yxT θ) = −g(−yxT θ) (yxT θ) = −g(−yxT θ)yxk .
∂θk ∂θk
Thus, a stochastic gradient procedure for minimization of J(θ) iteratively
performs the following for iterations t = 1, 2, . . ., where αt is a stepsize at
time t:
1. Choose an example i ∈ {1, . . . , m} uniformly at random
2. Perform the gradient update
T
θ(t+1) = θ(t) − αt · ∇θ ϕlogistic (y (i) x(i) θ(t) )
T
= θ(t) + αt g(−y (i) x(i) θ(t) )y (i) x(i) = θ(t) + αt hθ(t) (−y (i) x(i) )y (i) x(i) .

6
This update is intuitive: if our current hypothesis hθ(t) assigns probability
close to 1 for the incorrect label −y (i) , then we try to reduce the loss by
moving θ in the direction of y (i) x(i) . Conversely, if our current hypothesis
hθ(t) assigns probability close to 0 for the incorrect label −y (i) , the update
essentially does nothing.

7
 CS229 Lecture Notes
Andrew Ng

Part IV
Generative Learning algorithms
So far, we’ve mainly been talking about learning algorithms that model
p(y|x; θ), the conditional distribution of y given x. For instance, logistic
regression modeled p(y|x; θ) as hθ (x) = g(θT x) where g is the sigmoid func-
tion. In these notes, we’ll talk about a different type of learning algorithm.
Consider a classification problem in which we want to learn to distinguish
between elephants (y = 1) and dogs (y = 0), based on some features of
an animal. Given a training set, an algorithm like logistic regression or
the perceptron algorithm (basically) tries to find a straight line—that is, a
decision boundary—that separates the elephants and dogs. Then, to classify
a new animal as either an elephant or a dog, it checks on which side of the
decision boundary it falls, and makes its prediction accordingly.
Here’s a different approach. First, looking at elephants, we can build a
model of what elephants look like. Then, looking at dogs, we can build a
separate model of what dogs look like. Finally, to classify a new animal, we
can match the new animal against the elephant model, and match it against
the dog model, to see whether the new animal looks more like the elephants
or more like the dogs we had seen in the training set.
Algorithms that try to learn p(y|x) directly (such as logistic regression),
or algorithms that try to learn mappings directly from the space of inputs X
to the labels {0, 1}, (such as the perceptron algorithm) are called discrim-
inative learning algorithms. Here, we’ll talk about algorithms that instead
try to model p(x|y) (and p(y)). These algorithms are called generative
learning algorithms. For instance, if y indicates whether an example is a
dog (0) or an elephant (1), then p(x|y = 0) models the distribution of dogs’
features, and p(x|y = 1) models the distribution of elephants’ features.
After modeling p(y) (called the class priors) and p(x|y), our algorithm

1
 2

can then use Bayes rule to derive the posterior distribution on y given x:
p(x|y)p(y)
p(y|x) = .
p(x)
Here, the denominator is given by p(x) = p(x|y = 1)p(y = 1) + p(x|y =
0)p(y = 0) (you should be able to verify that this is true from the standard
properties of probabilities), and thus can also be expressed in terms of the
quantities p(x|y) and p(y) that we’ve learned. Actually, if were calculating
p(y|x) in order to make a prediction, then we don’t actually need to calculate
the denominator, since
p(x|y)p(y)
arg max p(y|x) = arg max
y y p(x)
= arg max p(x|y)p(y).
y

1 Gaussian discriminant analysis

The first generative learning algorithm that we’ll look at is Gaussian discrim-
inant analysis (GDA). In this model, we’ll assume that p(x|y) is distributed
according to a multivariate normal distribution. Let’s talk briefly about the
properties of multivariate normal distributions before moving on to the GDA
model itself.

1.1 The multivariate normal distribution

The multivariate normal distribution in d-dimensions, also called the multi-
variate Gaussian distribution, is parameterized by a mean vector µ ∈ Rd
and a covariance matrix Σ ∈ Rd×d , where Σ ≥ 0 is symmetric and positive
semi-definite. Also written “N (µ, Σ)”, its density is given by:
 
1 1 T −1
p(x; µ, Σ) = exp − (x − µ) Σ (x − µ) .
(2π)d/2 |Σ|1/2 2
In the equation above, “|Σ|” denotes the determinant of the matrix Σ.
For a random variable X distributed N (µ, Σ), the mean is (unsurpris-
ingly) given by µ: Z
E[X] = x p(x; µ, Σ)dx = µ
x
The covariance of a vector-valued random variable Z is defined as Cov(Z) =
E[(Z − E[Z])(Z − E[Z])T ]. This generalizes the notion of the variance of a
 3

real-valued random variable. The covariance can also be defined as Cov(Z) =

E[ZZ T ] − (E[Z])(E[Z])T . (You should be able to prove to yourself that these
two definitions are equivalent.) If X ∼ N (µ, Σ), then

Cov(X) = Σ.

Here are some examples of what the density of a Gaussian distribution

looks like:

0.25 0.25 0.25

0.2 0.2 0.2

0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

3 3 3
2 2 2
3 3 3
1 2 1 2 1 2
0 1 0 1 0 1
−1 0 −1 0 −1 0
−1 −1 −1
−2 −2 −2
−2 −2 −2
−3 −3 −3 −3 −3 −3

The left-most figure shows a Gaussian with mean zero (that is, the 2x1
zero-vector) and covariance matrix Σ = I (the 2x2 identity matrix). A Gaus-
sian with zero mean and identity covariance is also called the standard nor-
mal distribution. The middle figure shows the density of a Gaussian with
zero mean and Σ = 0.6I; and in the rightmost figure shows one with , Σ = 2I.
We see that as Σ becomes larger, the Gaussian becomes more “spread-out,”
and as it becomes smaller, the distribution becomes more “compressed.”
Let’s look at some more examples.

0.25 0.25 0.25

0.2 0.2 0.2

0.15 0.15 0.15

0.1 0.1 0.1

0.05 0.05 0.05

3 3 3

2 2 2

1 1 1

0 0 0
3 3 3
−1 2 −1 2 −1 2
1 1 1
−2 0 −2 0 −2 0
−1 −1 −1
−3 −2 −3 −2 −3 −2
−3 −3 −3

The figures above show Gaussians with mean 0, and with covariance
matrices respectively
     
1 0 1 0.5 1 0.8
Σ= ; Σ= ; Σ= .
0 1 0.5 1 0.8 1

The leftmost figure shows the familiar standard normal distribution, and we
see that as we increase the off-diagonal entry in Σ, the density becomes more
 4

“compressed” towards the 45◦ line (given by x1 = x2 ). We can see this more
clearly when we look at the contours of the same three densities:
3 3 3

2 2 2

1 1 1

0 0 0

−1 −1 −1

−2 −2 −2

−3 −3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

Here’s one last set of examples generated by varying Σ:

3 3 3

2 2 2

1 1 1

0 0 0

−1 −1 −1

−2 −2 −2

−3 −3 −3
−3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3 −3 −2 −1 0 1 2 3

The plots above used, respectively,

     
1 -0.5 1 -0.8 3 0.8
Σ= ; Σ= ; Σ= .
-0.5 1 -0.8 1 0.8 1

From the leftmost and middle figures, we see that by decreasing the off-
diagonal elements of the covariance matrix, the density now becomes “com-
pressed” again, but in the opposite direction. Lastly, as we vary the pa-
rameters, more generally the contours will form ellipses (the rightmost figure
showing an example).
As our last set of examples, fixing Σ = I, by varying µ, we can also move
the mean of the density around.
 5

0.25 0.25

0.2 0.2

0.15 0.15

0.1 0.1

0.05 0.05

3 3
2 2
3 3
1 2 1 2
0 1 0 1
−1 0 −1 0
−1 −1
−2 −2
−2 −2
−3 −3 −3 −3

0.25

0.2

0.15

0.1

0.05

3
2
3
1 2
0 1
−1 0
−1
−2
−2
−3 −3

The figures above were generated using Σ = I, and respectively

     
1 -0.5 -1
µ= ; µ= ; µ= .
0 0 -1.5

1.2 The Gaussian Discriminant Analysis model

When we have a classification problem in which the input features x are
continuous-valued random variables, we can then use the Gaussian Discrim-
inant Analysis (GDA) model, which models p(x|y) using a multivariate nor-
mal distribution. The model is:

y ∼ Bernoulli(φ)
x|y = 0 ∼ N (µ0 , Σ)
x|y = 1 ∼ N (µ1 , Σ)
 6

Writing out the distributions, this is:

p(y) = φy (1 − φ)1−y
 
1 1 T −1
p(x|y = 0) = exp − (x − µ0 ) Σ (x − µ0 )
(2π)d/2 |Σ|1/2 2
 
1 1 T −1
p(x|y = 1) = exp − (x − µ1 ) Σ (x − µ1 )
(2π)d/2 |Σ|1/2 2

Here, the parameters of our model are φ, Σ, µ0 and µ1 . (Note that while
there’re two different mean vectors µ0 and µ1 , this model is usually applied
using only one covariance matrix Σ.) The log-likelihood of the data is given
by
n
Y
`(φ, µ0 , µ1 , Σ) = log p(x(i) , y (i) ; φ, µ0 , µ1 , Σ)
i=1
n
Y
= log p(x(i) |y (i) ; µ0 , µ1 , Σ)p(y (i) ; φ).
i=1

By maximizing ` with respect to the parameters, we find the maximum like-

lihood estimate of the parameters (see problem set 1) to be:
n
1X
φ = 1{y (i) = 1}
n i=1
Pn (i)
i=1 1{y = 0}x(i)
µ0 = P n
1{y (i) = 0}
Pn i=1 (i)
i=1 1{y = 1}x(i)
µ1 = P n (i) = 1}
i=1 1{y
n
1 X (i)
Σ = (x − µy(i) )(x(i) − µy(i) )T .
n i=1

Pictorially, what the algorithm is doing can be seen in as follows:

 7

−1

−2

−3

−4

−5

−6

−7
−2 −1 0 1 2 3 4 5 6 7

Shown in the figure are the training set, as well as the contours of the
two Gaussian distributions that have been fit to the data in each of the
two classes. Note that the two Gaussians have contours that are the same
shape and orientation, since they share a covariance matrix Σ, but they have
different means µ0 and µ1 . Also shown in the figure is the straight line
giving the decision boundary at which p(y = 1|x) = 0.5. On one side of
the boundary, we’ll predict y = 1 to be the most likely outcome, and on the
other side, we’ll predict y = 0.

1.3 Discussion: GDA and logistic regression

The GDA model has an interesting relationship to logistic regression. If we
view the quantity p(y = 1|x; φ, µ0 , µ1 , Σ) as a function of x, we’ll find that it
can be expressed in the form
1
p(y = 1|x; φ, Σ, µ0 , µ1 ) = ,
1 + exp(−θT x)

where θ is some appropriate function of φ, Σ, µ0 , µ1 .1 This is exactly the form

that logistic regression—a discriminative algorithm—used to model p(y =
1|x).
1
This uses the convention of redefining the x(i) ’s on the right-hand-side to be (n + 1)-
(i)
dimensional vectors by adding the extra coordinate x0 = 1; see problem set 1.
 8

When would we prefer one model over another? GDA and logistic regres-
sion will, in general, give different decision boundaries when trained on the
same dataset. Which is better?
We just argued that if p(x|y) is multivariate gaussian (with shared Σ),
then p(y|x) necessarily follows a logistic function. The converse, however,
is not true; i.e., p(y|x) being a logistic function does not imply p(x|y) is
multivariate gaussian. This shows that GDA makes stronger modeling as-
sumptions about the data than does logistic regression. It turns out that
when these modeling assumptions are correct, then GDA will find better fits
to the data, and is a better model. Specifically, when p(x|y) is indeed gaus-
sian (with shared Σ), then GDA is asymptotically efficient. Informally,
this means that in the limit of very large training sets (large n), there is no
algorithm that is strictly better than GDA (in terms of, say, how accurately
they estimate p(y|x)). In particular, it can be shown that in this setting,
GDA will be a better algorithm than logistic regression; and more generally,
even for small training set sizes, we would generally expect GDA to better.
In contrast, by making significantly weaker assumptions, logistic regres-
sion is also more robust and less sensitive to incorrect modeling assumptions.
There are many different sets of assumptions that would lead to p(y|x) taking
the form of a logistic function. For example, if x|y = 0 ∼ Poisson(λ0 ), and
x|y = 1 ∼ Poisson(λ1 ), then p(y|x) will be logistic. Logistic regression will
also work well on Poisson data like this. But if we were to use GDA on such
data—and fit Gaussian distributions to such non-Gaussian data—then the
results will be less predictable, and GDA may (or may not) do well.
To summarize: GDA makes stronger modeling assumptions, and is more
data efficient (i.e., requires less training data to learn “well”) when the mod-
eling assumptions are correct or at least approximately correct. Logistic
regression makes weaker assumptions, and is significantly more robust to
deviations from modeling assumptions. Specifically, when the data is in-
deed non-Gaussian, then in the limit of large datasets, logistic regression will
almost always do better than GDA. For this reason, in practice logistic re-
gression is used more often than GDA. (Some related considerations about
discriminative vs. generative models also apply for the Naive Bayes algo-
rithm that we discuss next, but the Naive Bayes algorithm is still considered
a very good, and is certainly also a very popular, classification algorithm.)
 9

2 Naive Bayes
In GDA, the feature vectors x were continuous, real-valued vectors. Let’s
now talk about a different learning algorithm in which the xj ’s are discrete-
valued.
For our motivating example, consider building an email spam filter using
machine learning. Here, we wish to classify messages according to whether
they are unsolicited commercial (spam) email, or non-spam email. After
learning to do this, we can then have our mail reader automatically filter
out the spam messages and perhaps place them in a separate mail folder.
Classifying emails is one example of a broader set of problems called text
classification.
Let’s say we have a training set (a set of emails labeled as spam or non-
spam). We will cover two algorithms, the difference between which is mostly
the ways to represent an email as feature vectors. to represent an email.
Subsection 2.1 covers the so-called multi-variate Bernoulli event model, which
was not covered in the class, and can be considered as optional reading.
You could consider directly start reading from Section 2.3. To make the
subsections self-contained, Section 2.3 and Section 2.1 contains overlapping
materials.

2.1 Multi-variate Bernoulli event model (skipped in

Spring 2019 offering)
In this section, we will represent an email via a feature vector whose length
is equal to the number of words in the dictionary. Specifically, if an email
contains the j-th word of the dictionary, then we will set xj = 1; otherwise,
we let xj = 0. For instance, the vector
 
1 a

 0 
 aardvark
 0  aardwolf
 ..  ..
x= . .
 

1 buy
 
 
 ..  ..
 .  .
0 zygmurgy
 10

is used to represent an email that contains the words “a” and “buy,” but not
“aardvark,” “aardwolf” or “zygmurgy.”2 The set of words encoded into the
feature vector is called the vocabulary, so the dimension of x is equal to
the size of the vocabulary.
Having chosen our feature vector, we now want to build a generative
model. So, we have to model p(x|y). But if we have, say, a vocabulary of
50000 words, then x ∈ {0, 1}50000 (x is a 50000-dimensional vector of 0’s and
1’s), and if we were to model x explicitly with a multinomial distribution over
the 250000 possible outcomes, then we’d end up with a (250000 −1)-dimensional
parameter vector. This is clearly too many parameters.
To model p(x|y), we will therefore make a very strong assumption. We will
assume that the xi ’s are conditionally independent given y. This assumption
is called the Naive Bayes (NB) assumption, and the resulting algorithm is
called the Naive Bayes classifier. For instance, if y = 1 means spam email;
“buy” is word 2087 and “price” is word 39831; then we are assuming that if
I tell you y = 1 (that a particular piece of email is spam), then knowledge
of x2087 (knowledge of whether “buy” appears in the message) will have no
effect on your beliefs about the value of x39831 (whether “price” appears).
More formally, this can be written p(x2087 |y) = p(x2087 |y, x39831 ). (Note that
this is not the same as saying that x2087 and x39831 are independent, which
would have been written “p(x2087 ) = p(x2087 |x39831 )”; rather, we are only
assuming that x2087 and x39831 are conditionally independent given y.)
We now have:
p(x1 , . . . , x50000 |y)
= p(x1 |y)p(x2 |y, x1 )p(x3 |y, x1 , x2 ) · · · p(x50000 |y, x1 , . . . , x49999 )
= p(x1 |y)p(x2 |y)p(x3 |y) · · · p(x50000 |y)
Y d
= p(xj |y)
j=1

The first equality simply follows from the usual properties of probabilities,
2
Actually, rather than looking through an English dictionary for the list of all English
words, in practice it is more common to look through our training set and encode in our
feature vector only the words that occur at least once there. Apart from reducing the
number of words modeled and hence reducing our computational and space requirements,
this also has the advantage of allowing us to model/include as a feature many words
that may appear in your email (such as “cs229”) but that you won’t find in a dictionary.
Sometimes (as in the homework), we also exclude the very high frequency words (which
will be words like “the,” “of,” “and”; these high frequency, “content free” words are called
stop words) since they occur in so many documents and do little to indicate whether an
email is spam or non-spam.
 11

and the second equality used the NB assumption. We note that even though
the Naive Bayes assumption is an extremely strong assumptions, the resulting
algorithm works well on many problems.
Our model is parameterized by φj|y=1 = p(xj = 1|y = 1), φj|y=0 = p(xj =
1|y = 0), and φy = p(y = 1). As usual, given a training set {(x(i) , y (i) ); i =
1, . . . , n}, we can write down the joint likelihood of the data:
n
Y
L(φy , φj|y=0 , φj|y=1 ) = p(x(i) , y (i) ).
i=1

Maximizing this with respect to φy , φj|y=0 and φj|y=1 gives the maximum
likelihood estimates:
Pn (i) (i)
i=1 1{xj = 1 ∧ y = 1}
φj|y=1 = Pn (i) = 1}
i=1 1{y
Pn (i) (i)
i=1 1{xj = 1 ∧ y = 0}
φj|y=0 = Pn (i)
i=1 1{y = 0}
Pn (i)
i=1 1{y = 1}
φy =
n
In the equations above, the “∧” symbol means “and.” The parameters have
a very natural interpretation. For instance, φj|y=1 is just the fraction of the
spam (y = 1) emails in which word j does appear.
Having fit all these parameters, to make a prediction on a new example
with features x, we then simply calculate
p(x|y = 1)p(y = 1)
p(y = 1|x) =
p(x)
Q 
d
p(xj |y = 1) p(y = 1)
j=1
= Q  Q  ,
d d
j=1 p(xj |y = 1) p(y = 1) + j=1 p(xj |y = 0) p(y = 0)

and pick whichever class has the higher posterior probability.

Lastly, we note that while we have developed the Naive Bayes algorithm
mainly for the case of problems where the features xj are binary-valued, the
generalization to where xj can take values in {1, 2, . . . , kj } is straightforward.
Here, we would simply model p(xj |y) as multinomial rather than as Bernoulli.
Indeed, even if some original input attribute (say, the living area of a house,
as in our earlier example) were continuous valued, it is quite common to
discretize it—that is, turn it into a small set of discrete values—and apply
 12

Naive Bayes. For instance, if we use some feature xj to represent living area,
we might discretize the continuous values as follows:
Living area (sq. feet) < 400 400-800 800-1200 1200-1600 >1600
xi 1 2 3 4 5
Thus, for a house with living area 890 square feet, we would set the value
of the corresponding feature xj to 3. We can then apply the Naive Bayes
algorithm, and model p(xj |y) with a multinomial distribution, as described
previously. When the original, continuous-valued attributes are not well-
modeled by a multivariate normal distribution, discretizing the features and
using Naive Bayes (instead of GDA) will often result in a better classifier.

2.2 Laplace smoothing

This section depends on Section 2.1. If you skipped Section 2.1, then you
an read Section 2.3.1 where we introduce Laplace smoothing again in a way
that only depends on Section 2.3.
The Naive Bayes algorithm, as we have described, will work fairly well
for many problems. But there is a simple change that makes it work even
better, especially for text classification. Let’s briefly discuss a problem with
the algorithm in its current form, and then talk about how we can fix it.
Consider spam/email classification, and let’s suppose that, we are in the
year 20xx. After completing CS229 and having done excellent work on the
project, you decide around May 20xx to submit your work to the NeurlPS
conference for publication.3 Because you end up discussing the conference in
your emails, you also start getting messages with the word “neurips” in it.
But this is your first NeurIPS paper, and until now, you have not previously
seen any emails containing the word “neurips”; in particular “neurips” did
not ever appear in your training set of spam/non-spam emails. Assuming
that “neurips” was the 35000th word in the dictionary, your Naive Bayes
spam filter had picked its maximum likelihood estimates of the parameters
φ35000|y to be
Pn (i)
i=1 1{x35000 = 1 ∧ y (i) = 1}
φ35000|y=1 = Pn (i) = 1}
=0
i=1 1{y
Pn (i)
i=1 1{x35000 = 1 ∧ y (i) = 0}
φ35000|y=0 = Pn (i) = 0}
=0
i=1 1{y
3
NeurlPS is one of the top machine learning conferences. The deadline for submitting
a paper is typically in May-June.
 13

I.e., because it has never seen “neurips” before in either spam or non-spam
training examples, it thinks the probability of seeing it in either type of email
is zero. Hence, when trying to decide if one of these messages containing
“neurips” is spam, it calculates the class posterior probabilities, and obtains
Qd
j=1 p(xj |y = 1)p(y = 1)
p(y = 1|x) = Qd Qd
j=1 p(xj |y = 1)p(y = 1) + j=1 p(xj |y = 0)p(y = 0)
0
= .
0

This is because each of the terms “ dj=1 p(xj |y)” includes a term p(x35000 |y) =
Q
0 that is multiplied into it. Hence, our algorithm obtains 0/0, and doesn’t
know how to make a prediction.
Stating the problem more broadly, it is statistically a bad idea to esti-
mate the probability of some event to be zero just because you haven’t seen
it before in your finite training set. Take the problem of estimating the mean
of a multinomial random variable z taking values in {1, . . . , k}. We can pa-
rameterize our multinomial with φj = p(z = j). Given a set of n independent
observations {z (1) , . . . , z (n) }, the maximum likelihood estimates are given by
Pn
1{z (i) = j}
φj = i=1 .
n
As we saw previously, if we were to use these maximum likelihood estimates,
then some of the φj ’s might end up as zero, which was a problem. To avoid
this, we can use Laplace smoothing, which replaces the above estimate
with
1 + ni=1 1{z (i) = j}
P
φj = .
n+k
Here,
Pk we’ve added 1 to the numerator, and k to the denominator. Note that
j=1 φj = 1 still holds (check this yourself!), which is a desirable property
since the φj ’s are estimates for probabilities that we know must sum to 1.
Also, φj 6= 0 for all values of j, solving our problem of probabilities being
estimated as zero. Under certain (arguably quite strong) conditions, it can
be shown that the Laplace smoothing actually gives the optimal estimator
of the φj ’s.
Returning to our Naive Bayes classifier, with Laplace smoothing, we
 14

therefore obtain the following estimates of the parameters:

Pn (i)
1+ 1{x = 1 ∧ y (i) = 1}
φj|y=1 = Pn j
i=1
2 + i=1 1{y (i) = 1}
(i)
1 + ni=1 1{xj = 1 ∧ y (i) = 0}
P
φj|y=0 =
2 + ni=1 1{y (i) = 0}
P

(In practice, it usually doesn’t matter much whether we apply Laplace smooth-
ing to φy or not, since we will typically have a fair fraction each of spam and
non-spam messages, so φy will be a reasonable estimate of p(y = 1) and will
be quite far from 0 anyway.)

2.3 Multinomial event model for text classification

Let’s talk about another model that is specifically for text classification. To
make the subsection self-contained, we start from scratch again with the
basic concepts.
We consider a dictionary or vocabulary of words that are un-
der considerations. We assume all the sentences contain only words
in the vocabulary. For example, we may have a vocabulary V =
{a, aardvark, aardwolf, · · · , zygumrgy}. Given an email that consists of a
sequence of d words, we will represent the email by a vector x = (x1 , . . . , xd )
of dimension d; note that d can vary for different emails. Here xj is the
identity of the j-th word of the email — if the j-word of the email is the k-th
word in the vocabulary, then xj = k. For instance, if an email starts with
“A NeurIPS . . . ,” then x1 = 1 (“a” is the first word in the dictionary), and
x2 = 35000 (if “neurips” is the 35000th word in the dictionary).
Having chosen our feature vector, we now want to build a genera-
tive model. So, we have to model p(x|y). But if we have, say, a
vocabulary of 50000 words, and each document has 100 words. then
x ∈ {1, 2, · · · , 50000}100 (x is 100-dimensional vector of values from
{1, 2, · · · , 50000}), and if we were to model x explicitly with a multinomial
distribution over the 50000100 possible outcomes, then we’d end up with a
(50000100 − 1)-dimensional parameter vector. This is clearly too many pa-
rameters.
To model p(x|y), we will therefore make a very strong assumption. We will
assume that the xi ’s are conditionally independent given y. This assumption
is called the Naive Bayes (NB) assumption, and the resulting algorithm
is called the Naive Bayes classifier. For instance, if y = 1 means spam
 15

email; then we are assuming that if I tell you y = 1 (that a particular piece of
email is spam), then the knowledge of the first word doesn’t have any effect
on our belief about the second word (or the beliefs about any other words).
We now have:

p(x1 , . . . , xd |y)
= p(x1 |y)p(x2 |y, x1 )p(x3 |y, x1 , x2 ) · · · p(xd |y, x1 , . . . , xd ))
= p(x1 |y)p(x2 |y)p(x3 |y) · · · p(xd |y)
Y d
= p(xj |y)
j=1

The first equality simply follows from the usual properties of probabilities,
and the second equality used the NB assumption. We note that even though
the Naive Bayes assumption is an extremely strong assumptions, the resulting
algorithm works well on many problems.
The parameters for our model are φy = p(y = 1), φk|y=1 = p(xj = k|y =
1) (for any j) and φk|y=0 = p(xj = k|y = 0). Note that we have assumed that
p(xj |y) is the same for all values of j (i.e., that the distribution according
to which a word is generated does not depend on its position j within the
email).
If we are given a training set {(x(i) , y (i) ); i = 1, . . . , n} where x(i) =
(i) (i) (i)
(x1 , x2 , . . . , xdi ) (here, di is the number of words in the i-training example),
the likelihood of the data is given by
n
Y
L(φy , φk|y=0 , φk|y=1 ) = p(x(i) , y (i) )
i=1
n di
!
Y Y (i)
= p(xj |y; φk|y=0 , φk|y=1 ) p(y (i) ; φy ).
i=1 j=1

Maximizing this yields the maximum likelihood estimates of the parameters:

Pn Pdi (i) (i)
i=1 j=1 1{xj = k ∧ y = 1}
φk|y=1 = P n (i) = 1}d
i=1 1{y i
Pn Pdi (i) (i)
i=1 j=1 1{xj = k ∧ y = 0}
φk|y=0 = Pn (i) = 0}d
i=1 1{y i
Pn (i)
i=1 1{y = 1}
φy = .
n
 16

2.3.1 Laplace smoothing: a self-contained introduction

The Naive Bayes algorithm as we have described it will work fairly well
for many problems, but there is a simple change that makes it work much
better, especially for text classification. Let’s briefly discuss a problem with
the algorithm above in Section 2.3, and then talk about how we can fix it.
Consider the second word in the vocabulary, aardvark, and suppose that
it never appears in the training dataset. Then, the estimator in Section 2.3
will give
Pn Pdi (i) (i)
i=1 j=1 1{xj = 2 ∧ y = 1}
φ2|y=1 = P n (i) = 1}d
=0
i=1 1{y i
Pn Pdi (i) (i)
i=1 j=1 1{xj = 2 ∧ y = 0}
φ2|y=0 = Pn (i) = 0}d
=0
i=1 1{y i

Now suppose at prediction time, we have an email x that contains the

word aardvark as the first word. Towards using Bayes rule, we will first need
to compute the probability of seeing this email p(x). Since we’ve estimated
that the chance of seeing the word aardvark at any position conditioned
on y = 0 to be zero, the probability of seeing the whole email conditioned
on y = 0 is also zero because the naive bayes assumption. Similarly, the
probability of seeing the email is also zero. Now using the Bayes rule

Qd
j=1 p(xj |y = 1)p(y = 1)
p(y = 1|x) = Qd Qd
j=1 p(xj |y = 1)p(y = 1) + j=1 p(xj |y = 0)p(y = 0)
0
= .
0
This is because each of the terms “ dj=1 p(xj |y)” includes a term p(x1 =
Q
2|y) = 0 that is multiplied into it. Hence, our algorithm obtains 0/0, and
doesn’t know how to make a prediction.
Stating the problem more broadly, it is statistically a bad idea to esti-
mate the probability of some event to be zero just because you haven’t seen
it before in your finite training set. Take the problem of estimating the mean
of a multinomial random variable z taking values in {1, . . . , k}. We can pa-
rameterize our multinomial with φj = p(z = j). Given a set of n independent
observations {z (1) , . . . , z (n) }, the maximum likelihood estimates are given by
Pn
1{z (i) = j}
φj = i=1 .
n
 17

As we saw previously, if we were to use these maximum likelihood estimates,

then some of the φj ’s might end up as zero, which was a problem. To avoid
this, we can use Laplace smoothing, which replaces the above estimate
with
1 + ni=1 1{z (i) = j}
P
φj = .
n+k
Here,
Pk we’ve added 1 to the numerator, and k to the denominator. Note that
j=1 φj = 1 still holds (check this yourself!), which is a desirable property
since the φj ’s are estimates for probabilities that we know must sum to 1.
Also, φj 6= 0 for all values of j, solving our problem of probabilities being
estimated as zero. Under certain (arguably quite strong) conditions, it can
be shown that the Laplace smoothing actually gives the optimal estimator
of the φj ’s.
Returning to our Naive Bayes classifier, if we were to apply Laplace
smoothing (which is needed in practice for good performance) when esti-
mating φk|y=0 and φk|y=1 , we add 1 to the numerators and |V | to the denom-
inators, and obtain:
Pn Pdi (i) (i)
1+ i=1 j=1 1{xj = k ∧ y = 1}
φk|y=1 = Pn (i)
|V | + i=1 1{y = 1}di
Pn Pdi (i)
1 + i=1 j=1 1{xj = k ∧ y (i) = 0}
φk|y=0 = .
|V | + ni=1 1{y (i) = 0}di
P

While not necessarily the very best classification algorithm, the Naive Bayes
classifier often works surprisingly well. It is often also a very good “first thing
to try,” given its simplicity and ease of implementation.
 CS229 Supplemental Lecture notes
John Duchi

1 General loss functions

Building off of our interpretations of supervised learning as (1) choosing a
representation for our problem, (2) choosing a loss function, and (3) minimiz-
ing the loss, let us consider a slightly more general formulation for supervised
learning. In the supervised learning settings we have considered thus far, we
have input data x ∈ Rn and targets y from a space Y. In linear regression,
this corresponded to y ∈ R, that is, Y = R, for logistic regression and other
binary classification problems, we had y ∈ Y = {−1, 1}, and for multiclass
classification we had y ∈ Y = {1, 2, . . . , k} for some number k of classes.
For each of these problems, we made predictions based on θT x for some
vector θ, and we constructed a loss function L : R × Y → R, where L(θT x, y)
measures the loss we suffer for predicting θT x. For logistic regression, we use
the logistic loss
T
L(z, y) = log(1 + e−yz ) or L(θT x, y) = log(1 + e−yθ x ).
For linear regression we use the squared error
1 1
L(z, y) = (z − y)2 or L(θT x, y) = (θT x − y)2 .
2 2
For multiclass classification, we had a slight variant, where we let Θ =
[θ1 · · · θk ] for θi ∈ Rn , and used the loss L : Rk × {1, . . . , k} → R
k
! k
!
X X
L(z, y) = log exp(zi − zy ) or L(ΘT x, y) = log exp(xT (θi − θy )) ,
i=1 i=1

the idea being that we wish to have >θyT x θiT x

for all i 6= k. Given a training
(i) (i)
set of pairs {x , y }, choose θ by minimizing the empirical risk
m
1 X
J(θ) = L(θT x(i) , y (i) ). (1)
m i=1

1
2 The representer theorem
Let us consider a slight variant of choosing θ to minimize the risk (1). In
many situations—for reasons that we will study more later in the class—it
is useful to add regularization to the risk J. We add regularization for many
reasons: often, it makes problem (1) easier to solve numerically, and also
it can allow the θ we get out of minimizing the risk (1) able to generalize
better to unseen data. Generally, regularization is taken to be of the form
r(θ) = kθk or r(θ) = kθk2 for some norm k·k on Rn . The most common
choice is so-called ℓ2 -regularization, in which we choose
λ
r(θ) = kθk22 ,
2
√
where we recall that kθk2 = θT θ is the Euclidean norm, or length, of the
vector θ. This gives rise to the regularized risk, which is
m
1 X λ
Jλ (θ) = L(θT x(i) , y (i) ) + kθk22 . (2)
m i=1 2

Let us consider the structure of any θ that minimizes the risk (2). We
assume—as we often do—that for each fixed target value y ∈ Y, the function
L(z, y) is convex in z. (This is the case for linear regression and binary
and multiclass logistic regression, as well as a number of other losses we will
consider.) It turns out that under these assumptions, we may always write
the solutions to the problem (2) as a linear combination of the input variables
x(i) . More precisely, we have the following theorem, known as the representer
theorem.

Theorem 2.1. Suppose in the definition of the regularized risk (2) that λ ≥
0. Then there is a minimizer of the regularized risk (2) that can be written
m
X
θ= αi x(i)
i=1

for some real-valued weights αi .

Proof For intuition, we give a proof of the result in the case that L(z, y),
when viewed as a function of z, is differentiable and λ > 0. In Appendix A,

2
we present a more general statement of the theorem as well as a rigorous
proof.
∂
Let L′ (z, y) = ∂z L(z, y) denote the derivative of the loss with respect to
z. Then by the chain rule, we have the gradient identity
1
∇θ L(θT x, y) = L′ (θT x, y)x and ∇θ kθk22 = θ,
2
where ∇θ denotes taking a gradient with respect to θ. As the risk must have
0 gradient at all stationary points (including the minimizer), we can write
m
1 X ′ T (i) (i) (i)
∇Jλ (θ) = L (θ x , y )x + λθ = ~0.
m i=1

In particular, letting wi = L′ (θT x(i) , y (i) ), as L′ (θT x(i) , y (i) ) is a scalar (which
depends on θ, but no matter what θ is, wi is still a real number), we have
n
1X
θ=− wi x(i) .
λ i=1

Set αi = − wλi to get the result.

3 Nonlinear features and kernels

Based on the representer theorem, Theorem 2.1, we see that we can always
write the vector θ as a linear combination of the data {x(i) }m
i=1 . Importantly,
this means we can always make predictions
m
X
θ T x = xT θ = αi xT x(i) .
i=1

That is, in any learning algorithm, we may can replace all appearances of
P (i) T
θT x with m i=1 αi x x, and then minimize directly over α ∈ Rm .
Let us consider this idea in somewhat more generality. In our discussion
of linear regression, we had a problem in which the input x was the living
area of a house, and we considered performing regression using the features x,
x2 and x3 (say) to obtain a cubic function. To distinguish between these two

3
sets of variables, we’ll call the “original” input value the input attributes
of a problem (in this case, x, the living area). When that is mapped to
some new set of quantities that are then passed to the learning algorithm,
we’ll call those new quantities the input features. (Unfortunately, different
authors use different terms to describe these two things, but we’ll try to use
this terminology consistently in these notes.) We will also let φ denote the
feature mapping, which maps from the attributes to the features. For
instance, in our example, we had
 
x
φ(x) = x2  .
x3

Rather than applying a learning algorithm using the original input at-
tributes x, we may instead want to learn using some features φ(x). To do so,
we simply need to go over our previous algorithm, and replace x everywhere
in it with φ(x).
Since the algorithm can be written entirely in terms of the inner prod-
ucts hx, zi, this means that we would replace all those inner products with
hφ(x), φ(z)i. Specificically, given a feature mapping φ, we define the corre-
sponding kernel to be
K(x, z) = φ(x)T φ(z).
Then, everywhere we previously had hx, zi in our algorithm, we could simply
replace it with K(x, z), and our algorithm would now be learning using the
features φ. Let us write this out more carefully.
P We saw(i)by the representer
theorem (Theorem 2.1) that we can write θ = m i=1 αi φ(x ) for some weights
αi . Then we can write the (regularized) risk

Jλ (θ) = Jλ (α)
m  m  m 2
1 X (i) T
X
(j) (i) λ X
= L φ(x ) αj φ(x ), y + αi φ(x(i) )
m i=1 j=1
2 i=1 2
m  m  m m
1 X X λ XX
= L αj φ(x(i) )T φ(x(j) ), y (i) + αi αj φ(x(i) )T φ(x(j) )
m i=1 j=1
2 i=1 j=1
m  m 
1 X X λX
= L αj K(x(i) , x(j) ), y (i) + αi αi K(x(i) , x(j) ).
m i=1 j=1
2 i,j

4
That is, we can write the entire loss function to be minimized in terms of the
kernel matrix
K = [K(x(i) , x(j) )]m
i,j=1 ∈ R
m×m
.
Now, given φ, we could easily compute K(x, z) by finding φ(x) and φ(z)
and taking their inner product. But what’s more interesting is that often,
K(x, z) may be very inexpensive to calculate, even though φ(x) itself may be
very expensive to calculate (perhaps because it is an extremely high dimen-
sional vector). In such settings, by using in our algorithm an efficient way to
calculate K(x, z), we can learn in the high dimensional feature space space
given by φ, but without ever having to explicitly find or represent vectors
φ(x). As a few examples, some kernels (corresponding to infinite-dimensional
vectors φ) include
 
1 2
K(x, z) = exp − 2 kx − zk2 ,
2τ
known as the Gaussian or Radial Basis Function (RBF) kernel and applicable
to data in any dimension, or the min-kernel (applicable when x ∈ R, defined
by
K(x, z) = min{x, z}.
See also the lecture notes on Support Vector Machines (SVMs) for more on
these kernel machines.

4 Stochastic gradient descent for kernelized

machine learning
If we let K ∈ Rm×m denote the kernel matrix, and for shorthand define the
vectors  
K(x(i) , x(1) )
 K(x(i) , x(2) ) 
K (i) = 
 
.. ,
 . 
(i) (m)
K(x , x )
so that K = [K (1) K (2) · · · K (m) ], then we may write the regularized risk in
a consise form as
m
1 X T λ
Jλ (α) = L(K (i) α, y (i) ) + αT Kα.
m i=1 2

5
Now, let us consider taking a stochastic gradient of the above risk Jλ . That is,
we wish to construct a (simple to compute) random vector whose expectation
is ∇Jλ (α), which does not have too much variance. To do so, we first compute
the gradient of Jλ (α) on its own. We calculate the gradient of individual loss
terms by noting that
T T
∇α L(K (i) α, y (i) ) = L′ (K (i) α, y (i) )K (i) ,
while   m
λ T X
∇α α Kα = λKα = λ K (i) αi .
2 i=1
Thus, we have
m m
1 X ′ (i) T X
∇α Jλ (α) = L (K α, y (i) )K (i) + λ K (i) αi .
m i=1 i=1

Thus, if we choose a random index i ∈ {1, . . . , m}, we have that

T
L′ (K (i) α, y (i) )K (i) + mλK (i) αi
is a stochastic gradient for Jλ (α). This gives us a stochastic gradient proce-
dure for Kernel supervised learning problems, shown in Figure 1. One minor

Input: A loss function L, kernel matrix K = [K (1) · · · K (m) ], and labels

{y (i) }m
i=1 , and sequence of positive stepsizes η1 , η2 , η3 , . . .

Iterate for t = 1, 2, . . .

(i) Choose index i ∈ {1, . . . , m} uniformly at random

(ii) Update
h i
′ (i) T (i) (i) (i)
α := α − ηt L (K α, y )K + mλK αi .

Figure 1: Stochastic gradient descent for kernel supervised learning problems.

remark is in order regarding Algorithm 1: because we multiply the λK (i) αi

term by m to keep the gradient unbiased, it is important that λ > 0 not be
too large, as the algorithm can be a bit unstable
√ otherwise. In addition, a
common choice of stepsize is to use ηt = 1/ t, or a constant multiple thereof.

6
5 Support vector machines
Now we discuss (one approach) to Support Vector Machines (SVM)s, which
apply to binary classification problems with labels y ∈ {−1, 1}, and a partic-
ular choice of loss function L. In particular, for the support vector machine,
we use the margin-based loss function

L(z, y) = [1 − yz]+ = max{0, 1 − yz}. (3)

So, in a sense, SVMs are nothing but a special case of the general theoret-
ical results we have described above. In particular, we have the empirical
regularized risk
m
1 Xh T
i λ
Jλ (α) = 1 − y (i) K (i) α + αT Kα,
m i=1 + 2

where the matrix K = [K (1) · · · K (m) ] is defined by Kij = K(x(i) , x(j) ).

In the lecture notes, you can see another way of deriving the support
vector machine and a description of why we actually call them support vector
machines.

6 An example
In this section, we consider a particular example kernel, known as the Gaus-
sian or Radial Basis Function (RBF) kernel. This kernel is defined by
 
1 2
K(x, z) = exp − 2 kx − zk2 , (4)
2τ

where τ > 0 is a parameter controlling the bandwidth of the kernel. Intu-

itively, for τ very small, we will have K(x, z) ≈ 0 unless x ≈ z, that is, x and
z are very close, in which case we have K(x, z) ≈ 1. However, for τ large,
then we have a much smoother kernel function K. The feature function φ
for this kernel is infinite dimensional.1 That said, it is possible to gain some
1
If you have seen characteristic functions or Fourier transforms, then you might rec-
ognize the RBF kernel as the Fourier transform of the Gaussian distribution with mean
zero and variance τ 2 . That is, in Rn , let W ∼ N(0, τ 2 In×n ), so that W has density

7
1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

τ = .1 τ = .2

Figure 2: Small bandwidths τ for Gaussian kernel K(x, z) =

exp(− 2τ12 kx − zk22 ).

intuition for the kernel by considering the classifications it makes on a new

example x: we have prediction
m m  
X
(i)
X 1 (i) 2
K(x , x)αi = exp − 2 x − x 2 αi ,
i=1 i=1
2τ
kwk2 √
p(w) = (2πτ12 )n/2 exp(− 2τ 22 ). Let i = −1 be the imaginary unit, then for any vector v
we have
Z Z  
1 1 2
E[exp(iv T W )] = exp(iv T w)p(w)dw = exp iv T
w − kwk 2 dw
(2πτ 2 )n/2 2τ 2
 
1 2
= exp − 2 kvk2 .
2τ
T
Thus, if we define the “vector” (really, function) φ(x, w) = eix w
and let a∗ be the complex
conjugate of a ∈ C, then we have
 
∗ ixT W −ixT W T 1 2
E[φ(x, W )φ(z, W ) ] = E[e e ] = E[exp(iW (x − z))] = exp − 2 kx − zk2 .
2τ

In particular, we see that K(x, z) is the inner product in a space of functions that are
integrable against p(w).

8
1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

τ = .4 τ = .8

Figure 3: Medium bandwidths τ for Gaussian kernel K(x, z) =

exp(− 2τ12 kx − zk22 ).

so that this becomes something like a weighting depending on how close x

is to each x(i) —that is, the contribution of weight αi is multiplied by the
similarity of x to x(i) as determined by the kernel function.
In Figures 2, 3, and 4, we show the results of training 6 different kernel
classifiers by minimizing
m h i
X T λ
Jλ (α) = 1 − y (i) K (i) α + αT Kα,
i=1
+ 2

with m = 200 and λ = 1/m, for different values of τ in the kernel (4). We
plot the training data (positive examples as blue x’s and negative examples
as red o’s) as well as the decision surface of the resulting classifier. That is,
we plot the lines defined by
( m
)
X
x ∈ R2 : K(x, x(i) )αi = 0 ,
i=1
P
the regions where the learned classifier makes a prediction m
givingP (i)
i=1 K(x, x )αi >
m (i)
0 and i=1 K(x, x )αi < 0. From the figure, we see that for large τ , we have

9
1.4 1.4

1.2 1.2

1 1

0.8 0.8

0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

τ = 1.6 τ = 3.2

Figure 4: Large bandwidths τ for Gaussian kernel K(x, z) =

exp(− 2τ 2 kx − zk22 ).
1

a very simple classifier: it is nearly linear, while for τ = .1, the classifier has
substantial variability and is highly non-linear. For reference, in Figure 5, we
plot the optimal classifier along with the training data; the optimal classifier
minimizes the misclassification error given infinite training data.

A A more general representer theorem

In this section, we present a more general version of the representer theorem
along with a rigorous proof. Let r : R → R be any non-decreasing function
of its argument, and consider the regularized risk
m
1 X T
Jr (θ) = L(x(i) θ, y (i) ) + r(kθk2 ). (5)
m i=1

Often, we take r(t) = λ2 t2 , which corresponds to the common choice of ℓ2 -

regularization, but the next theorem makes clear that this is not necessary
for the representer theorem. Indeed, we could simply take r(t) = 0 for all t,
and the theorem still holds.

10
 1.4

1.2

0.8

0.6

0.4

0.2

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1

Figure 5: Optimal classifier along with training data.

in Rn ). Let θ ∈ Rn be any vector. Then

Theorem A.1 (Representer theorem P
there exists some α ∈ R and θ = m
m (α)
i=1 αi x
(i)
such that

Jr (θ(α) ) ≤ Jr (θ).

In particular, there is no loss of generality in always assuming we can write

the optimization problem to minimize J(θ) by only considering θ in the span
of the data.
Proof Our proof relies on some machinery from linear algebra, which
allows us to keep it concise, but feel free to ask questions if it is too concise.
The vectors {x(i) }m n
i=1 are in R , and as a consequence there is some sub-
n
space V of R such that
X
m 
(i)
V = βi x : βi ∈ R .
i=1

Then V has an orthonormal basis {v1 , . . . , vn0 } for vectors vi ∈ Rn ,P

where the
size (dimension) of the basis is n0 ≤ n. Thus we can write V = { ni=1 0
bi v i :
bi ∈ R}, where we recall that orthonormality means that the vectors vi
satisfy kvi k2 = 1 and viT vj = 0 for i 6= j. There is also an orthogonal
subspace V ⊥ = {u ∈ Rn : uT v = 0 for all v ∈ V }, which has an orthonormal

11
basis of size n⊥ = n − n0 ≥ 0, which we write as {u1 , . . . , un⊥ } ⊂ Rn . By
construction they satisfy uTi vj = 0 for all i, j.
Because θ ∈ Rn , we know that we can write it uniquely as
n0
X n⊥
X
θ= ν i vi + µi ui , where νi ∈ R and µi ∈ R,
i=1 i=1

and the values µ, ν are unique. Now, we know that by definition of the space
V as the span of {x(i) }m m
i=1 , there exists α ∈ R such that
n0
X m
X
ν i vi = αi x(i) ,
i=1 i=1

so that we have m n⊥
X X
(i)
θ= αi x + µ i ui .
i=1 i=1
(α)
Pm
Define θ = i=1 αi x . Now, for any data point x(j) , we have
(i)

uTi x(j) = 0 for all i = 1, . . . , n⊥ ,

so that uTi θ(α) = 0. As a consquence, we have
n⊥
X 2 n⊥
X n⊥
X 2
(α) 2 2
kθk22 = θ (α)
+ µ i ui = θ 2
+2 µi uTi θ(α) + µ i ui ≥ θ(α) 2
,
i=1 2 i=1 i=1 2
| {z }
=0
(6a)
and we also have
T
θ(α) x(i) = θT x(i) (6b)
for all points x(i) .
That is, by using kθk2 ≥ θ(α) 2
and equality (6b), we have
m m
1 X (6b) 1 X T
Jr (θ) = L(θT x(i) , y (i) ) + r(kθk2 ) = L(θ(α) x(i) , y (i) ) + r(kθk2 )
m i=1 m i=1
m
(6a) 1 X T
≥ L(θ(α) x(i) , y (i) ) + r( θ(α) 2 )
m i=1
= Jr (θ(α) ).
This is the desired result.

12
 CS229 Lecture Notes
Andrew Ng
updated by Tengyu Ma on April 21, 2019

Part V
Kernel Methods
1.1 Feature maps
Recall that in our discussion about linear regression, we considered the prob-
lem of predicting the price of a house (denoted by y) from the living area of
the house (denoted by x), and we fit a linear function of x to the training
data. What if the price y can be more accurately represented as a non-linear
function of x? In this case, we need a more expressive family of models than
linear models.
We start by considering fitting cubic functions y = θ3 x3 + θ2 x2 + θ1 x + θ0 .
It turns out that we can view the cubic function as a linear function over
the a different set of feature variables (defined below). Concretely, let the
function φ : R → R4 be defined as

 
1
 x  4
 x2  ∈ R .
φ(x) =  (1)


x3

Let θ ∈ R4 be the vector containing θ0 , θ1 , θ2 , θ3 as entries. Then we can

rewrite the cubic function in x as:

θ3 x3 + θ2 x2 + θ1 x + θ0 = θT φ(x)

Thus, a cubic function of the variable x can be viewed as a linear function

over the variables φ(x). To distinguish between these two sets of variables,

1
 2

in the context of kernel methods, we will call the “original” input value the
input attributes of a problem (in this case, x, the living area). When the
original input is mapped to some new set of quantities φ(x), we will call those
new quantities the features variables. (Unfortunately, different authors use
different terms to describe these two things in different contexts.) We will
call φ a feature map, which maps the attributes to the features.

1.2 LMS (least mean squares) with features

We will derive the gradient descent algorithm for fitting the model θT φ(x).
First recall that for ordinary least square problem where we were to fit θT x,
the batch gradient descent update is (see the first lecture note for its deriva-
tion):
n
X
y (i) − hθ (x(i) ) x(i)


θ := θ + α
i=1
Xn
y (i) − θT x(i) x(i) .


:= θ + α (2)
i=1

Let φ : Rd → Rp be a feature map that maps attribute x (in Rd ) to the

features φ(x) in Rp . (In the motivating example in the previous subsection,
we have d = 1 and p = 4.) Now our goal is to fit the function θT φ(x), with
θ being a vector in Rp instead of Rd . We can replace all the occurrences of
x(i) in the algorithm above by φ(x(i) ) to obtain the new update:

n
X
y (i) − θT φ(x(i) ) φ(x(i) )


θ := θ + α (3)
i=1

Similarly, the corresponding stochastic gradient descent update rule is

θ := θ + α y (i) − θT φ(x(i) ) φ(x(i) )

(4)

1.3 LMS with the kernel trick

The gradient descent update, or stochastic gradient update above becomes
computationally expensive when the features φ(x) is high-dimensional. For
example, consider the direct extension of the feature map in equation (1) to
 3

high-dimensional input x: suppose x ∈ Rd , and let φ(x) be the vector that

contains all the monomials of x with degree ≤ 3
 
1
 x1 
 
 x2 
 .. 
 
 . 
 2 
 x1 
 
 x1 x2 
 
φ(x) =  x1 x3  . (5)
 
 .. 
 . 
 x2 x1 
 
 . 
 .. 
 
 x3 
 1 
 x2 x 
 1 2 
..
.

The dimension of the features φ(x) is on the order of d3 .1 This is a pro-

hibitively long vector for computational purpose — when d = 1000, each
update requires at least computing and storing a 10003 = 109 dimensional
vector, which is 106 times slower than the update rule for for ordinary least
squares updates (2).
It may appear at first that such d3 runtime per update and memory usage
are inevitable, because the vector θ itself is of dimension p ≈ d3 , and we may
need to update every entry of θ and store it. However, we will introduce the
kernel trick with which we will not need to store θ explicitly, and the runtime
can be significantly improved.
For simplicity, we assume the initialize the value θ = 0, and we focus
on the iterative update (3). The main observation is that at any time, θ
can be represented as a linear combination of the vectors φ(x(1) ), . . . , φ(x(n) ).
Indeed,
Pn we can show this inductively as follows. At initialization, θ = 0 =
(i)
i=1 0 · φ(x ). Assume at some point, θ can be represented as

n
X
θ= βi φ(x(i) ) (6)
i=1

1
Here, for simplicity, we include all the monomials with repetitions (so that, e.g., x1 x2 x3
and x2 x3 x1 both appear in φ(x)). Therefore, there are totally 1 + d + d2 + d3 entries in
φ(x).
 4

for some β1 , . . . , βn ∈ R. Then we claim that in the next round, θ is still a

linear combination of φ(x(1) ), . . . , φ(x(n) ) because
n
X
y (i) − θT φ(x(i) ) φ(x(i) )


θ := θ + α
i=1
n
X n
X
(i)
y (i) − θT φ(x(i) ) φ(x(i) )


= βi φ(x ) + α
i=1 i=1
n
X
(βi + α y (i) − θT φ(x(i) ) ) φ(x(i) )


= (7)
i=1
| {z }
new βi

You may realize that our general strategy is to implicitly represent the p-
dimensional vector θ by a set of coefficients β1 , . . . , βn . Towards doing this,
we derive the update rule of the coefficients β1 , . . . , βn . Using the equation
above, we see that the new βi depends on the old one via

βi := βi + α y (i) − θT φ(x(i) )


(8)

HerePwe still have the old θ on the RHS of the equation. Replacing θ by
θ = nj=1 βj φ(x(j) ) gives

n
!
(j) T
X
∀i ∈ {1, . . . , n}, βi := βi + α y (i) − βj φ(x ) φ(x(i) )
j=1

T
We often rewrite φ(x(j) ) φ(x(i) ) as hφ(x(j) ), φ(x(i) )i to emphasize that it’s the
inner product of the two feature vectors. Viewing βi ’s as the new representa-
tion of θ, we have successfully translated the batch gradient descent algorithm
into an algorithm that updates the value of β iteratively. It may appear that
at every iteration, we still need to compute the values of hφ(x(j) ), φ(x(i) )i for
all pairs of i, j, each of which may take roughly O(p) operation. However,
two important properties come to rescue:

1. We can pre-compute the pairwise inner products hφ(x(j) ), φ(x(i) )i for all
pairs of i, j before the loop starts.

2. For the feature map φ defined in (5) (or many other interesting fea-
ture maps), computing hφ(x(j) ), φ(x(i) )i can be efficient and does not
 5

necessarily require computing φ(x(i) ) explicitly. This is because:

d
X X X
hφ(x), φ(z)i = 1 + xi zi + xi xj zi zj + xi xj xk zi zj zk
i=1 i,j∈{1,...,d} i,j,k∈{1,...,d}
d d
!2 d
!3
X X X
=1+ xi zi + xi zi + xi zi
i=1 i=1 i=1
= 1 + hx, zi + hx, zi + hx, zi3
2
(9)

Therefore, to compute hφ(x), φ(z)i, we can first compute hx, zi with

O(d) time and then take another constant number of operations to com-
pute 1 + hx, zi + hx, zi2 + hx, zi3 .

As you will see, the inner products between the features hφ(x), φ(z)i are
essential here. We define the Kernel corresponding to the feature map φ as
a function that maps X × X → R satisfying: 2

K(x, z) , hφ(x), φ(z)i (10)

To wrap up the discussion, we write the down the final algorithm as

follows:

1. Compute all the values K(x(i) , x(j) ) , hφ(x(i) ), φ(x(j) )i using equa-
tion (9) for all i, j ∈ {1, . . . , n}. Set β := 0.

2. Loop:
n
!
X
(i) (i) (j)
∀i ∈ {1, . . . , n}, βi := βi + α y − βj K(x , x ) (11)
j=1

Or in vector notation, letting K be the n × n matrix with Kij =

K(x , x(j) ), we have
(i)

β := β + α(~y − Kβ)

With the algorithm above, we can update the representation β of the

vector θ efficiently with O(n) time per update. Finally, we need to show that
2
Recall that X is the space of the input x. In our running example, X = Rd
 6

the knowledge of the representation β suffices to compute the prediction

θT φ(x). Indeed, we have
n n
X T X
θT φ(x) = βi φ(x(i) ) φ(x) = βi K(x(i) , x) (12)
i=1 i=1

You may realize that fundamentally all we need to know about the feature
map φ(·) is encapsulated in the corresponding kernel function K(·, ·). We
will expand on this in the next section.

1.4 Properties of kernels

In the last subsection, we started with an explicitly defined feature map φ,
which induces the kernel function K(x, z) , hφ(x), φ(z)i. Then we saw that
the kernel function is so intrinsic so that as long as the kernel function is
defined, the whole training algorithm can be written entirely in the language
of the kernel without referring to the feature map φ, so can the prediction of
a test example x (equation (12).)
Therefore, it would be tempted to define other kernel function K(·, ·)
and run the algorithm (11). Note that the algorithm (11) does not need to
explicitly access the feature map φ, and therefore we only need to ensure the
existence of the feature map φ, but do not necessarily need to be able to
explicitly write φ down.
What kinds of functions K(·, ·) can correspond to some feature map φ? In
other words, can we tell if there is some feature mapping φ so that K(x, z) =
φ(x)T φ(z) for all x, z?
If we can answer this question by giving a precise characterization of valid
kernel functions, then we can completely change the interface of selecting
feature maps φ to the interface of selecting kernel function K. Concretely,
we can pick a function K, verify that it satisfies the characterization (so
that there exists a feature map φ that K corresponds to), and then we can
run update rule (11). The benefit here is that we don’t have to be able
to compute φ or write it down analytically, and we only need to know its
existence. We will answer this question at the end of this subsection after
we go through several concrete examples of kernels.
Suppose x, z ∈ Rd , and let’s first consider the function K(·, ·) defined as:

K(x, z) = (xT z)2 .

 7

We can also write this as

d
! d
!
X X
K(x, z) = xi zi xj zj
i=1 j=1
d X
X d
= xi xj zi zj
i=1 j=1
d
X
= (xi xj )(zi zj )
i,j=1

Thus, we see that K(x, z) = hφ(x), φ(z)i is the kernel function that corre-
sponds to the the feature mapping φ given (shown here for the case of d = 3)
by  
x1 x1
 x1 x2 
 
 x1 x3 
 
 x2 x1 
 
φ(x) =  x2 x2  .

 x2 x3 
 
 x3 x1 
 
 x3 x2 
x3 x3
Revisiting the computational efficiency perspective of kernel, note that whereas
calculating the high-dimensional φ(x) requires O(d2 ) time, finding K(x, z)
takes only O(d) time—linear in the dimension of the input attributes.
For another related example, also consider K(·, ·) defined by

K(x, z) = (xT z + c)2

d d
X X √ √
= (xi xj )(zi zj ) + ( 2cxi )( 2czi ) + c2 .
i,j=1 i=1

(Check this yourself.) This function K is a kernel function that corresponds

 8

to the feature mapping (again shown for d = 3)

 
x1 x1
 x1 x2 
 
 x1 x3 
 
 x2 x1 
 
 x2 x2 
 
 x2 x3 
 
 x3 x1  ,
φ(x) =  
 x3 x2 
 
 x3 x3 
 √ 
 2cx1 
 √ 
 2cx2 
 √ 
 2cx3 
c

and the parameter c controls the relative weighting between the xi (first
order) and the xi xj (second order) terms.
More broadly, the kernel K(x, z) = (xT z + c)k corresponds to a feature
mapping to an d+k


k
feature space, corresponding of all monomials of the
form xi1 xi2 . . . xik that are up to order k. However, despite working in this
O(dk )-dimensional space, computing K(x, z) still takes only O(d) time, and
hence we never need to explicitly represent feature vectors in this very high
dimensional feature space.

Kernels as similarity metrics. Now, let’s talk about a slightly different

view of kernels. Intuitively, (and there are things wrong with this intuition,
but nevermind), if φ(x) and φ(z) are close together, then we might expect
K(x, z) = φ(x)T φ(z) to be large. Conversely, if φ(x) and φ(z) are far apart—
say nearly orthogonal to each other—then K(x, z) = φ(x)T φ(z) will be small.
So, we can think of K(x, z) as some measurement of how similar are φ(x)
and φ(z), or of how similar are x and z.
Given this intuition, suppose that for some learning problem that you’re
working on, you’ve come up with some function K(x, z) that you think might
be a reasonable measure of how similar x and z are. For instance, perhaps
you chose
||x − z||2
 
K(x, z) = exp − .
2σ 2
This is a reasonable measure of x and z’s similarity, and is close to 1 when
x and z are close, and near 0 when x and z are far apart. Does there exist
 9

a feature map φ such that the kernel K defined above satisfies K(x, z) =
φ(x)T φ(z)? In this particular example, the answer is yes. This kernel is called
the Gaussian kernel, and corresponds to an infinite dimensional feature
mapping φ. We will give a precise characterization about what properties
a function K needs to satisfy so that it can be a valid kernel function that
corresponds to some feature map φ.

Necessary conditions for valid kernels. Suppose for now that K is

indeed a valid kernel corresponding to some feature mapping φ, and we will
first see what properties it satisfies. Now, consider some finite set of n points
(not necessarily the training set) {x(1) , . . . , x(n) }, and let a square, n-by-n
matrix K be defined so that its (i, j)-entry is given by Kij = K(x(i) , x(j) ).
This matrix is called the kernel matrix. Note that we’ve overloaded the
notation and used K to denote both the kernel function K(x, z) and the
kernel matrix K, due to their obvious close relationship.
Now, if K is a valid kernel, then Kij = K(x(i) , x(j) ) = φ(x(i) )T φ(x(j) ) =
φ(x(j) )T φ(x(i) ) = K(x(j) , x(i) ) = Kji , and hence K must be symmetric. More-
over, letting φk (x) denote the k-th coordinate of the vector φ(x), we find that
for any vector z, we have
XX
z T Kz = zi Kij zj
i j
XX
= zi φ(x(i) )T φ(x(j) )zj
i j
XX X
= zi φk (x(i) )φk (x(j) )zj
i j k
XXX
= zi φk (x(i) )φk (x(j) )zj
k i j
!2
X X
= zi φk (x(i) )
k i
≥ 0.
2
P P
The second-to-last step uses the fact that i,j ai aj = ( i ai ) for ai =
zi φk (x(i) ). Since z was arbitrary, this shows that K is positive semi-definite
(K ≥ 0).
Hence, we’ve shown that if K is a valid kernel (i.e., if it corresponds to
some feature mapping φ), then the corresponding kernel matrix K ∈ Rn×n
is symmetric positive semidefinite.
 10

Sufficient conditions for valid kernels. More generally, the condition

above turns out to be not only a necessary, but also a sufficient, condition
for K to be a valid kernel (also called a Mercer kernel). The following result
is due to Mercer.3

Theorem (Mercer). Let K : Rd × Rd 7→ R be given. Then for K

to be a valid (Mercer) kernel, it is necessary and sufficient that for any
{x(1) , . . . , x(n) }, (n < ∞), the corresponding kernel matrix is symmetric pos-
itive semi-definite.

Given a function K, apart from trying to find a feature mapping φ that

corresponds to it, this theorem therefore gives another way of testing if it is
a valid kernel. You’ll also have a chance to play with these ideas more in
problem set 2.
In class, we also briefly talked about a couple of other examples of ker-
nels. For instance, consider the digit recognition problem, in which given
an image (16x16 pixels) of a handwritten digit (0-9), we have to figure out
which digit it was. Using either a simple polynomial kernel K(x, z) = (xT z)k
or the Gaussian kernel, SVMs were able to obtain extremely good perfor-
mance on this problem. This was particularly surprising since the input
attributes x were just 256-dimensional vectors of the image pixel intensity
values, and the system had no prior knowledge about vision, or even about
which pixels are adjacent to which other ones. Another example that we
briefly talked about in lecture was that if the objects x that we are trying
to classify are strings (say, x is a list of amino acids, which strung together
form a protein), then it seems hard to construct a reasonable, “small” set of
features for most learning algorithms, especially if different strings have dif-
ferent lengths. However, consider letting φ(x) be a feature vector that counts
the number of occurrences of each length-k substring in x. If we’re consid-
ering strings of English letters, then there are 26k such strings. Hence, φ(x)
is a 26k dimensional vector; even for moderate values of k, this is probably
too big for us to efficiently work with. (e.g., 264 ≈ 460000.) However, using
(dynamic programming-ish) string matching algorithms, it is possible to ef-
ficiently compute K(x, z) = φ(x)T φ(z), so that we can now implicitly work
in this 26k -dimensional feature space, but without ever explicitly computing
feature vectors in this space.
3
Many texts present Mercer’s theorem in a slightly more complicated form involving
L2 functions, but when the input attributes take values in Rd , the version given here is
equivalent.
 11

Application of kernel methods: We’ve seen the application of kernels

to linear regression. In the next part, we will introduce the support vector
machines to which kernels can be directly applied. dwell too much longer on
it here. In fact, the idea of kernels has significantly broader applicability than
linear regression and SVMs. Specifically, if you have any learning algorithm
that you can write in terms of only inner products hx, zi between input
attribute vectors, then by replacing this with K(x, z) where K is a kernel,
you can “magically” allow your algorithm to work efficiently in the high
dimensional feature space corresponding to K. For instance, this kernel trick
can be applied with the perceptron to derive a kernel perceptron algorithm.
Many of the algorithms that we’ll see later in this class will also be amenable
to this method, which has come to be known as the “kernel trick.”

Part VI
Support Vector Machines
This set of notes presents the Support Vector Machine (SVM) learning al-
gorithm. SVMs are among the best (and many believe are indeed the best)
“off-the-shelf” supervised learning algorithms. To tell the SVM story, we’ll
need to first talk about margins and the idea of separating data with a large
“gap.” Next, we’ll talk about the optimal margin classifier, which will lead
us into a digression on Lagrange duality. We’ll also see kernels, which give
a way to apply SVMs efficiently in very high dimensional (such as infinite-
dimensional) feature spaces, and finally, we’ll close off the story with the
SMO algorithm, which gives an efficient implementation of SVMs.

2 Margins: Intuition
We’ll start our story on SVMs by talking about margins. This section will
give the intuitions about margins and about the “confidence” of our predic-
tions; these ideas will be made formal in Section 4.
Consider logistic regression, where the probability p(y = 1|x; θ) is mod-
eled by hθ (x) = g(θT x). We then predict “1” on an input x if and only if
hθ (x) ≥ 0.5, or equivalently, if and only if θT x ≥ 0. Consider a positive
training example (y = 1). The larger θT x is, the larger also is hθ (x) = p(y =
1|x; θ), and thus also the higher our degree of “confidence” that the label is 1.
Thus, informally we can think of our prediction as being very confident that
 12

y = 1 if θT x  0. Similarly, we think of logistic regression as confidently

predicting y = 0, if θT x  0. Given a training set, again informally it seems
that we’d have found a good fit to the training data if we can find θ so that
θT x(i)  0 whenever y (i) = 1, and θT x(i)  0 whenever y (i) = 0, since this
would reflect a very confident (and correct) set of classifications for all the
training examples. This seems to be a nice goal to aim for, and we’ll soon
formalize this idea using the notion of functional margins.
For a different type of intuition, consider the following figure, in which x’s
represent positive training examples, o’s denote negative training examples,
a decision boundary (this is the line given by the equation θT x = 0, and
is also called the separating hyperplane) is also shown, and three points
have also been labeled A, B and C.

A0
1

B0
1
C0
1

Notice that the point A is very far from the decision boundary. If we are
asked to make a prediction for the value of y at A, it seems we should be
quite confident that y = 1 there. Conversely, the point C is very close to
the decision boundary, and while it’s on the side of the decision boundary
on which we would predict y = 1, it seems likely that just a small change to
the decision boundary could easily have caused out prediction to be y = 0.
Hence, we’re much more confident about our prediction at A than at C. The
point B lies in-between these two cases, and more broadly, we see that if
a point is far from the separating hyperplane, then we may be significantly
more confident in our predictions. Again, informally we think it would be
nice if, given a training set, we manage to find a decision boundary that
allows us to make all correct and confident (meaning far from the decision
boundary) predictions on the training examples. We’ll formalize this later
using the notion of geometric margins.
 13

3 Notation
To make our discussion of SVMs easier, we’ll first need to introduce a new
notation for talking about classification. We will be considering a linear
classifier for a binary classification problem with labels y and features x.
From now, we’ll use y ∈ {−1, 1} (instead of {0, 1}) to denote the class labels.
Also, rather than parameterizing our linear classifier with the vector θ, we
will use parameters w, b, and write our classifier as

hw,b (x) = g(wT x + b).

Here, g(z) = 1 if z ≥ 0, and g(z) = −1 otherwise. This “w, b” notation

allows us to explicitly treat the intercept term b separately from the other
parameters. (We also drop the convention we had previously of letting x0 = 1
be an extra coordinate in the input feature vector.) Thus, b takes the role of
what was previously θ0 , and w takes the role of [θ1 . . . θd ]T .
Note also that, from our definition of g above, our classifier will directly
predict either 1 or −1 (cf. the perceptron algorithm), without first going
through the intermediate step of estimating p(y = 1) (which is what logistic
regression does).

4 Functional and geometric margins

Let’s formalize the notions of the functional and geometric margins. Given a
training example (x(i) , y (i) ), we define the functional margin of (w, b) with
respect to the training example as

γ̂ (i) = y (i) (wT x(i) + b).

Note that if y (i) = 1, then for the functional margin to be large (i.e., for
our prediction to be confident and correct), we need wT x(i) + b to be a large
positive number. Conversely, if y (i) = −1, then for the functional margin
to be large, we need wT x(i) + b to be a large negative number. Moreover, if
y (i) (wT x(i) + b) > 0, then our prediction on this example is correct. (Check
this yourself.) Hence, a large functional margin represents a confident and a
correct prediction.
For a linear classifier with the choice of g given above (taking values in
{−1, 1}), there’s one property of the functional margin that makes it not a
very good measure of confidence, however. Given our choice of g, we note that
if we replace w with 2w and b with 2b, then since g(wT x + b) = g(2wT x + 2b),
 14

this would not change hw,b (x) at all. I.e., g, and hence also hw,b (x), depends
only on the sign, but not on the magnitude, of wT x + b. However, replacing
(w, b) with (2w, 2b) also results in multiplying our functional margin by a
factor of 2. Thus, it seems that by exploiting our freedom to scale w and b,
we can make the functional margin arbitrarily large without really changing
anything meaningful. Intuitively, it might therefore make sense to impose
some sort of normalization condition such as that ||w||2 = 1; i.e., we might
replace (w, b) with (w/||w||2 , b/||w||2 ), and instead consider the functional
margin of (w/||w||2 , b/||w||2 ). We’ll come back to this later.
Given a training set S = {(x(i) , y (i) ); i = 1, . . . , n}, we also define the
function margin of (w, b) with respect to S as the smallest of the functional
margins of the individual training examples. Denoted by γ̂, this can therefore
be written:
γ̂ = min γ̂ (i) .
i=1,...,n

Next, let’s talk about geometric margins. Consider the picture below:

A w

γ (i)

The decision boundary corresponding to (w, b) is shown, along with the

vector w. Note that w is orthogonal (at 90◦ ) to the separating hyperplane.
(You should convince yourself that this must be the case.) Consider the
point at A, which represents the input x(i) of some training example with
label y (i) = 1. Its distance to the decision boundary, γ (i) , is given by the line
segment AB.
How can we find the value of γ (i) ? Well, w/||w|| is a unit-length vector
pointing in the same direction as w. Since A represents x(i) , we therefore
find that the point B is given by x(i) − γ (i) · w/||w||. But this point lies on
 15

the decision boundary, and all points x on the decision boundary satisfy the
equation wT x + b = 0. Hence,
 
T (i) (i) w
w x −γ + b = 0.
||w||

Solving for γ (i) yields

T
wT x(i) + b

(i) w b
γ = = x(i) + .
||w|| ||w|| ||w||

This was worked out for the case of a positive training example at A in the
figure, where being on the “positive” side of the decision boundary is good.
More generally, we define the geometric margin of (w, b) with respect to a
training example (x(i) , y (i) ) to be
 T !
w b
γ (i) = y (i) x(i) + .
||w|| ||w||

Note that if ||w|| = 1, then the functional margin equals the geometric
margin—this thus gives us a way of relating these two different notions of
margin. Also, the geometric margin is invariant to rescaling of the parame-
ters; i.e., if we replace w with 2w and b with 2b, then the geometric margin
does not change. This will in fact come in handy later. Specifically, because
of this invariance to the scaling of the parameters, when trying to fit w and b
to training data, we can impose an arbitrary scaling constraint on w without
changing anything important; for instance, we can demand that ||w|| = 1, or
|w1 | = 5, or |w1 + b| + |w2 | = 2, and any of these can be satisfied simply by
rescaling w and b.
Finally, given a training set S = {(x(i) , y (i) ); i = 1, . . . , n}, we also define
the geometric margin of (w, b) with respect to S to be the smallest of the
geometric margins on the individual training examples:

γ = min γ (i) .
i=1,...,n

5 The optimal margin classifier

Given a training set, it seems from our previous discussion that a natural
desideratum is to try to find a decision boundary that maximizes the (ge-
ometric) margin, since this would reflect a very confident set of predictions
 16

on the training set and a good “fit” to the training data. Specifically, this
will result in a classifier that separates the positive and the negative training
examples with a “gap” (geometric margin).
For now, we will assume that we are given a training set that is linearly
separable; i.e., that it is possible to separate the positive and negative ex-
amples using some separating hyperplane. How will we find the one that
achieves the maximum geometric margin? We can pose the following opti-
mization problem:

maxγ,w,b γ
s.t. y (i) (wT x(i) + b) ≥ γ, i = 1, . . . , n
||w|| = 1.

I.e., we want to maximize γ, subject to each training example having func-

tional margin at least γ. The ||w|| = 1 constraint moreover ensures that the
functional margin equals to the geometric margin, so we are also guaranteed
that all the geometric margins are at least γ. Thus, solving this problem will
result in (w, b) with the largest possible geometric margin with respect to the
training set.
If we could solve the optimization problem above, we’d be done. But the
“||w|| = 1” constraint is a nasty (non-convex) one, and this problem certainly
isn’t in any format that we can plug into standard optimization software to
solve. So, let’s try transforming the problem into a nicer one. Consider:
γ̂
maxγ̂,w,b
||w||
s.t. y (i) (wT x(i) + b) ≥ γ̂, i = 1, . . . , n

Here, we’re going to maximize γ̂/||w||, subject to the functional margins all
being at least γ̂. Since the geometric and functional margins are related by
γ = γ̂/||w|, this will give us the answer we want. Moreover, we’ve gotten rid
of the constraint ||w|| = 1 that we didn’t like. The downside is that we now
γ̂
have a nasty (again, non-convex) objective ||w|| function; and, we still don’t
have any off-the-shelf software that can solve this form of an optimization
problem.
Let’s keep going. Recall our earlier discussion that we can add an arbi-
trary scaling constraint on w and b without changing anything. This is the
key idea we’ll use now. We will introduce the scaling constraint that the
functional margin of w, b with respect to the training set must be 1:

γ̂ = 1.
 17

Since multiplying w and b by some constant results in the functional margin

being multiplied by that same constant, this is indeed a scaling constraint,
and can be satisfied by rescaling w, b. Plugging this into our problem above,
and noting that maximizing γ̂/||w|| = 1/||w|| is the same thing as minimizing
||w||2 , we now have the following optimization problem:
1
minw,b ||w||2
2
s.t. y (i) (wT x(i) + b) ≥ 1, i = 1, . . . , n
We’ve now transformed the problem into a form that can be efficiently
solved. The above is an optimization problem with a convex quadratic ob-
jective and only linear constraints. Its solution gives us the optimal mar-
gin classifier. This optimization problem can be solved using commercial
quadratic programming (QP) code.4
While we could call the problem solved here, what we will instead do is
make a digression to talk about Lagrange duality. This will lead us to our
optimization problem’s dual form, which will play a key role in allowing us to
use kernels to get optimal margin classifiers to work efficiently in very high
dimensional spaces. The dual form will also allow us to derive an efficient
algorithm for solving the above optimization problem that will typically do
much better than generic QP software.

6 Lagrange duality (optional reading)

Let’s temporarily put aside SVMs and maximum margin classifiers, and talk
about solving constrained optimization problems.
Consider a problem of the following form:
minw f (w)
s.t. hi (w) = 0, i = 1, . . . , l.
Some of you may recall how the method of Lagrange multipliers can be used
to solve it. (Don’t worry if you haven’t seen it before.) In this method, we
define the Lagrangian to be
l
X
L(w, β) = f (w) + βi hi (w)
i=1
4
You may be familiar with linear programming, which solves optimization problems
that have linear objectives and linear constraints. QP software is also widely available,
which allows convex quadratic objectives and linear constraints.
 18

Here, the βi ’s are called the Lagrange multipliers. We would then find
and set L’s partial derivatives to zero:
∂L ∂L
= 0; = 0,
∂wi ∂βi
and solve for w and β.
In this section, we will generalize this to constrained optimization prob-
lems in which we may have inequality as well as equality constraints. Due to
time constraints, we won’t really be able to do the theory of Lagrange duality
justice in this class,5 but we will give the main ideas and results, which we
will then apply to our optimal margin classifier’s optimization problem.
Consider the following, which we’ll call the primal optimization problem:
minw f (w)
s.t. gi (w) ≤ 0, i = 1, . . . , k
hi (w) = 0, i = 1, . . . , l.
To solve it, we start by defining the generalized Lagrangian
k
X l
X
L(w, α, β) = f (w) + αi gi (w) + βi hi (w).
i=1 i=1

Here, the αi ’s and βi ’s are the Lagrange multipliers. Consider the quantity
θP (w) = max L(w, α, β).
α,β : αi ≥0

Here, the “P” subscript stands for “primal.” Let some w be given. If w
violates any of the primal constraints (i.e., if either gi (w) > 0 or hi (w) 6= 0
for some i), then you should be able to verify that
k
X l
X
θP (w) = max f (w) + αi gi (w) + βi hi (w) (13)
α,β : αi ≥0
i=1 i=1
= ∞. (14)
Conversely, if the constraints are indeed satisfied for a particular value of w,
then θP (w) = f (w). Hence,

f (w) if w satisfies primal constraints
θP (w) =
∞ otherwise.
5
Readers interested in learning more about this topic are encouraged to read, e.g., R.
T. Rockarfeller (1970), Convex Analysis, Princeton University Press.
 19

Thus, θP takes the same value as the objective in our problem for all val-
ues of w that satisfies the primal constraints, and is positive infinity if the
constraints are violated. Hence, if we consider the minimization problem
min θP (w) = min max L(w, α, β),
w w α,β : αi ≥0

we see that it is the same problem (i.e., and has the same solutions as) our
original, primal problem. For later use, we also define the optimal value of
the objective to be p∗ = minw θP (w); we call this the value of the primal
problem.
Now, let’s look at a slightly different problem. We define
θD (α, β) = min L(w, α, β).
w
Here, the “D” subscript stands for “dual.” Note also that whereas in the
definition of θP we were optimizing (maximizing) with respect to α, β, here
we are minimizing with respect to w.
We can now pose the dual optimization problem:
max θD (α, β) = max min L(w, α, β).
α,β : αi ≥0 α,β : αi ≥0 w

This is exactly the same as our primal problem shown above, except that the
order of the “max” and the “min” are now exchanged. We also define the
optimal value of the dual problem’s objective to be d∗ = maxα,β : αi ≥0 θD (w).
How are the primal and the dual problems related? It can easily be shown
that
d∗ = max min L(w, α, β) ≤ min max L(w, α, β) = p∗ .
α,β : αi ≥0 w w α,β : αi ≥0

(You should convince yourself of this; this follows from the “max min” of a
function always being less than or equal to the “min max.”) However, under
certain conditions, we will have
d∗ = p∗ ,
so that we can solve the dual problem in lieu of the primal problem. Let’s
see what these conditions are.
Suppose f and the gi ’s are convex,6 and the hi ’s are affine.7 Suppose
further that the constraints gi are (strictly) feasible; this means that there
exists some w so that gi (w) < 0 for all i.
6
When f has a Hessian, then it is convex if and only if the Hessian is positive semi-
definite. For instance, f (w) = wT w is convex; similarly, all linear (and affine) functions
are also convex. (A function f can also be convex without being differentiable, but we
won’t need those more general definitions of convexity here.)
7
I.e., there exists ai , bi , so that hi (w) = aTi w + bi . “Affine” means the same thing as
linear, except that we also allow the extra intercept term bi .
 20

Under our above assumptions, there must exist w∗ , α∗ , β ∗ so that w∗ is the

solution to the primal problem, α∗ , β ∗ are the solution to the dual problem,
and moreover p∗ = d∗ = L(w∗ , α∗ , β ∗ ). Moreover, w∗ , α∗ and β ∗ satisfy the
Karush-Kuhn-Tucker (KKT) conditions, which are as follows:

∂
L(w∗ , α∗ , β ∗ ) = 0, i = 1, . . . , d (15)
∂wi
∂
L(w∗ , α∗ , β ∗ ) = 0, i = 1, . . . , l (16)
∂βi
αi∗ gi (w∗ ) = 0, i = 1, . . . , k (17)
gi (w∗ ) ≤ 0, i = 1, . . . , k (18)
α∗ ≥ 0, i = 1, . . . , k (19)

Moreover, if some w∗ , α∗ , β ∗ satisfy the KKT conditions, then it is also a

solution to the primal and dual problems.
We draw attention to Equation (17), which is called the KKT dual
complementarity condition. Specifically, it implies that if αi∗ > 0, then
gi (w∗ ) = 0. (I.e., the “gi (w) ≤ 0” constraint is active, meaning it holds with
equality rather than with inequality.) Later on, this will be key for showing
that the SVM has only a small number of “support vectors”; the KKT dual
complementarity condition will also give us our convergence test when we
talk about the SMO algorithm.

7 Optimal margin classifiers

Note: The equivalence of optimization problem (20) and the optimization
problem (24), and the relationship between the primary and dual variables in
equation (22) are the most important take home messages of this section.

Previously, we posed the following (primal) optimization problem for find-

ing the optimal margin classifier:
1
minw,b ||w||2 (20)
2
s.t. y (i) (wT x(i) + b) ≥ 1, i = 1, . . . , n

We can write the constraints as

gi (w) = −y (i) (wT x(i) + b) + 1 ≤ 0.

 21

We have one such constraint for each training example. Note that from the
KKT dual complementarity condition, we will have αi > 0 only for the train-
ing examples that have functional margin exactly equal to one (i.e., the ones
corresponding to constraints that hold with equality, gi (w) = 0). Consider
the figure below, in which a maximum margin separating hyperplane is shown
by the solid line.

The points with the smallest margins are exactly the ones closest to the
decision boundary; here, these are the three points (one negative and two pos-
itive examples) that lie on the dashed lines parallel to the decision boundary.
Thus, only three of the αi ’s—namely, the ones corresponding to these three
training examples—will be non-zero at the optimal solution to our optimiza-
tion problem. These three points are called the support vectors in this
problem. The fact that the number of support vectors can be much smaller
than the size the training set will be useful later.
Let’s move on. Looking ahead, as we develop the dual form of the prob-
lem, one key idea to watch out for is that we’ll try to write our algorithm
in terms of only the inner product hx(i) , x(j) i (think of this as (x(i) )T x(j) )
between points in the input feature space. The fact that we can express our
algorithm in terms of these inner products will be key when we apply the
kernel trick.
When we construct the Lagrangian for our optimization problem we have:
n
1 2
X
αi y (i) (wT x(i) + b) − 1 .
 
L(w, b, α) = ||w|| − (21)
2 i=1

Note that there’re only “αi ” but no “βi ” Lagrange multipliers, since the
problem has only inequality constraints.
 22

Let’s find the dual form of the problem. To do so, we need to first
minimize L(w, b, α) with respect to w and b (for fixed α), to get θD , which
we’ll do by setting the derivatives of L with respect to w and b to zero. We
have: n
X
∇w L(w, b, α) = w − αi y (i) x(i) = 0
i=1

This implies that

n
X
w= αi y (i) x(i) . (22)
i=1

As for the derivative with respect to b, we obtain

n
∂ X
L(w, b, α) = αi y (i) = 0. (23)
∂b i=1

If we take the definition of w in Equation (22) and plug that back into
the Lagrangian (Equation 21), and simplify, we get
n n n
X 1 X (i) (j) X
L(w, b, α) = αi − y y αi αj (x(i) )T x(j) − b αi y (i) .
i=1
2 i,j=1 i=1

But from Equation (23), the last term must be zero, so we obtain
n n
X 1 X (i) (j)
L(w, b, α) = αi − y y αi αj (x(i) )T x(j) .
i=1
2 i,j=1

Recall that we got to the equation above by minimizing L with respect to w

and b. Putting this together with the constraints αi ≥ 0 (that we always had)
and the constraint (23), we obtain the following dual optimization problem:
n n
X 1 X (i) (j)
maxα W (α) = αi − y y αi αj hx(i) , x(j) i. (24)
i=1
2 i,j=1
s.t. αi ≥ 0, i = 1, . . . , n
Xn
αi y (i) = 0,
i=1

You should also be able to verify that the conditions required for p∗ = d∗
and the KKT conditions (Equations 15–19) to hold are indeed satisfied in
our optimization problem. Hence, we can solve the dual in lieu of solving
 23

the primal problem. Specifically, in the dual problem above, we have a

maximization problem in which the parameters are the αi ’s. We’ll talk later
about the specific algorithm that we’re going to use to solve the dual problem,
but if we are indeed able to solve it (i.e., find the α’s that maximize W (α)
subject to the constraints), then we can use Equation (22) to go back and
find the optimal w’s as a function of the α’s. Having found w∗ , by considering
the primal problem, it is also straightforward to find the optimal value for
the intercept term b as

∗
maxi:y(i) =−1 w∗ T x(i) + mini:y(i) =1 w∗ T x(i)
b =− . (25)
2
(Check for yourself that this is correct.)
Before moving on, let’s also take a more careful look at Equation (22),
which gives the optimal value of w in terms of (the optimal value of) α.
Suppose we’ve fit our model’s parameters to a training set, and now wish to
make a prediction at a new point input x. We would then calculate wT x + b,
and predict y = 1 if and only if this quantity is bigger than zero. But
using (22), this quantity can also be written:

n
!T
X
wT x + b = αi y (i) x(i) x+b (26)
i=1
n
X
= αi y (i) hx(i) , xi + b. (27)
i=1

Hence, if we’ve found the αi ’s, in order to make a prediction, we have to

calculate a quantity that depends only on the inner product between x and
the points in the training set. Moreover, we saw earlier that the αi ’s will all
be zero except for the support vectors. Thus, many of the terms in the sum
above will be zero, and we really need to find only the inner products between
x and the support vectors (of which there is often only a small number) in
order calculate (27) and make our prediction.
By examining the dual form of the optimization problem, we gained sig-
nificant insight into the structure of the problem, and were also able to write
the entire algorithm in terms of only inner products between input feature
vectors. In the next section, we will exploit this property to apply the ker-
nels to our classification problem. The resulting algorithm, support vector
machines, will be able to efficiently learn in very high dimensional spaces.
 24

8 Regularization and the non-separable case

(optional reading)
The derivation of the SVM as presented so far assumed that the data is
linearly separable. While mapping data to a high dimensional feature space
via φ does generally increase the likelihood that the data is separable, we
can’t guarantee that it always will be so. Also, in some cases it is not clear
that finding a separating hyperplane is exactly what we’d want to do, since
that might be susceptible to outliers. For instance, the left figure below
shows an optimal margin classifier, and when a single outlier is added in the
upper-left region (right figure), it causes the decision boundary to make a
dramatic swing, and the resulting classifier has a much smaller margin.

To make the algorithm work for non-linearly separable datasets as well

as be less sensitive to outliers, we reformulate our optimization (using `1
regularization) as follows:
n
1 X
minγ,w,b ||w||2 + C ξi
2 i=1

s.t. y (i) (wT x(i) + b) ≥ 1 − ξi , i = 1, . . . , n

ξi ≥ 0, i = 1, . . . , n.

Thus, examples are now permitted to have (functional) margin less than 1,
and if an example has functional margin 1 − ξi (with ξ > 0), we would pay
a cost of the objective function being increased by Cξi . The parameter C
 25

controls the relative weighting between the twin goals of making the ||w||2
small (which we saw earlier makes the margin large) and of ensuring that
most examples have functional margin at least 1.
As before, we can form the Lagrangian:
n n n
1 X X   X
L(w, b, ξ, α, r) = wT w + C ξi − αi y (i) (xT w + b) − 1 + ξi − ri ξi .
2 i=1 i=1 i=1

Here, the αi ’s and ri ’s are our Lagrange multipliers (constrained to be ≥ 0).

We won’t go through the derivation of the dual again in detail, but after
setting the derivatives with respect to w and b to zero as before, substituting
them back in, and simplifying, we obtain the following dual form of the
problem:
n n
X 1 X (i) (j)
maxα W (α) = αi − y y αi αj hx(i) , x(j) i
i=1
2 i,j=1
s.t. 0 ≤ αi ≤ C, i = 1, . . . , n
Xn
αi y (i) = 0,
i=1

As before, we also have that w can be expressed in terms of the αi ’s

as given in Equation (22), so that after solving the dual problem, we can
continue to use Equation (27) to make our predictions. Note that, somewhat
surprisingly, in adding `1 regularization, the only change to the dual problem
is that what was originally a constraint that 0 ≤ αi has now become 0 ≤
αi ≤ C. The calculation for b∗ also has to be modified (Equation 25 is no
longer valid); see the comments in the next section/Platt’s paper.
Also, the KKT dual-complementarity conditions (which in the next sec-
tion will be useful for testing for the convergence of the SMO algorithm)
are:

αi = 0 ⇒ y (i) (wT x(i) + b) ≥ 1 (28)

αi = C ⇒ y (i) (wT x(i) + b) ≤ 1 (29)
0 < αi < C ⇒ y (i) (wT x(i) + b) = 1. (30)

Now, all that remains is to give an algorithm for actually solving the dual
problem, which we will do in the next section.
 26

9 The SMO algorithm (optional reading)

The SMO (sequential minimal optimization) algorithm, due to John Platt,
gives an efficient way of solving the dual problem arising from the derivation
of the SVM. Partly to motivate the SMO algorithm, and partly because it’s
interesting in its own right, let’s first take another digression to talk about
the coordinate ascent algorithm.

9.1 Coordinate ascent

Consider trying to solve the unconstrained optimization problem

max W (α1 , α2 , . . . , αn ).
α

Here, we think of W as just some function of the parameters αi ’s, and for now
ignore any relationship between this problem and SVMs. We’ve already seen
two optimization algorithms, gradient ascent and Newton’s method. The
new algorithm we’re going to consider here is called coordinate ascent:

Loop until convergence: {

For i = 1, . . . , n, {
αi := arg maxα̂i W (α1 , . . . , αi−1 , α̂i , αi+1 , . . . , αn ).
}

Thus, in the innermost loop of this algorithm, we will hold all the variables
except for some αi fixed, and reoptimize W with respect to just the parameter
αi . In the version of this method presented here, the inner-loop reoptimizes
the variables in order α1 , α2 , . . . , αn , α1 , α2 , . . .. (A more sophisticated version
might choose other orderings; for instance, we may choose the next variable
to update according to which one we expect to allow us to make the largest
increase in W (α).)
When the function W happens to be of such a form that the “arg max”
in the inner loop can be performed efficiently, then coordinate ascent can be
a fairly efficient algorithm. Here’s a picture of coordinate ascent in action:
 27

2.5

1.5

0.5

−0.5

−1

−1.5

−2

−2 −1.5 −1 −0.5 0 0.5 1 1.5 2 2.5

The ellipses in the figure are the contours of a quadratic function that
we want to optimize. Coordinate ascent was initialized at (2, −2), and also
plotted in the figure is the path that it took on its way to the global maximum.
Notice that on each step, coordinate ascent takes a step that’s parallel to one
of the axes, since only one variable is being optimized at a time.

9.2 SMO
We close off the discussion of SVMs by sketching the derivation of the SMO
algorithm. Some details will be left to the homework, and for others you
may refer to the paper excerpt handed out in class.
Here’s the (dual) optimization problem that we want to solve:
n n
X 1 X (i) (j)
maxα W (α) = αi − y y αi αj hx(i) , x(j) i. (31)
i=1
2 i,j=1
s.t. 0 ≤ αi ≤ C, i = 1, . . . , n (32)
Xn
αi y (i) = 0. (33)
i=1

Let’s say we have set of αi ’s that satisfy the constraints (32-33). Now,
suppose we want to hold α2 , . . . , αn fixed, and take a coordinate ascent step
and reoptimize the objective with respect to α1 . Can we make any progress?
The answer is no, because the constraint (33) ensures that
n
X
α1 y (1) = − αi y (i) .
i=2
 28

Or, by multiplying both sides by y (1) , we equivalently have

n
X
α1 = −y (1) αi y (i) .
i=2

(This step used the fact that y (1) ∈ {−1, 1}, and hence (y (1) )2 = 1.) Hence,
α1 is exactly determined by the other αi ’s, and if we were to hold α2 , . . . , αn
fixed, then we can’t make any change to α1 without violating the con-
straint (33) in the optimization problem.
Thus, if we want to update some subject of the αi ’s, we must update at
least two of them simultaneously in order to keep satisfying the constraints.
This motivates the SMO algorithm, which simply does the following:
Repeat till convergence {
1. Select some pair αi and αj to update next (using a heuristic that
tries to pick the two that will allow us to make the biggest progress
towards the global maximum).
2. Reoptimize W (α) with respect to αi and αj , while holding all the
other αk ’s (k 6= i, j) fixed.
}
To test for convergence of this algorithm, we can check whether the KKT
conditions (Equations 28-30) are satisfied to within some tol. Here, tol is
the convergence tolerance parameter, and is typically set to around 0.01 to
0.001. (See the paper and pseudocode for details.)
The key reason that SMO is an efficient algorithm is that the update to
αi , αj can be computed very efficiently. Let’s now briefly sketch the main
ideas for deriving the efficient update.
Let’s say we currently have some setting of the αi ’s that satisfy the con-
straints (32-33), and suppose we’ve decided to hold α3 , . . . , αn fixed, and
want to reoptimize W (α1 , α2 , . . . , αn ) with respect to α1 and α2 (subject to
the constraints). From (33), we require that
n
X
(1) (2)
α1 y + α2 y =− αi y (i) .
i=3

Since the right hand side is fixed (as we’ve fixed α3 , . . . αn ), we can just let
it be denoted by some constant ζ:
α1 y (1) + α2 y (2) = ζ. (34)
We can thus picture the constraints on α1 and α2 as follows:
 29

H α1y(1)+ α2y(2)=ζ
α2

L
α1 C

From the constraints (32), we know that α1 and α2 must lie within the box
[0, C] × [0, C] shown. Also plotted is the line α1 y (1) + α2 y (2) = ζ, on which we
know α1 and α2 must lie. Note also that, from these constraints, we know
L ≤ α2 ≤ H; otherwise, (α1 , α2 ) can’t simultaneously satisfy both the box
and the straight line constraint. In this example, L = 0. But depending on
what the line α1 y (1) + α2 y (2) = ζ looks like, this won’t always necessarily be
the case; but more generally, there will be some lower-bound L and some
upper-bound H on the permissible values for α2 that will ensure that α1 , α2
lie within the box [0, C] × [0, C].
Using Equation (34), we can also write α1 as a function of α2 :

α1 = (ζ − α2 y (2) )y (1) .

(Check this derivation yourself; we again used the fact that y (1) ∈ {−1, 1} so
that (y (1) )2 = 1.) Hence, the objective W (α) can be written

W (α1 , α2 , . . . , αn ) = W ((ζ − α2 y (2) )y (1) , α2 , . . . , αn ).

Treating α3 , . . . , αn as constants, you should be able to verify that this is

just some quadratic function in α2 . I.e., this can also be expressed in the
form aα22 + bα2 + c for some appropriate a, b, and c. If we ignore the “box”
constraints (32) (or, equivalently, that L ≤ α2 ≤ H), then we can easily
maximize this quadratic function by setting its derivative to zero and solving.
We’ll let α2new,unclipped denote the resulting value of α2 . You should also be
able to convince yourself that if we had instead wanted to maximize W with
respect to α2 but subject to the box constraint, then we can find the resulting
value optimal simply by taking α2new,unclipped and “clipping” it to lie in the
 30

[L, H] interval, to get

if α2new,unclipped > H

 H
α2new = αnew,unclipped if L ≤ α2new,unclipped ≤ H
 2
L if α2new,unclipped < L

Finally, having found the α2new , we can use Equation (34) to go back and find
the optimal value of α1new .
There’re a couple more details that are quite easy but that we’ll leave you
to read about yourself in Platt’s paper: One is the choice of the heuristics
used to select the next αi , αj to update; the other is how to update b as the
SMO algorithm is run.
 CS229 Lecture notes
Andrew Ng

Part VI
Learning Theory
1 Bias/variance tradeoff
When talking about linear regression, we discussed the problem of whether
to fit a “simple” model such as the linear “y = θ0 +θ1 x,” or a more “complex”
model such as the polynomial “y = θ0 + θ1 x + · · · θ5 x5 .” We saw the following
example:
4.5 4.5 4.5

4 4 4

3.5 3.5 3.5

3 3 3

2.5 2.5 2.5

y

2 2 2

1.5 1.5 1.5

1 1 1

0.5 0.5 0.5

0 0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
x x x

Fitting a 5th order polynomial to the data (rightmost figure) did not
result in a good model. Specifically, even though the 5th order polynomial
did a very good job predicting y (say, prices of houses) from x (say, living
area) for the examples in the training set, we do not expect the model shown
to be a good one for predicting the prices of houses not in the training set. In
other words, what’s has been learned from the training set does not generalize
well to other houses. The generalization error (which will be made formal
shortly) of a hypothesis is its expected error on examples not necessarily in
the training set.
Both the models in the leftmost and the rightmost figures above have
large generalization error. However, the problems that the two models suffer
from are very different. If the relationship between y and x is not linear,

1
 2

then even if we were fitting a linear model to a very large amount of training
data, the linear model would still fail to accurately capture the structure
in the data. Informally, we define the bias of a model to be the expected
generalization error even if we were to fit it to a very (say, infinitely) large
training set. Thus, for the problem above, the linear model suffers from large
bias, and may underfit (i.e., fail to capture structure exhibited by) the data.
Apart from bias, there’s a second component to the generalization error,
consisting of the variance of a model fitting procedure. Specifically, when
fitting a 5th order polynomial as in the rightmost figure, there is a large risk
that we’re fitting patterns in the data that happened to be present in our
small, finite training set, but that do not reflect the wider pattern of the
relationship between x and y. This could be, say, because in the training set
we just happened by chance to get a slightly more-expensive-than-average
house here, and a slightly less-expensive-than-average house there, and so
on. By fitting these “spurious” patterns in the training set, we might again
obtain a model with large generalization error. In this case, we say the model
has large variance.1
Often, there is a tradeoff between bias and variance. If our model is too
“simple” and has very few parameters, then it may have large bias (but small
variance); if it is too “complex” and has very many parameters, then it may
suffer from large variance (but have smaller bias). In the example above,
fitting a quadratic function does better than either of the extremes of a first
or a fifth order polynomial.

2 Preliminaries
In this set of notes, we begin our foray into learning theory. Apart from
being interesting and enlightening in its own right, this discussion will also
help us hone our intuitions and derive rules of thumb about how to best
apply learning algorithms in different settings. We will also seek to answer
a few questions: First, can we make formal the bias/variance tradeoff that
was just discussed? The will also eventually lead us to talk about model
selection methods, which can, for instance, automatically decide what order
polynomial to fit to a training set. Second, in machine learning it’s really
1
In these notes, we will not try to formalize the definitions of bias and variance beyond
this discussion. While bias and variance are straightforward to define formally for, e.g.,
linear regression, there have been several proposals for the definitions of bias and variance
for classification, and there is as yet no agreement on what is the “right” and/or the most
useful formalism.
 3

generalization error that we care about, but most learning algorithms fit their
models to the training set. Why should doing well on the training set tell us
anything about generalization error? Specifically, can we relate error on the
training set to generalization error? Third and finally, are there conditions
under which we can actually prove that learning algorithms will work well?
We start with two simple but very useful lemmas.
Lemma. (The union bound). Let A1 , A2 , . . . , Ak be k different events (that
may not be independent). Then
P (A1 ∪ · · · ∪ Ak ) ≤ P (A1 ) + . . . + P (Ak ).
In probability theory, the union bound is usually stated as an axiom
(and thus we won’t try to prove it), but it also makes intuitive sense: The
probability of any one of k events happening is at most the sums of the
probabilities of the k different events.
Lemma. (Hoeffding inequality) Let Z1 , . . . , Zm be m independent and iden-
tically distributed (iid) random variables drawn from a Bernoulli(φ)Pdistri-
bution. I.e., P (Zi = 1) = φ, and P (Zi = 0) = 1 − φ. Let φ̂ = (1/m) mi=1 Zi
be the mean of these random variables, and let any γ > 0 be fixed. Then
P (|φ − φ̂| > γ) ≤ 2 exp(−2γ 2 m)
This lemma (which in learning theory is also called the Chernoff bound)
says that if we take φ̂—the average of m Bernoulli(φ) random variables—to
be our estimate of φ, then the probability of our being far from the true value
is small, so long as m is large. Another way of saying this is that if you have
a biased coin whose chance of landing on heads is φ, then if you toss it m
times and calculate the fraction of times that it came up heads, that will be
a good estimate of φ with high probability (if m is large).
Using just these two lemmas, we will be able to prove some of the deepest
and most important results in learning theory.
To simplify our exposition, lets restrict our attention to binary classifica-
tion in which the labels are y ∈ {0, 1}. Everything we’ll say here generalizes
to other, including regression and multi-class classification, problems.
We assume we are given a training set S = {(x(i) , y (i) ); i = 1, . . . , m}
of size m, where the training examples (x(i) , y (i) ) are drawn iid from some
probability distribution D. For a hypothesis h, we define the training error
(also called the empirical risk or empirical error in learning theory) to
be m
1 X
ε̂(h) = 1{h(x(i) ) 6= y (i) }.
m i=1
 4

This is just the fraction of training examples that h misclassifies. When we

want to make explicit the dependence of ε̂(h) on the training set S, we may
also write this a ε̂S (h). We also define the generalization error to be

ε(h) = P(x,y)∼D (h(x) 6= y).

I.e. this is the probability that, if we now draw a new example (x, y) from
the distribution D, h will misclassify it.
Note that we have assumed that the training data was drawn from the
same distribution D with which we’re going to evaluate our hypotheses (in
the definition of generalization error). This is sometimes also referred to as
one of the PAC assumptions.2
Consider the setting of linear classification, and let hθ (x) = 1{θ T x ≥ 0}.
What’s a reasonable way of fitting the parameters θ? One approach is to try
to minimize the training error, and pick

θ̂ = arg min ε̂(hθ ).

θ

We call this process empirical risk minimization (ERM), and the resulting
hypothesis output by the learning algorithm is ĥ = hθ̂ . We think of ERM
as the most “basic” learning algorithm, and it will be this algorithm that we
focus on in these notes. (Algorithms such as logistic regression can also be
viewed as approximations to empirical risk minimization.)
In our study of learning theory, it will be useful to abstract away from
the specific parameterization of hypotheses and from issues such as whether
we’re using a linear classifier. We define the hypothesis class H used by a
learning algorithm to be the set of all classifiers considered by it. For linear
classification, H = {hθ : hθ (x) = 1{θ T x ≥ 0}, θ ∈ Rn+1 } is thus the set of
all classifiers over X (the domain of the inputs) where the decision boundary
is linear. More broadly, if we were studying, say, neural networks, then we
could let H be the set of all classifiers representable by some neural network
architecture.
Empirical risk minimization can now be thought of as a minimization over
the class of functions H, in which the learning algorithm picks the hypothesis:

ĥ = arg min ε̂(h)

h∈H

2
PAC stands for “probably approximately correct,” which is a framework and set of
assumptions under which numerous results on learning theory were proved. Of these, the
assumption of training and testing on the same distribution, and the assumption of the
independently drawn training examples, were the most important.
 5

3 The case of finite H

Lets start by considering a learning problem in which we have a finite hy-
pothesis class H = {h1 , . . . , hk } consisting of k hypotheses. Thus, H is just a
set of k functions mapping from X to {0, 1}, and empirical risk minimization
selects ĥ to be whichever of these k functions has the smallest training error.
We would like to give guarantees on the generalization error of ĥ. Our
strategy for doing so will be in two parts: First, we will show that ε̂(h) is a
reliable estimate of ε(h) for all h. Second, we will show that this implies an
upper-bound on the generalization error of ĥ.
Take any one, fixed, hi ∈ H. Consider a Bernoulli random variable Z
whose distribution is defined as follows. We’re going to sample (x, y) ∼ D.
Then, we set Z = 1{hi (x) 6= y}. I.e., we’re going to draw one example,
and let Z indicate whether hi misclassifies it. Similarly, we also define Zj =
1{hi (x(j) ) 6= y (j) }. Since our training set was drawn iid from D, Z and the
Zj ’s have the same distribution.
We see that the misclassification probability on a randomly drawn example—
that is, ε(h)—is exactly the expected value of Z (and Zj ). Moreover, the
training error can be written
m
1 X
ε̂(hi ) = Zj .
m j=1

Thus, ε̂(hi ) is exactly the mean of the m random variables Zj that are drawn
iid from a Bernoulli distribution with mean ε(hi ). Hence, we can apply the
Hoeffding inequality, and obtain

P (|ε(hi ) − ε̂(hi )| > γ) ≤ 2 exp(−2γ 2 m).

This shows that, for our particular hi , training error will be close to
generalization error with high probability, assuming m is large. But we
don’t just want to guarantee that ε(hi ) will be close to ε̂(hi ) (with high
probability) for just only one particular hi . We want to prove that this will
be true for simultaneously for all h ∈ H. To do so, let Ai denote the event
that |ε(hi ) − ε̂(hi )| > γ. We’ve already show that, for any particular Ai , it
holds true that P (Ai ) ≤ 2 exp(−2γ 2 m). Thus, using the union bound, we
 6

have that

P (∃ h ∈ H.|ε(hi ) − ε̂(hi )| > γ) = P (A1 ∪ · · · ∪ Ak )

Xk
≤ P (Ai )
i=1
Xk
≤ 2 exp(−2γ 2 m)
i=1
= 2k exp(−2γ 2 m)

If we subtract both sides from 1, we find that

P (¬∃ h ∈ H.|ε(hi ) − ε̂(hi )| > γ) = P (∀h ∈ H.|ε(hi ) − ε̂(hi )| ≤ γ)

≥ 1 − 2k exp(−2γ 2 m)

(The “¬” symbol means “not.”) So, with probability at least 1−2k exp(−2γ 2 m),
we have that ε(h) will be within γ of ε̂(h) for all h ∈ H. This is called a uni-
form convergence result, because this is a bound that holds simultaneously
for all (as opposed to just one) h ∈ H.
In the discussion above, what we did was, for particular values of m and
γ, given a bound on the probability that, for some h ∈ H, |ε(h) − ε̂(h)| > γ.
There are three quantities of interest here: m, γ, and the probability of error;
we can bound either one in terms of the other two.
For instance, we can ask the following question: Given γ and some δ > 0,
how large must m be before we can guarantee that with probability at least
1 − δ, training error will be within γ of generalization error? By setting
δ = 2k exp(−2γ 2 m) and solving for m, [you should convince yourself this is
the right thing to do!], we find that if
1 2k
m≥ log ,
2γ 2 δ
then with probability at least 1 − δ, we have that |ε(h) − ε̂(h)| ≤ γ for all
h ∈ H. (Equivalently, this show that the probability that |ε(h) − ε̂(h)| > γ
for some h ∈ H is at most δ.) This bound tells us how many training
examples we need in order make a guarantee. The training set size m that
a certain method or algorithm requires in order to achieve a certain level of
performance is also called the algorithm’s sample complexity.
The key property of the bound above is that the number of training
examples needed to make this guarantee is only logarithmic in k, the number
of hypotheses in H. This will be important later.
 7

Similarly, we can also hold m and δ fixed and solve for γ in the previous
equation, and show [again, convince yourself that this is right!] that with
probability 1 − δ, we have that for all h ∈ H,
r
1 2k
|ε̂(h) − ε(h)| ≤ log .
2m δ
Now, lets assume that uniform convergence holds, i.e., that |ε(h)− ε̂(h)| ≤
γ for all h ∈ H. What can we prove about the generalization of our learning
algorithm that picked ĥ = arg minh∈H ε̂(h)?
Define h∗ = arg minh∈H ε(h) to be the best possible hypothesis in H. Note
that h∗ is the best that we could possibly do given that we are using H, so
it makes sense to compare our performance to that of h∗ . We have:

ε(ĥ) ≤ ε̂(ĥ) + γ
≤ ε̂(h∗ ) + γ
≤ ε(h∗ ) + 2γ

The first line used the fact that |ε(ĥ)− ε̂(ĥ)| ≤ γ (by our uniform convergence
assumption). The second used the fact that ĥ was chosen to minimize ε̂(h),
and hence ε̂(ĥ) ≤ ε̂(h) for all h, and in particular ε̂(ĥ) ≤ ε̂(h∗ ). The third
line used the uniform convergence assumption again, to show that ε̂(h∗ ) ≤
ε(h∗ ) + γ. So, what we’ve shown is the following: If uniform convergence
occurs, then the generalization error of ĥ is at most 2γ worse than the best
possible hypothesis in H!
Lets put all this together into a theorem.
Theorem. Let |H| = k, and let any m, δ be fixed. Then with probability at
least 1 − δ, we have that
  r
1 2k
ε(ĥ) ≤ min ε(h) + 2 log .
h∈H 2m δ

√
This is proved by letting γ equal the · term, using our previous argu-
ment that uniform convergence occurs with probability at least 1 − δ, and
then noting that uniform convergence implies ε(h) is at most 2γ higher than
ε(h∗ ) = minh∈H ε(h) (as we showed previously).
This also quantifies what we were saying previously saying about the
bias/variance tradeoff in model selection. Specifically, suppose we have some
hypothesis class H, and are considering switching to some much larger hy-
pothesis class H0 ⊇ H. If we switch to H0 , then the first term minh ε(h)
 8

can only decrease (since we’d then be taking a min over a larger set of func-
tions). Hence, by learning using a larger hypothesis class,
√ our “bias” can
only decrease. However, if k increases, then the second 2 · term would also
increase. This increase corresponds to our “variance” increasing when we use
a larger hypothesis class.
By holding γ and δ fixed and solving for m like we did before, we can
also obtain the following sample complexity bound:
Corollary. Let |H| = k, and let any δ, γ be fixed. Then for ε(ĥ) ≤
minh∈H ε(h) + 2γ to hold with probability at least 1 − δ, it suffices that

1 2k
m ≥ log
2γ 2 δ
 
1 k
= O log ,
γ2 δ

4 The case of infinite H

We have proved some useful theorems for the case of finite hypothesis classes.
But many hypothesis classes, including any parameterized by real numbers
(as in linear classification) actually contain an infinite number of functions.
Can we prove similar results for this setting?
Lets start by going through something that is not the “right” argument.
Better and more general arguments exist, but this will be useful for honing
our intuitions about the domain.
Suppose we have an H that is parameterized by d real numbers. Since we
are using a computer to represent real numbers, and IEEE double-precision
floating point (double’s in C) uses 64 bits to represent a floating point num-
ber, this means that our learning algorithm, assuming we’re using double-
precision floating point, is parameterized by 64d bits. Thus, our hypothesis
class really consists of at most k = 264d different hypotheses. From the Corol-
lary at the end of the previous section, we therefore find that, to guarantee
ε(ĥ) ≤ ε(h∗ )+ 2γ, with to hold
 with probability at least 1 − δ, it suffices
1 264d
that m ≥ O γ 2 log δ = O γd2 log 1δ = Oγ,δ (d). (The γ, δ subscripts are
to indicate that the last big-O is hiding constants that may depend on γ and
δ.) Thus, the number of training examples needed is at most linear in the
parameters of the model.
The fact that we relied on 64-bit floating point makes this argument not
entirely satisfying, but the conclusion is nonetheless roughly correct: If what
we’re going to do is try to minimize training error, then in order to learn
 9

“well” using a hypothesis class that has d parameters, generally we’re going
to need on the order of a linear number of training examples in d.
(At this point, it’s worth noting that these results were proved for an al-
gorithm that uses empirical risk minimization. Thus, while the linear depen-
dence of sample complexity on d does generally hold for most discriminative
learning algorithms that try to minimize training error or some approxima-
tion to training error, these conclusions do not always apply as readily to
discriminative learning algorithms. Giving good theoretical guarantees on
many non-ERM learning algorithms is still an area of active research.)
The other part of our previous argument that’s slightly unsatisfying is
that it relies on the parameterization of H. Intuitively, this doesn’t seem like
it should matter: We had written the class of linear classifiers as hθ (x) =
1{θ0 + θ1 x1 + · · · θn xn ≥ 0}, with n + 1 parameters θ0 , . . . , θn . But it could
also be written hu,v (x) = 1{(u20 − v02 ) + (u21 − v12 )x1 + · · · (u2n − vn2 )xn ≥ 0}
with 2n + 2 parameters ui , vi . Yet, both of these are just defining the same
H: The set of linear classifiers in n dimensions.
To derive a more satisfying argument, lets define a few more things.
Given a set S = {x(i) , . . . , x(d) } (no relation to the training set) of points
x(i) ∈ X , we say that H shatters S if H can realize any labeling on S.
I.e., if for any set of labels {y (1) , . . . , y (d) }, there exists some h ∈ H so that
h(x(i) ) = y (i) for all i = 1, . . . d.
Given a hypothesis class H, we then define its Vapnik-Chervonenkis
dimension, written VC(H), to be the size of the largest set that is shattered
by H. (If H can shatter arbitrarily large sets, then VC(H) = ∞.)
For instance, consider the following set of three points:



x2


x1
Can the set H of linear classifiers in two dimensions (h(x) = 1{θ0 +θ1 x1 +
θ2 x2 ≥ 0}) can shatter the set above? The answer is yes. Specifically, we
 10

see that, for any of the eight possible labelings of these points, we can find a
linear classifier that obtains “zero training error” on them:

x2 x2 x2 x2

x1 x1 x1 x1

x2 x2 x2 x2

x1 x1 x1 x1

Moreover, it is possible to show that there is no set of 4 points that this

hypothesis class can shatter. Thus, the largest set that H can shatter is of
size 3, and hence VC(H) = 3.
Note that the VC dimension of H here is 3 even though there may be
sets of size 3 that it cannot shatter. For instance, if we had a set of three
points lying in a straight line (left figure), then there is no way to find a linear
separator for the labeling of the three points shown below (right figure):



x2 
x2

x1 x1
In order words, under the definition of the VC dimension, in order to
prove that VC(H) is at least d, we need to show only that there’s at least
one set of size d that H can shatter.
The following theorem, due to Vapnik, can then be shown. (This is, many
would argue, the most important theorem in all of learning theory.)
 11

Theorem. Let H be given, and let d = VC(H). Then with probability at

least 1 − δ, we have that for all h ∈ H,
r !
d m 1 1
|ε(h) − ε̂(h)| ≤ O log + log .
m d m δ

Thus, with probability at least 1 − δ, we also have that:

r !
∗ d m 1 1
ε(ĥ) ≤ ε(h ) + O log + log .
m d m δ

In other words, if a hypothesis class has finite VC dimension, then uniform

convergence occurs as m becomes large. As before, this allows us to give a
bound on ε(h) in terms of ε(h∗ ). We also have the following corollary:
Corollary. For |ε(h) − ε̂(h)| ≤ γ to hold for all h ∈ H (and hence ε(ĥ) ≤
ε(h∗ ) + 2γ) with probability at least 1 − δ, it suffices that m = Oγ,δ (d).
In other words, the number of training examples needed to learn “well”
using H is linear in the VC dimension of H. It turns out that, for “most”
hypothesis classes, the VC dimension (assuming a “reasonable” parameter-
ization) is also roughly linear in the number of parameters. Putting these
together, we conclude that (for an algorithm that tries to minimize training
error) the number of training examples needed is usually roughly linear in
the number of parameters of H.
 CS229 Bias-Variance and Error Analysis
Yoann Le Calonnec
October 2, 2017

1 The Bias-Variance Tradeoff

Assume you are given a well fitted machine learning model fˆ that you want to apply on
some test dataset. For instance, the model could be a linear regression whose parameters
were computed using some training set different from your test set. For each point x in your
test set, you want to predict the associated target y ∈ R, and compute the mean squared
error (MSE)

E(x,y)∼test set |fˆ(x) − y|2

You now realize that this MSE is too high, and try to find an explanation to this result:

• Overfitting: the model is too closely related to the examples in the training set and
doesn’t generalize well to other examples.

• Underfitting: the model didn’t gather enough information from the training set, and
doesn’t capture the link between the features x and the target y.

• The data is simply noisy, that is the model is neither overfitting or underfitting, and
the high MSE is simply due to the amount of noise in the dataset.

Our intuition can be formalized by the Bias-Variance tradeoff.

Assume that the points in your training/test set are all taken from a similar distribution,
with
yi = f (xi ) + i , where the noise i satisfies E(i ) = 0, Var(i ) = σ 2
and your goal is to compute f . By looking at your training set, you obtain an estimate fˆ.
Now use this estimate with your test set, meaning that for each example j in the test set,
your prediction for yj = f (xj ) + j is fˆ(xj ). Here, xj is a fixed real number (or vector if the
feature space is multi-dimensional) thus f (xj ) is fixed, and j is a real random variable with
mean 0 and variance σ 2 . The crucial observation is that fˆ(xj ) is random since it depends on
the values i from the training set. That’s why talking about the bias E(fˆ(x) − f (x)) and
the variance of fˆ makes sense.
We can now compute our MSE on the test set by computing the following expectation
with respect to the possible training sets (since fˆ is a random variable function of the choice
of the traning set)

1
  
Test MSE = E (y − fˆ(x))2
 
= E ( + f (x) − fˆ(x))2
 
= E(2 ) + E (f (x) − fˆ(x))2
 2  
2 ˆ ˆ
= σ + E(f (x) − f (x)) + Var f (x) − f (x)
 2  
2 ˆ
= σ + Bias f (x) + Var f (x) ˆ

There is nothing we can do about the first term σ 2 as we can not predict the noise  by
definition. The bias term is due to underfitting, meaning that on average, fˆ does not predict
f . The last term is closely related to overfitting, the prediction fˆ is too close from the values
ytrain and varies a lot with the choice of our training set.
To sum up, we can understand our MSE as follows

High Bias ←→ Underfitting

High Variance ←→ Overfitting
Large σ 2 ←→ Noisy data

Hence, when analyzing the performance of a machine learning algorithm, we must always
ask ourselves how to reduce the bias without increasing the variance, and respectively how to
reduce the variance without increasing the bias. Most of the time, reducing one will increase
the other, and there is a tradeoff between bias and variance.

2 Error Analysis
Even though understanding whether our poor test error is due to high bias or high variance
is important, knowing which parts of the machine learning algorithm lead to this error or
score is crucial.
Consider the machine learning pipeline on figure 1.
The algorithms is divided into several steps

1. The inputs are taken from a camera image

2. Preprocessing to remove the background on the image. For instance, if the image are
taken from a security camera, the background is always the same, and we could remove
it easily by keeping the pixels that changed on the image.

3. Detect the position of the face.

4. Detect the eyes - Detect the nose - Detect the mouth

2
 Figure 1: Face recognition pipeline

5. Final logistic regression step to predict the label

If you biuld a complicated system like this one, you might want to figure out how much
error is attributable to each of the components, how good is each of these green boxes.
Indeed, if one of these boxes is really problematic, you might want to spend more time
trying to improve the performance of that one green box. How do you decide what part to
focus on?
One thing we can do is plug in the ground-truth for each component, and see how
accuracy changes. Let’s say the overall accuracy of the system is 85% (pretty bad). You
can now take your development set and manually give it the perfect background removal,
that is, instead of using your background removal algorithm, manually specify the perfect
background removal yourself (using photoshop for instance), and look at how much that
affect the performance of the overall system.
Now let’s say the accuracy only improves by 0.1%. This gives us an upperbound, that
is even if we worked for years on background removal, it wouldn’t help our system by more
than 0.1%.
Now let’s give the pipeline the perfect face detection by specifying the position of the
face manually, see how much we improve the performance, and so on.
The results are specified in the table 1.
Looking at the table, we know that working on the background removal won’t help much.
It also tells us where the biggest jumps are. We notice that having an accurate face detection
mechanism really improves the performance, and similarly, the eyes really help making the
prediction more accurate.
Error analysis is also useful when publishing a paper, since it’s a convenient way to

3
 Component Accuracy
Overall system 85%
Preprocess (remove background) 85.1%
Face detection 91%
Eyes segmentation 95%
Nose segmentation 96%
Mouth segmentation 97%
Logistic regression 100%

Table 1: Accuracy when providing the system with the perfect component

analyze the error of an algorithm and explain which parts should be improved.

Ablative analysis
While error analysis tries to explain the difference between current performance and perfect
performance, ablative analysis tries to explain the difference between some baseline (much
poorer) performance and current performance.
For instance, suppose you have built a good anti-spam classifier by adding lots of clever
features to logistic regression

• Spelling correction

• Sender host features

• Email header features

• Email text parser features

• Javascript parser

• Features from embedded images

and your question is: How much did each of these components really help?
In this example, let’s say that simple logistic regression without any clever features gets
94% performance, but when adding these clever features, we get 99.9% performance. In
abaltive analysis, what we do is start from the current level of performance 99.9%, and
slowly take away all of these features to see how it affects performance. The results are
provided in table 2.
When presenting the results in a paper, ablative analysis really helps analyzing the fea-
tures that helped decreasing the misclassification rate. Instead of simply giving the loss/error
rate of the algorithm, we can provide evidence that some specific features are actually more
important than others.

4
 Component Accuracy
Overall system 99.9%
Spelling correction 99.0%
Sender host features 98.9%
Email header features 98.9%
Email text parser features 95%
Javascript parser 94.5%
Features from images 94.0%

Table 2: Accuracy when removing feature from logistic regression

Analyze your mistakes

Assume you are given a dataset with pictures of animals, and your goal is to identify pictures
of cats that you would eventually send to the members of a community of cat lovers. You
notice that there are many pictures of dogs in the original dataset, and wonders whether
you should build a special algorithm to identify the pictures of dogs and avoid sending dogs
pictures to cat lovers or not.
One thing you can do is take a 100 examples from your development set that are mis-
classified, and count up how many of these 100 mistakes are dogs. If 5% of them are dogs,
then even if you come up with a solution to identidy your dogs, your error would only go
down by 5%, that is your accuracy would go up from 90% to 90.5%. However, if 50 of these
100 errors are dogs, then you could improve your accuracy to reach 95%.
By analyzing your mistakes, you can focus on what’s really important. If you notice that
80 out of your 100 mistakes are blurry images, then work hard on classifying correctly these
blurry images. If you notice that 70 out of the 100 errors are great cats, then focus on this
specific task of identifying great cats.
In brief, do not waste your time improving parts of your algorithm that won’t really help
decreasing your error rate, and focus on what really matters.

5
 Some Calculations from Bias Variance
Christopher Ré
May 7, 2019

This note contains a reprise of the eigenvalue arguments to understand how

variance is reduced by regularization. We also ¿ describe different ways regu-
larization can occur including from the ¿ algorithm or initialization. This note
contains some additional calculations from the lecture and Piazza, just so that
we have typeset versions of them. They contain no new information over the
lecture, but they do supplement the notes.
Recall we have a design matrix X ∈ Rn×d and labels y ∈ Rn . We are inter-
ested in the underdetermined case n < d so that rank(X) ≤ n < d. We consider
the following optimization problem for least squares with a regularization pa-
rameter λ ≥ 0:
1 λ
`(θ; λ) = min kXθ − yk2 + kθk2
θ∈Rd 2 2

Normal Equations Computing derivatives as we did for the normal equa-

tions, we see that:

∇θ `(θ; λ) = X T (Xθ − y) + λθ = (X T X + λI)θ − X T y

By setting ∇θ `(θ, λ) = 0 we can solve for the θ̂ that minimizes the above prob-
lem. Explicitly, we have:

−1
θ̂ = X T X + λI XT y (1)

To see that the inverse in Eq. 1 exists, we observe that X T X is a symmetric,

real d×d matrix so it has d eigenvalues (some may be 0). Moreover, it is positive
semidefinite, and we capture this by writing eig(X T X) = {σ12 , . . . , σd2 }. Now,
inspired by the regularized problem, we examine:

eig(X T X + λI) = σ12 + λ, . . . , σd2 + λ



Since σi2 ≥ 0 for all i ∈ [d], if we set λ > 0 then X T X + λI is full rank, and the
inverse of (X T X + λI) exists. In turn, this means there is a unique such θ̂.

Variance Recall that in bias-variance, we are concerned with the variance of

θ̂ as we sample the training set. We want to argue that as the regularization
parameter λ increases, the variance in the fitted θ̂ decreases. We won’t carry

1
out the full formal argument, but it suffices to make one observation that is
immediate from Eq. 1: the variance of θ̂ is proportional to the eigenvalues of
(X T X + λI)−1 . To see this, observe that the eigenvalues of an inverse are just
the inverse of the eigenvalues:
 

T

−1  1 1
eig X X + λI = , ..., 2
σ12 + λ σd + λ

Now, condition on the points we draw, namely X. Then, recall that ran-
domness is in the label noise (recall the linear regression model y ∼ Xθ∗ +
N (0, τ 2 I) = N (Xθ∗ , τ 2 I)).
Recall a fact about the multivariate normal distribution:

if y ∼ N (µ, Σ) then Ay ∼ N (Aµ, AΣAT )

Using linearity, we can verify that the expectation of θ̂ is

E[θ̂] = E[(X T X + λI)−1 X T y]

= E[(X T X + λI)−1 X T (Xθ∗ + N (0, τ 2 , I))]
= E[(X T X + λI)−1 X T (Xθ∗ )]
= (X T X + λI)−1 (X T X)θ∗ (essentially a “shrunk” θ∗ )

The last line above suggests that the more regularization we add (larger the λ),
the more the estimated θ̂ will be shrunk towards 0. In other words, regulariza-
tion adds bias (towards zero in this case). Though we paid the cost of higher
bias, we gain by reducing the variance of θ̂. To see this bias-variance tradeoff
concretely, observe the covariance matrix of θ̂:

C := Cov[θ̂]
= (X T X + λI)−1 X T (τ 2 I) X(X T X + λI)−1

and
τ 2 σ12 τ 2 σd2
 
eig(C) = , . . . ,
(σ12 + λ)2 (σd2 + λ)2

Notice that the entire spectrum of the covariance is a decreasing function of λ.

By decomposing in the eigenvalue basis, we can see that actually E[kθ̂ − θ∗ k2 ]
is a decreasing function of λ, as desired.

Gradient Descent We show that you can initialize gradient descent in a way
that effectively regularizes undetermined least squares–even with no regulariza-
tion penalty (λ = 0). Our first observation is that any point x ∈ Rd can be
decomposed into two orthogonal components x0 , x1 such that

x = x0 + x1 and x0 ∈ Null(X T ) and Range(X).

2
Recall that Null(X T ) and Range(X) are orthogonal subspaces by the fundamen-
tal theory of linear algebra. We write P0 for the projection on the null and P1
for the projection on the range, then x0 = P0 (x) and x1 = P1 (x).
If one initializes at a point θ then, we observe that the gradient is orthogonal
to the null space. That is, if g(θ) = X T (Xθ−y) then g T P0 (v) = 0 for any v ∈ Rd .
But, then:

P0 (θ(t+1) ) = P0 (θt − αg(θ(t) )) = P0 (θt ) − αP0 g(θ(t) ) = P0 (θ(t) )

That is, no learning happens in the null. Whatever portion is in the null that
we initialize stays there throughout execution.
A key property of the Moore-Penrose pseudoinverse, is that if θ̂ = (X T X)+ X T y
then P0 (θ̂) = 0. Hence, the gradient descent solution initialized at θ0 can be
written θ̂ + P0 (θ0 ). Two immediate observations:

• Using the Moore-Penrose inverse acts as regularization, because it selects

the solution θ̂.
• So does gradient descent–provided that we initialize at θ0 = 0. This
is particularly interesting, as many modern machine learning techniques
operate in these underdetermined regimes.

We’ve argued that there are many ways to find equivalent solutions, and that
this allows us to understand the effect on the model fitting procedure as regu-
larization. Thus, there are many ways to find that equivalent solution. Many
modern methods of machine learning including dropout and data augmentation
are not penalty, but their effect is understood as regularization. One contrast
with the above methods is that they often depend on some property of the data
or for how much they effectively regularization. In some sense, they adapt to
the data. A final comment is that in the same sense above, adding more data
regularizes the model as well!

3
 CS229 Lecture notes
Andrew Ng

Part VI
Regularization and model
selection
Suppose we are trying select among several different models for a learning
problem. For instance, we might be using a polynomial regression model
hθ (x) = g(θ0 + θ1 x + θ2 x2 + · · · + θk xk ), and wish to decide if k should be
0, 1, . . . , or 10. How can we automatically select a model that represents
a good tradeoff between the twin evils of bias and variance1 ? Alternatively,
suppose we want to automatically choose the bandwidth parameter τ for
locally weighted regression, or the parameter C for our `1 -regularized SVM.
How can we do that?
For the sake of concreteness, in these notes we assume we have some
finite set of models M = {M1 , . . . , Md } that we’re trying to select among.
For instance, in our first example above, the model Mi would be an i-th
order polynomial regression model. (The generalization to infinite M is not
hard.2 ) Alternatively, if we are trying to decide between using an SVM, a
neural network or logistic regression, then M may contain these models.
1
Given that we said in the previous set of notes that bias and variance are two very
different beasts, some readers may be wondering if we should be calling them “twin” evils
here. Perhaps it’d be better to think of them as non-identical twins. The phrase “the
fraternal twin evils of bias and variance” doesn’t have the same ring to it, though.
2
If we are trying to choose from an infinite set of models, say corresponding to the
possible values of the bandwidth τ ∈ R+ , we may discretize τ and consider only a finite
number of possible values for it. More generally, most of the algorithms described here
can all be viewed as performing optimization search in the space of models, and we can
perform this search over infinite model classes as well.

1
 2

1 Cross validation
Lets suppose we are, as usual, given a training set S. Given what we know
about empirical risk minimization, here’s what might initially seem like a
algorithm, resulting from using empirical risk minimization for model selec-
tion:

1. Train each model Mi on S, to get some hypothesis hi .

2. Pick the hypotheses with the smallest training error.

This algorithm does not work. Consider choosing the order of a poly-
nomial. The higher the order of the polynomial, the better it will fit the
training set S, and thus the lower the training error. Hence, this method will
always select a high-variance, high-degree polynomial model, which we saw
previously is often poor choice.
Here’s an algorithm that works better. In hold-out cross validation
(also called simple cross validation), we do the following:

1. Randomly split S into Strain (say, 70% of the data) and Scv (the remain-
ing 30%). Here, Scv is called the hold-out cross validation set.

2. Train each model Mi on Strain only, to get some hypothesis hi .

3. Select and output the hypothesis hi that had the smallest error ε̂Scv (hi )
on the hold out cross validation set. (Recall, ε̂Scv (h) denotes the empir-
ical error of h on the set of examples in Scv .)

By testing on a set of examples Scv that the models were not trained on,
we obtain a better estimate of each hypothesis hi ’s true generalization error,
and can then pick the one with the smallest estimated generalization error.
Usually, somewhere between 1/4 − 1/3 of the data is used in the hold out
cross validation set, and 30% is a typical choice.
Optionally, step 3 in the algorithm may also be replaced with selecting
the model Mi according to arg mini ε̂Scv (hi ), and then retraining Mi on the
entire training set S. (This is often a good idea, with one exception being
learning algorithms that are be very sensitive to perturbations of the initial
conditions and/or data. For these methods, Mi doing well on Strain does not
necessarily mean it will also do well on Scv , and it might be better to forgo
this retraining step.)
The disadvantage of using hold out cross validation is that it “wastes”
about 30% of the data. Even if we were to take the optional step of retraining
 3

the model on the entire training set, it’s still as if we’re trying to find a good
model for a learning problem in which we had 0.7m training examples, rather
than m training examples, since we’re testing models that were trained on
only 0.7m examples each time. While this is fine if data is abundant and/or
cheap, in learning problems in which data is scarce (consider a problem with
m = 20, say), we’d like to do something better.
Here is a method, called k-fold cross validation, that holds out less
data each time:
1. Randomly split S into k disjoint subsets of m/k training examples each.
Lets call these subsets S1 , . . . , Sk .

2. For each model Mi , we evaluate it as follows:

For j = 1, . . . , k
Train the model Mi on S1 ∪ · · · ∪ Sj−1 ∪ Sj+1 ∪ · · · Sk (i.e., train
on all the data except Sj ) to get some hypothesis hij .
Test the hypothesis hij on Sj , to get ε̂Sj (hij ).
The estimated generalization error of model Mi is then calculated
as the average of the ε̂Sj (hij )’s (averaged over j).

3. Pick the model Mi with the lowest estimated generalization error, and
retrain that model on the entire training set S. The resulting hypothesis
is then output as our final answer.
A typical choice for the number of folds to use here would be k = 10.
While the fraction of data held out each time is now 1/k—much smaller
than before—this procedure may also be more computationally expensive
than hold-out cross validation, since we now need train to each model k
times.
While k = 10 is a commonly used choice, in problems in which data is
really scarce, sometimes we will use the extreme choice of k = m in order
to leave out as little data as possible each time. In this setting, we would
repeatedly train on all but one of the training examples in S, and test on that
held-out example. The resulting m = k errors are then averaged together to
obtain our estimate of the generalization error of a model. This method has
its own name; since we’re holding out one training example at a time, this
method is called leave-one-out cross validation.
Finally, even though we have described the different versions of cross vali-
dation as methods for selecting a model, they can also be used more simply to
evaluate a single model or algorithm. For example, if you have implemented
 4

some learning algorithm and want to estimate how well it performs for your
application (or if you have invented a novel learning algorithm and want to
report in a technical paper how well it performs on various test sets), cross
validation would give a reasonable way of doing so.

2 Feature Selection
One special and important case of model selection is called feature selection.
To motivate this, imagine that you have a supervised learning problem where
the number of features n is very large (perhaps n  m), but you suspect that
there is only a small number of features that are “relevant” to the learning
task. Even if you use the a simple linear classifier (such as the perceptron)
over the n input features, the VC dimension of your hypothesis class would
still be O(n), and thus overfitting would be a potential problem unless the
training set is fairly large.
In such a setting, you can apply a feature selection algorithm to reduce the
number of features. Given n features, there are 2n possible feature subsets
(since each of the n features can either be included or excluded from the
subset), and thus feature selection can be posed as a model selection problem
over 2n possible models. For large values of n, it’s usually too expensive to
explicitly enumerate over and compare all 2n models, and so typically some
heuristic search procedure is used to find a good feature subset. The following
search procedure is called forward search:

1. Initialize F = ∅.

2. Repeat {

(a) For i = 1, . . . , n if i 6∈ F, let Fi = F ∪ {i}, and use some ver-

sion of cross validation to evaluate features Fi . (I.e., train your
learning algorithm using only the features in Fi , and estimate its
generalization error.)
(b) Set F to be the best feature subset found on step (a).

3. Select and output the best feature subset that was evaluated during the
entire search procedure.
 5

The outer loop of the algorithm can be terminated either when F =

{1, . . . , n} is the set of all features, or when |F| exceeds some pre-set thresh-
old (corresponding to the maximum number of features that you want the
algorithm to consider using).
This algorithm described above one instantiation of wrapper model
feature selection, since it is a procedure that “wraps” around your learning
algorithm, and repeatedly makes calls to the learning algorithm to evaluate
how well it does using different feature subsets. Aside from forward search,
other search procedures can also be used. For example, backward search
starts off with F = {1, . . . , n} as the set of all features, and repeatedly deletes
features one at a time (evaluating single-feature deletions in a similar manner
to how forward search evaluates single-feature additions) until F = ∅.
Wrapper feature selection algorithms often work quite well, but can be
computationally expensive given how that they need to make many calls to
the learning algorithm. Indeed, complete forward search (terminating when
F = {1, . . . , n}) would take about O(n2 ) calls to the learning algorithm.
Filter feature selection methods give heuristic, but computationally
much cheaper, ways of choosing a feature subset. The idea here is to compute
some simple score S(i) that measures how informative each feature xi is about
the class labels y. Then, we simply pick the k features with the largest scores
S(i).
One possible choice of the score would be define S(i) to be (the absolute
value of) the correlation between xi and y, as measured on the training data.
This would result in our choosing the features that are the most strongly
correlated with the class labels. In practice, it is more common (particularly
for discrete-valued features xi ) to choose S(i) to be the mutual information
MI(xi , y) between xi and y:
X X p(xi , y)
MI(xi , y) = p(xi , y) log .
p(xi )p(y)
xi ∈{0,1} y∈{0,1}

(The equation above assumes that xi and y are binary-valued; more generally
the summations would be over the domains of the variables.) The probabil-
ities above p(xi , y), p(xi ) and p(y) can all be estimated according to their
empirical distributions on the training set.
To gain intuition about what this score does, note that the mutual infor-
mation can also be expressed as a Kullback-Leibler (KL) divergence:
MI(xi , y) = KL (p(xi , y)||p(xi )p(y))
You’ll get to play more with KL-divergence in Problem set #3, but infor-
mally, this gives a measure of how different the probability distributions
 6

p(xi , y) and p(xi )p(y) are. If xi and y are independent random variables,
then we would have p(xi , y) = p(xi )p(y), and the KL-divergence between the
two distributions will be zero. This is consistent with the idea if xi and y
are independent, then xi is clearly very “non-informative” about y, and thus
the score S(i) should be small. Conversely, if xi is very “informative” about
y, then their mutual information MI(xi , y) would be large.
One final detail: Now that you’ve ranked the features according to their
scores S(i), how do you decide how many features k to choose? Well, one
standard way to do so is to use cross validation to select among the possible
values of k. For example, when applying naive Bayes to text classification—
a problem where n, the vocabulary size, is usually very large—using this
method to select a feature subset often results in increased classifier accuracy.

3 Bayesian statistics and regularization

In this section, we will talk about one more tool in our arsenal for our battle
against overfitting.
At the beginning of the quarter, we talked about parameter fitting using
maximum likelihood (ML), and chose our parameters according to
m
Y
θML = arg max p(y (i) |x(i) ; θ).
θ
i=1

Throughout our subsequent discussions, we viewed θ as an unknown param-

eter of the world. This view of the θ as being constant-valued but unknown
is taken in frequentist statistics. In the frequentist this view of the world, θ
is not random—it just happens to be unknown—and it’s our job to come up
with statistical procedures (such as maximum likelihood) to try to estimate
this parameter.
An alternative way to approach our parameter estimation problems is to
take the Bayesian view of the world, and think of θ as being a random
variable whose value is unknown. In this approach, we would specify a
prior distribution p(θ) on θ that expresses our “prior beliefs” about the
parameters. Given a training set S = {(x(i) , y (i) )}m
i=1 , when we are asked to
make a prediction on a new value of x, we can then compute the posterior
 7

distribution on the parameters

p(S|θ)p(θ)
p(θ|S) =
p(S)
Qm (i) (i)


i=1 p(y |x , θ) p(θ)
= R Qm (1)
θ
( i=1 p(y (i) |x(i) , θ)p(θ)) dθ

In the equation above, p(y (i) |x(i) , θ) comes from whatever model you’re using
for your learning problem. For example, if you are using Bayesian logistic re-
(i) (i)
gression, then you might choose p(y (i) |x(i) , θ) = hθ (x(i) )y (1−hθ (x(i) ))(1−y ) ,
where hθ (x(i) ) = 1/(1 + exp(−θ T x(i) )).3
When we are given a new test example x and asked to make it prediction
on it, we can compute our posterior distribution on the class label using the
posterior distribution on θ:
Z
p(y|x, S) = p(y|x, θ)p(θ|S)dθ (2)
θ

In the equation above, p(θ|S) comes from Equation (1). Thus, for example,
if the goal is to the predict the expected value of y given x, then we would
output4 Z
E[y|x, S] = yp(y|x, S)dy
y

The procedure that we’ve outlined here can be thought of as doing “fully
Bayesian” prediction, where our prediction is computed by taking an average
with respect to the posterior p(θ|S) over θ. Unfortunately, in general it is
computationally very difficult to compute this posterior distribution. This is
because it requires taking integrals over the (usually high-dimensional) θ as
in Equation (1), and this typically cannot be done in closed-form.
Thus, in practice we will instead approximate the posterior distribution
for θ. One common approximation is to replace our posterior distribution for
θ (as in Equation 2) with a single point estimate. The MAP (maximum
a posteriori) estimate for θ is given by
m
Y
θMAP = arg max p(y (i) |x(i) , θ)p(θ). (3)
θ
i=1

3
Since we are now viewing θ as a random variable, it is okay to condition on it value,
and write “p(y|x, θ)” instead of “p(y|x; θ).”
4
The integral below would be replaced by a summation if y is discrete-valued.
 8

Note that this is the same formulas as for the ML (maximum likelihood)
estimate for θ, except for the prior p(θ) term at the end.
In practical applications, a common choice for the prior p(θ) is to assume
that θ ∼ N (0, τ 2 I). Using this choice of prior, the fitted parameters θMAP
will have smaller norm than that selected by maximum likelihood. (See
Problem Set #3.) In practice, this causes the Bayesian MAP estimate to be
less susceptible to overfitting than the ML estimate of the parameters. For
example, Bayesian logistic regression turns out to be an effective algorithm for
text classification, even though in text classification we usually have n  m.
 CS229 Lecture Notes
Andrew Ng and Kian Katanforoosh

Deep Learning
We now begin our study of deep learning. In this set of notes, we give an
overview of neural networks, discuss vectorization and discuss training neural
networks with backpropagation.

1 Neural Networks
We will start small and slowly build up a neural network, step by step. Recall
the housing price prediction problem from before: given the size of the house,
we want to predict the price.
Previously, we fitted a straight line to the graph. Now, instead of fitting a
straight line, we wish prevent negative housing prices by setting the absolute
minimum price as zero. This produces a “kink” in the graph as shown in
Figure 1.
Our goal is to input some input x into a function f (x) that outputs the
price of the house y. Formally, f : x → y. One of the simplest possible
neural networks is to define f (x) as a single “neuron” in the network where
f (x) = max(ax + b, 0), for some coefficients a, b. What f (x) does is return a
single value: (ax + b) or zero, whichever is greater. In the context of neural
networks, this function is called a ReLU (pronounced “ray-lu”), or rectified
linear unit. A more complex neural network may take the single neuron
described above and “stack” them together such that one neuron passes its
output as input into the next neuron, resulting in a more complex function.
Let us now deepen the housing prediction example. In addition to the size
of the house, suppose that you know the number of bedrooms, the zip code
Scribe: Albert Haque

1
 2

housing prices
1000

900

800

700
price (in $1000)
600

500

400

300

200

100

500 1000 1500 2000 2500 3000 3500 4000 4500 5000
square feet

Figure 1: Housing prices with a “kink” in the graph.

and the wealth of the neighborhood. Building neural networks is analogous

to Lego bricks: you take individual bricks and stack them together to build
complex structures. The same applies to neural networks: we take individual
neurons and stack them together to create complex neural networks.
Given these features (size, number of bedrooms, zip code, and wealth),
we might then decide that the price of the house depends on the maximum
family size it can accommodate. Suppose the family size is a function of
the size of the house and number of bedrooms (see Figure 2). The zip code
may provide additional information such as how walkable the neighborhood
is (i.e., can you walk to the grocery store or do you need to drive everywhere).
Combining the zip code with the wealth of the neighborhood may predict
the quality of the local elementary school. Given these three derived features
(family size, walkable, school quality), we may conclude that the price of the
home ultimately depends on these three features.
Size
Fam
ily S
# Bedrooms ize
Price
y
Walkable
Zip Code
ity
ol Qual
Scho
Wealth

Figure 2: Diagram of a small neural network for predicting housing prices.

 3

We have described this neural network as if you (the reader) already have
the insight to determine these three factors ultimately affect the housing
price. Part of the magic of a neural network is that all you need are the
input features x and the output y while the neural network will figure out
everything in the middle by itself. The process of a neural network learning
the intermediate features is called end-to-end learning.
Following the housing example, formally, the input to a neural network is
a set of input features x1 , x2 , x3 , x4 . We connect these four features to three
neurons. These three ”internal” neurons are called hidden units. The goal for
the neural network is to automatically determine three relevant features such
that the three features predict the price of a house. The only thing we must
provide to the neural network is a sufficient number of training examples
(x(i) , y (i) ). Often times, the neural network will discover complex features
which are very useful for predicting the output but may be difficult for a
human to understand since it does not have a “common” meaning. This is
why some people refer to neural networks as a black box, as it can be difficult
to understand the features it has invented.
Let us formalize this neural network representation. Suppose we have
three input features x1 , x2 , x3 which are collectively called the input layer,
four hidden units which are collectively called the hidden layer and one out-
put neuron called the output layer. The term hidden layer is called “hidden”
because we do not have the ground truth/training value for the hidden units.
This is in contrast to the input and output layers, both of which we know
the ground truth values from (x(i) , y (i) ).
The first hidden unit requires the input x1 , x2 , x3 and outputs a value
denoted by a1 . We use the letter a since it refers to the neuron’s “activation”
value. In this particular example, we have a single hidden layer but it is
[1]
possible to have multiple hidden layers. Let a1 denote the output value of
the first hidden unit in the first hidden layer. We use zero-indexing to refer
to the layer numbers. That is, the input layer is layer 0, the first hidden
layer is layer 1 and the output layer is layer 2. Again, more complex neural
networks may have more hidden layers. Given this mathematical notation,
[2]
the output of layer 2 is a1 . We can unify our notation:
[0]
x 1 = a1 (1.1)
[0]
x 2 = a2 (1.2)
[0]
x 3 = a3 (1.3)
To clarify, foo[1] with brackets denotes anything associated with layer 1, x(i)
[`]
with parenthesis refers to the ith training example, and aj refers to the
 4

activation of the j th unit in layer `. If we look at logistic regression g(x) as

a single neuron (see Figure 3):
1
g(x) =
1 + exp(−wT x)
The input to the logistic regression g(x) is three features x1 , x2 and x3 and it
outputs an estimated value of y. We can represent g(x) with a single neuron
in the neural network. We can break g(x) into two distinct computations:
(1) z = wT x + b and (2) a = σ(z) where σ(z) = 1/(1 + e−z ). Note the
notational difference: previously we used z = θT x but now we are using
z = wT x + b, where w is a vector. Later in these notes you will see capital W
to denote a matrix. The reasoning for this notational difference is conform
with standard neural network notation. More generally, a = g(z) where g(z)
is some activation function. Example activation functions include:
1
g(z) = (sigmoid) (1.4)
1 + e−z
g(z) = max(z, 0) (ReLU) (1.5)
ez − e−z
g(z) = z (tanh) (1.6)
e + e−z
In general, g(z) is a non-linear function.

x1

Estimated
x2
value of y

x3

Figure 3: Logistic regression as a single neuron.

Returning to our neural network from before, the first hidden unit in the first
hidden layer will perform the following computation:
[1] [1] T [1] [1] [1]
z1 = W1 x + b1 and a1 = g(z1 ) (1.7)
where W is a matrix of parameters and W1 refers to the first row of this
matrix. The parameters associated with the first hidden unit is the vector
 5

[1] [1]
W1 ∈ R3 and the scalar b1 ∈ R. For the second and third hidden units in
the first hidden layer, the computation is defined as:
[1] [1] T [1] [1] [1]
z2 = W2 x + b2 and a2 = g(z2 )
[1] [1] T [1] [1] [1]
z3 = W3 x + b3 and a3 = g(z3 )

where each hidden unit has its corresponding parameters W and b. Moving
on, the output layer performs the computation:
[2] [2] T [1] [2] [2] [2]
z1 = W1 a + b1 and a1 = g(z1 ) (1.8)

where a[1] is defined as the concatenation of all first layer activations:

 [1] 
a1
 [1] 
 a 
a[1] =  2[1]  (1.9)
 a3 
[1]
a4
[2]
The activation a1 from the second layer, which is a single scalar as defined by
[2] [2]
a1 = g(z1 ), represents the neural network’s final output prediction. Note
that for regression tasks, one typically does not apply a non-linear function
which is strictly positive (i.e., ReLU or sigmoid) because for some tasks, the
ground truth y value may in fact be negative.

2 Vectorization
In order to implement a neural network at a reasonable speed, one must be
careful when using for loops. In order to compute the hidden unit activations
in the first layer, we must compute z1 , ..., z4 and a1 , ..., a4 .
[1] [1] T [1] [1] [1]
z1 = W1 x + b1 and a1 = g(z1 ) (2.1)
.. .. ..
. . . (2.2)
[1] [1] T [1] [1] [1]
z4 = W4 x + b4 and a4 = g(z4 ) (2.3)

The most natural way to implement this in code is to use a for loop. One of
the treasures that deep learning has given to the field of machine learning is
that deep learning algorithms have high computational requirements. As a
result, code will run very slowly if you use for loops.
 6

This gave rise to vectorization. Instead of using for loops, vectorization

takes advantage of matrix algebra and highly optimized numerical linear
algebra packages (e.g., BLAS) to make neural network computations run
quickly. Before the deep learning era, a for loop may have been sufficient
on smaller datasets, but modern deep networks and state-of-the-art datasets
will be infeasible to run with for loops.

2.1 Vectorizing the Output Computation

We now present a method for computing z1 , ..., z4 without a for loop. Using
our matrix algebra, we can compute the activations:
[1] T
 [1]     [1] 
z1 — W1 —   b1
 .  T x1  b[1] 
 ..   — W2[1] — 
 
  x2  +  2 
 ..  =  ..  .  (2.4)
  
 .  .
  .. 
  x3
[1] [1] T [1]
z4 — W4 —
| {z } b4
3×1
| {z } | {z } x∈R | {z }
[1]
z ∈R 4×1 [1]
W ∈R 4×3 b ∈ R4×1
[1]

Where the Rd×n beneath each matrix indicates the dimensions. Expressing
this in matrix notation: z [1] = W [1] x + b[1] . To compute a[1] without a
for loop, we can leverage vectorized libraries in Matlab, Octave, or Python
which compute a[1] = g(z [1] ) very fast by performing parallel element-wise
operations. Mathematically, we defined the sigmoid function g(z) as:
1
g(z) = where z ∈ R (2.5)
1 + e−z
However, the sigmoid function can be defined not only for scalars but also
vectors. In a Matlab/Octave-like pseudocode, we can define the sigmoid as:

g(z) = 1 ./ (1+exp(-z)) where z ∈ Rd (2.6)

where ./ denotes element-wise division. With this vectorized implementa-

tion, a[1] = g(z [1] ) can be computed quickly.
To summarize the neural network so far, given an input x ∈ R3 , we com-
pute the hidden layer’s activations with z [1] = W [1] x + b[1] and a[1] = g(z [1] ).
To compute the output layer’s activations (i.e., neural network output):

z [2] = W
|{z}
[2] [1]
b[2]
a + |{z}
|{z} |{z} and a[2] = g(|{z}
|{z} z [2] ) (2.7)
1×1 1×4 4×1 1×1 1×1 1×1
 7

Why do we not use the identity function for g(z)? That is, why not use
g(z) = z? Assume for sake of argument that b[1] and b[2] are zeros. Using
Equation (2.7), we have:
z [2] = W [2] a[1] (2.8)
= W [2] g(z [1] ) by definition (2.9)
= W [2] z [1] since g(z) = z (2.10)
= W [2] W [1] x from Equation (2.4) (2.11)
[2] [1]
= W̃ x where W̃ = W W (2.12)
Notice how W [2] W [1] collapsed into W̃ . This is because applying a linear
function to another linear function will result in a linear function over the
original input (i.e., you can construct a W̃ such that W̃ x = W [2] W [1] x).
This loses much of the representational power of the neural network as often
times the output we are trying to predict has a non-linear relationship with
the inputs. Without non-linear activation functions, the neural network will
simply perform linear regression.

2.2 Vectorization Over Training Examples

Suppose you have a training set with three examples. The activations for
each example are as follows:
z [1](1) = W [1] x(1) + b[1]
z [1](2) = W [1] x(2) + b[1]
z [1](3) = W [1] x(3) + b[1]
Note the difference between square brackets [·], which refer to the layer num-
ber, and parenthesis (·), which refer to the training example number. In-
tuitively, one would implement this using a for loop. It turns out, we can
vectorize these operations as well. First, define:
 
| | |
X =  x(1) x(2) x(3)  (2.13)
| | |
Note that we are stacking training examples in columns and not rows. We
can then combine this into a single unified formulation:
 
| | |
Z [1] =  z [1](1) z [1](2) z [1](3)  = W [1] X + b[1] (2.14)
| | |
 8

You may notice that we are attempting to add b[1] ∈ R4×1 to W [1] X ∈
R4×3 . Strictly following the rules of linear algebra, this is not allowed. In
practice however, this addition is performed using broadcasting. We create
an intermediate b̃[1] ∈ R4×3 :
 
| | |
b̃[1] =  b[1] b[1] b[1]  (2.15)
| | |

We can then perform the computation: Z [1] = W [1] X + b̃[1] . Often times, it
is not necessary to explicitly construct b̃[1] . By inspecting the dimensions in
(2.14), you can assume b[1] ∈ R4×1 is correctly broadcast to W [1] X ∈ R4×3 .
Putting it together: Suppose we have a training set (x(1) , y (1) ), ..., (x(n) , y (n) )
where x(i) is a picture and y (i) is a binary label for whether the picture con-
tains a cat or not (i.e., 1=contains a cat). First, we initialize the parameters
W [1] , b[1] , W [2] , b[2] to small random numbers. For each example, we compute
the output “probability” from the sigmoid function a[2](i) . Second, using the
logistic regression log likelihood:
n 
X 
y (i) log a[2](i) + (1 − y (i) ) log(1 − a[2](i) ) (2.16)
i=1

Finally, we maximize this function using gradient ascent. This maximization

procedure corresponds to training the neural network.

3 Backpropagation
Instead of the housing example, we now have a new problem. Suppose we
wish to detect whether there is a soccer ball in an image or not. Given an
input image x(i) , we wish to output a binary prediction 1 if there is a ball in
the image and 0 otherwise.
Aside: Images can be represented as a matrix with number of elements
equal to the number of pixels. However, color images are digitally represented
as a volume (i.e., three-channels; or three matrices stacked on each other).
The number three is used because colors are represented as red-green-blue
(RGB) values. In the diagram below, we have a 64 × 64 × 3 image containing
a soccer ball. It is flattened into a single vector containing 12,288 elements.
A neural network model consists of two components: (i) the network
architecture, which defines how many layers, how many neurons, and how
the neurons are connected and (ii) the parameters (values; also known as
 9

weights). In this section, we will talk about how to learn the parameters.
First we will talk about parameter initialization, optimization and analyzing
these parameters.

3.1 Parameter Initialization

Consider a two layer neural network. On the left, the input is a flattened
(i)
image vector x(1) , ..., xd . In the first hidden layer, notice how all inputs are
connected to all neurons in the next layer. This is called a fully connected
layer.

The next step is to compute how many parameters are in this network. One
way of doing this is to compute the forward propagation by hand.
z [1] = W [1] x(i) + b[1] (3.1)
a[1] = g(z [1] ) (3.2)
z [2] = W [2] a[1] + b[2] (3.3)
[2] [2]
a = g(z ) (3.4)
[3] [3] [2] [3]
z =W a +b (3.5)
ŷ (i) = a[3] = g(z [3] ) (3.6)

We know that z [1] , a[1] ∈ R3×1 and z [2] , a[2] ∈ R2×1 and z [3] , a[3] ∈ R1×1 . As
of now, we do not know the size of W [1] . However, we can compute its size.
 10

We know that x ∈ Rd×1 . This leads us to the following

z [1] = W [1] x(i) = R3×1 Written as sizes: R3×1 = R?×? × Rd×1 (3.7)

Using matrix multiplication, we conclude that ?×? must be 3 × d. We also

conclude that the bias is of size 3 × 1 because its size must match W [1] x(i) .
We repeat this process for each hidden layer. This gives us:

W [2] ∈ R2×3 , b[2] ∈ R2×1 and W [3] ∈ R1×2 , b[3] ∈ R1×1 (3.8)

In total, we have 3n + 3 in the first layer, 2 × 3 + 2 in the second layer and

2 + 1 in the third layer. This gives us a total of 3n + 14 parameters.
Before we start training the neural network, we must select an initial
value for these parameters. We do not use the value zero as the initial value.
This is because the output of the first layer will always be the same since
W [1] x(i) + b[1] = 03×1 x(i) + 03×1 where 0d×n denotes a matrix of size n × m
filled with zeros. This will cause problems later on when we try to update
these parameters (i.e., the gradients will all be the same). The solution is to
randomly initialize the parameters to small values (e.g., normally distributed
around zero; N (0, 0.1)). Once the parameters have been initialized, we can
begin training the neural network with gradient descent.
The next step of the training process is to update the parameters. After a
single forward pass through the neural network, the output will be a predicted
value ŷ. We can then compute the loss L, in our case the log loss:
 
L(ŷ, y) = − (1 − y) log(1 − ŷ) + y log ŷ (3.9)

The loss function L(ŷ, y) produces a single scalar value. For short, we will
refer to the loss value as L. Given this value, we now must update all
parameters in layers of the neural network. For any given layer index `, we
update them:
∂L
W [`] = W [`] − α (3.10)
∂W [`]
∂L
b[`] = b[`] − α [`] (3.11)
∂b
where α is the learning rate. To proceed, we must compute the gradient with
respect to the parameters: ∂L/∂W [`] and ∂L/∂b[`] .
Remember, we made a decision to not set all parameters to zero. What if
we had initialized all parameters to be zero? We know that z [3] = W [3] a[2] +b[3]
 11

will evaluate to zero, because W [3] and b[3] are all zero. However, the output
of the neural network is defined as a[3] = g(z [3] ). Recall that g(·) is defined as
the sigmoid function. This means a[3] = g(0) = 0.5. Thus, no matter what
value of x(i) we provide, the network will output ŷ = 0.5.
What if we had initialized all parameters to be the same non-zero value?
In this case, consider the activations of the first layer:

a[1] = g(z [1] ) = g(W [1] x(i) + b[1] ) (3.12)

Each element of the activation vector a[1] will be the same (because W [1]
contains all the same values). This behavior will occur at all layers of the
neural network. As a result, when we compute the gradient, all neurons in
a layer will be equally responsible for anything contributed to the final loss.
We call this property symmetry. This means each neuron (within a layer)
will receive the exact same gradient update value (i.e., all neurons will learn
the same thing).
In practice, it turns out there is something better than random initializa-
tion. It is called Xavier/He initialization and initializes the weights:
r !
2
w[`] ∼ N 0, (3.13)
n[`] + n[`−1]

where n[`] is the number of neurons in layer `. This acts as a mini-normalization

technique. For a single layer, consider the variance of the input to the layer
as σ (in) and the variance of the output (i.e., activations) of a layer to be
σ (out) . Xavier/He initialization encourages σ (in) to be similar to σ (out) .

3.2 Optimization
Recall our neural network parameters: W [1] , b[1] , W [2] , b[2] , W [3] , b[3] . To up-
date them, we use stochastic gradient descent (SGD) using the update rules
in Equations (3.10) and (3.11). We will first compute the gradient with re-
spect to W [3] . The reason for this is that the influence of W [1] on the loss
is more complex than that of W [3] . This is because W [3] is “closer” to the
 12

output ŷ, in terms of number of computations.

 
∂L ∂
=− (1 − y) log(1 − ŷ) + y log ŷ (3.14)
∂W [3] ∂W [3]
 
∂ [3] [2] [3]
= −(1 − y) log 1 − g(W a + b ) (3.15)
∂W [3]
 
∂ [3] [2] [3]
−y log g(W a + b ) (3.16)
∂W [3]
1 0 [3] [2] [3] [2] T
= −(1 − y) (−1)g (W a + b )a (3.17)
1 − g(W [3] a[2] + b[3] )
1 T
−y [3] [2] [3]
g 0 (W [3] a[2] + b[3] )a[2] (3.18)
g(W a + b )
T T
= (1 − y)σ(W [3] a[2] + b[3] )a[2] − y(1 − σ(W [3] a[2] + b[3] ))a[2] (3.19)
T T
= (1 − y)a[3] a[2] − y(1 − a[3] )a[2] (3.20)
[2] T
= (a[3] − y)a (3.21)
Note that we are using sigmoid for g(·). The derivative of the sigmoid func-
tion: g 0 = σ 0 = σ(1 − σ). Additionally a[3] = σ(W [3] a[2] + b[3] ). At this point,
we have finished computing the gradient for one parameter, W [3] .
We will now compute the gradient for W [2] . Instead of deriving ∂L/∂W [2] ,
we can use the chain rule of calculus. We know that L depends on ŷ = a[3] .
∂L ∂L ?
[2]
= (3.22)
∂W ? ∂W [2]
If we look at the forward propagation, we know that loss L depends on
ŷ = a[3] . Using the chain rule, we can insert ∂a[3] /∂a[3] :
∂L ∂L ∂a[3] ?
= (3.23)
∂W [2] ∂a[3] ? ∂W [2]
We know that a[3] is directly related to z [3] .
∂L ∂L ∂a[3] ∂z [3] ?
= (3.24)
∂W [2] ∂a[3] ∂z [3] ? ∂W [2]
Furthermore we know that z [3] is directly related to a[2] . Note that we cannot
use W [2] or b[2] because a[2] is the only common element between Equations
(3.5) and (3.6). A common element is required for backpropagation.
∂L ∂L ∂a[3] ∂z [3] ∂a[2] ?
= (3.25)
∂W [2] ∂a[3] ∂z [3] ∂a[2] ? ∂W [2]
 13

Again, a[2] depends on z [2] , which z [2] directly depends on W [2] , which allows
us to complete the chain:
∂L ∂L ∂a[3] ∂z [3] ∂a[2] ∂z [2]
= (3.26)
∂W [2] ∂a[3] ∂z [3] ∂a[2] ∂z [2] ∂W [2]
Recall ∂L/∂W [3] :
∂L [3] [2] T
= (a − y)a (3.27)
∂W [3]
Since we computed ∂L/∂W [3] first, we know that a[2] = ∂z [3] /∂W [3] . Similarly
we have (a[3] − y) = ∂L/∂z [3] . These can help us compute ∂L/∂W [2] . We
substitute these values into Equation (3.26). This gives us:
∂L ∂L ∂a[3] ∂z [3] ∂a[2] ∂z [2]
= = (a[3] − y)W [3] g 0 (z [2] )a[1] (3.28)
∂W [2] |∂a[3]{z∂z [3]} |∂a [2] ∂z [2] ∂W [2]
{z } | {z } | {z }
(a[3] −y) W [3] g 0 (z [2] ) a[1]

While we have greatly simplified the process, we are not done yet. Because
we are computing derivatives in higher dimensions, the exact order of matrix
multiplication required to compute Equation (3.28) is not clear. We must
reorder the terms in Equation (3.28) such that the dimensions align. First,
we note the dimensions of all the terms:
∂L [3] 0 [2]
[2]
= (a[3] − y) W a[1]
g (z ) |{z} (3.29)
|∂W
|{z}
{z } | {z } | {z }
1×1 1×2 2×1 3×1
2×3

Notice how the terms do not align their shapes properly. We must rearrange
the terms by using properties of matrix algebra such that the matrix opera-
tions produce a result with the correct output shape. The correct ordering
is below:
∂L [3] T [1] T
[2]
=W ◦ g 0 (z [2] ) (a[3] − y) a (3.30)
|∂W
| {z } | {z } | {z } |{z}
{z } 2×1 2×1 1×1 1×3
2×3

We leave the remaining gradients as an exercise to the reader. In calculating

the gradients for the remaining parameters, it is important to use the inter-
mediate results we have computed for ∂L/∂W [2] and ∂L/∂W [3] , as these will
be directly useful for computing the gradient.
Returning to optimization, we previously discussed stochastic gradient
descent. Now we will talk about gradient descent. For any single layer `, the
update rule is defined as:
∂J
W [`] = W [`] − α (3.31)
∂W [`]
 14

n
1
L(i) and L(i) is the loss for a single exam-
P
where J is the cost function J = n
i=1
ple. The difference between the gradient descent update versus the stochastic
gradient descent version is that the cost function J gives more accurate gra-
dients whereas L(i) may be noisy. Stochastic gradient descent attempts to
approximate the gradient from (full) gradient descent. The disadvantage of
gradient descent is that it can be difficult to compute all activations for all
examples in a single forward or backwards propagation phase.
In practice, research and applications use mini-batch gradient descent.
This is a compromise between gradient descent and stochastic gradient de-
scent. In the case mini-batch gradient descent, the cost function Jmb is
defined as follows:
B
1 X (i)
Jmb = L (3.32)
B i=1
where B is the number of examples in the mini-batch.
There is another optimization method called momentum. Consider mini-
batch stochastic gradient. For any single layer `, the update rule is as follows:
(
∂J
vdW [`] = βvdW [`] + (1 − β) ∂W [`]
[`] [`]
(3.33)
W = W − αvdW [`]

Notice how there are now two stages instead of a single stage. The weight
update now depends on the cost J at this update step and the velocity vdW [`] .
The relative importance is controlled by β. Consider the analogy to a human
driving a car. While in motion, the car has momentum. If the car were to use
the brakes (or not push accelerator throttle), the car would continue moving
due to its momentum. Returning to optimization, the velocity vdW [`] will
keep track of the gradient over time. This technique has significantly helped
neural networks during the training phase.

3.3 Analyzing the Parameters

At this point, we have initialized the parameters and have optimized the
parameters. Suppose we evaluate the trained model and observe that it
achieves 96% accuracy on the training set but only 64% on the testing set.
Some solutions include: collecting more data, employing regularization, or
making the model shallower. Let us briefly look at regularization techniques.
 15

3.3.1 L2 Regularization
Let W below denote all the parameters in a model. In the case of neural
networks, you may think of applying the 2nd term to all layer weights W [`] .
For convenience, we simply write W . The L2 regularization adds another
term to the cost function:
λ
JL2 = J + ||W ||2 (3.34)
2
λX
=J+ |Wij |2 (3.35)
2 ij
λ
= J + WTW (3.36)
2
where J is the standard cost function from before, λ is an arbitrary value with
a larger value indicating more regularization and W contains all the weight
matrices, and where Equations (3.34), (3.35) and (3.36) are equivalent. The
update rule with L2 regularization becomes:
∂J λ ∂W T W
W =W −α −α (3.37)
∂W 2 ∂W
∂J
= (1 − αλ)W − α (3.38)
∂W
When we were updating our parameters using gradient descent, we did not
have the (1 − αλ)W term. This means with L2 regularization, every update
will include some penalization, depending on W . This penalization increases
the cost J, which encourages individual parameters to be small in magnitude,
which is a way to reduce overfitting.

3.3.2 Parameter Sharing

Recall logistic regression. It can be represented as a neural network, as
shown in Figure 3. The parameter vector θ = (θ1 , ..., θd ) must have the same
number of elements as the input vector x = (x1 , ..., xd ). In our image soccer
ball example, this means θ1 always looks at the top left pixel of the image
no matter what. However, we know that a soccer ball might appear in any
region of the image and not always the center. It is possible that θ1 was
never trained on a soccer ball in the top left of the image. As a result, during
test time, if an image of a soccer ball in the top left appears, the logistic
regression will likely predict no soccer ball. This is a problem.
This leads us to convolutional neural networks. Suppose θ is no longer a
vector but instead is a matrix. For our soccer ball example, suppose θ = R4×4 .
 16

For simplicity, we show the image as 64 × 64 but recall it is actually three-

dimensional and contains 3 channels. We now take our matrix of parameters

θ and slide it over the image. This is shown above by the thick square
in the upper left of the image. To compute the activation a, we compute
the element-wise product between θ and x1:4,1:4 , where the subscripts for x
indicate we are taking the top left 4 × 4 region in the image x. We then
collapse the matrix into a single scalar by summing all the elements resulting
from the element-wise product. Formally:
4 X
X 4
a= θij xij (3.39)
i=1 j=1

We then move this window slightly to the right in the image and repeat this
process. Once we have reached the end of the row, we start at the beginning
of the second row.

Once we have reached the end of the image, the parameters θ have “seen”
all pixels of the image: θ1 is no longer related to only the top left pixel. As a
result, whether the soccer ball appears in the bottom right or top left of the
image, the neural network will successfully detect the soccer ball.

Last Updated: April 24, 2019

CS229: Additional Notes on Backpropagation

1 Forward propagation
Recall that given input x, we define a[0] = x. Then for layer ` = 1, 2, . . . , N ,
where N is the number of layers of the network, we have
1. z [`] = W [`] a[`−1] + b[`]
2. a[`] = g [`] (z [`] )
In these notes we assume the nonlinearities g [`] are the same for all layers be-
sides layer N . This is because in the output layer we may be doing regression
[hence we might use g(x) = x] or binary classification [g(x) = sigmoid(x)] or
multiclass classification [g(x) = softmax(x)]. Hence we distinguish g [N ] from
g, and assume g is used for all layers besides layer N .
Finally, given the output of the network a[N ] , which we will more simply
denote as ŷ, we measure the loss J(W, b) = L(a[N ] , y) = L(ŷ, y). For example,
for real-valued regression we might use the squared loss
1
L(ŷ, y) = (ŷ − y)2
2
and for binary classification using logistic regression we use
L(ŷ, y) = −(y log ŷ + (1 − y) log(1 − ŷ))
or negative log-likelihood. Finally, for softmax regression over k classes, we
use the cross entropy loss
k
X
L(ŷ, y) = − 1{y = j} log ŷj
j=1

which is simply negative log-likelihood extended to the multiclass setting.

Note that ŷ is a k-dimensional vector in this case. If we use y to instead
denote the k-dimensional vector of zeros with a single 1 at the lth position,
where the true label is l, we can also express the cross entropy loss as
k
X
L(ŷ, y) = − yj log ŷj
j=1

1
 2

2 Backpropagation
Let’s define one more piece of notation that’ll be useful for backpropagation.1
We will define
δ [`] = ∇z[`] L(ŷ, y)
We can then define a three-step “recipe” for computing the gradients with
respect to every W [`] , b[`] as follows:
1. For output layer N , we have

δ [N ] = ∇z[N ] L(ŷ, y)

Sometimes we may want to compute ∇z[N ] L(ŷ, y) directly (e.g. if g [N ]

is the softmax function), whereas other times (e.g. when g [N ] is the
sigmoid function σ) we can apply the chain rule:

∇z[N ] L(ŷ, y) = ∇ŷ L(ŷ, y) ◦ (g [N ] )0 (z [N ] )

Note (g [N ] )0 (z [N ] ) denotes the elementwise derivative w.r.t. z [N ] .

2. For ` = N − 1, N − 2, . . . , 1, we have
>
δ [`] = (W [`+1] δ [`+1] ) ◦ g 0 (z [`] )

3. Finally, we can compute the gradients for layer ` as

>
∇W [`] J(W, b) = δ [`] a[`−1]
∇b[`] J(W, b) = δ [`]

where we use ◦ to indicate the elementwise product. Note the above proce-
dure is for a single training example.
You can try applying the above algorithm to logistic regression (N = 1,
g [1] is the sigmoid function σ) to sanity check steps (1) and (3). Recall that
σ 0 (z) = σ(z) ◦ (1 − σ(z)) and σ(z [1] ) is simply a[1] . Note that for logistic
regression, if x is a column vector in Rd×1 , then W [1] ∈ R1×d , and hence
∇W [1] J(W, b) ∈ R1×d . Example code for two layers is also given at:
http://cs229.stanford.edu/notes/backprop.py

1
These notes are closely adapted from:
http://ufldl.stanford.edu/tutorial/supervised/MultiLayerNeuralNetworks/
Scribe: Ziang Xie
 CS229 Lecture notes
Andrew Ng

The k-means clustering algorithm

In the clustering problem, we are given a training set {x(1) , . . . , x(m) }, and
want to group the data into a few cohesive “clusters.” Here, x(i) ∈ Rn
as usual; but no labels y (i) are given. So, this is an unsupervised learning
problem.
The k-means clustering algorithm is as follows:

1. Initialize cluster centroids µ1 , µ2 , . . . , µk ∈ Rn randomly.

2. Repeat until convergence: {

For every i, set

c(i) := arg min ||x(i) − µj ||2 .
j

For each j, set Pm (i)

i=1 1{c = j}x(i)
µj := P m (i) = j}
.
i=1 1{c

In the algorithm above, k (a parameter of the algorithm) is the number

of clusters we want to find; and the cluster centroids µj represent our current
guesses for the positions of the centers of the clusters. To initialize the cluster
centroids (in step 1 of the algorithm above), we could choose k training
examples randomly, and set the cluster centroids to be equal to the values of
these k examples. (Other initialization methods are also possible.)
The inner-loop of the algorithm repeatedly carries out two steps: (i)
“Assigning” each training example x(i) to the closest cluster centroid µj , and
(ii) Moving each cluster centroid µj to the mean of the points assigned to it.
Figure 1 shows an illustration of running k-means.

1
 2

(a) (b) (c)

(d) (e) (f)

Figure 1: K-means algorithm. Training examples are shown as dots, and

cluster centroids are shown as crosses. (a) Original dataset. (b) Random ini-
tial cluster centroids (in this instance, not chosen to be equal to two training
examples). (c-f) Illustration of running two iterations of k-means. In each
iteration, we assign each training example to the closest cluster centroid
(shown by “painting” the training examples the same color as the cluster
centroid to which is assigned); then we move each cluster centroid to the
mean of the points assigned to it. (Best viewed in color.) Images courtesy
Michael Jordan.

Is the k-means algorithm guaranteed to converge? Yes it is, in a certain

sense. In particular, let us define the distortion function to be:
X
m
J(c, µ) = ||x(i) − µc(i) ||2
i=1

Thus, J measures the sum of squared distances between each training exam-
ple x(i) and the cluster centroid µc(i) to which it has been assigned. It can
be shown that k-means is exactly coordinate descent on J. Specifically, the
inner-loop of k-means repeatedly minimizes J with respect to c while holding
µ fixed, and then minimizes J with respect to µ while holding c fixed. Thus,
J must monotonically decrease, and the value of J must converge. (Usu-
ally, this implies that c and µ will converge too. In theory, it is possible for
 3

k-means to oscillate between a few different clusterings—i.e., a few different

values for c and/or µ—that have exactly the same value of J, but this almost
never happens in practice.)
The distortion function J is a non-convex function, and so coordinate
descent on J is not guaranteed to converge to the global minimum. In other
words, k-means can be susceptible to local optima. Very often k-means will
work fine and come up with very good clusterings despite this. But if you
are worried about getting stuck in bad local minima, one common thing to
do is run k-means many times (using different random initial values for the
cluster centroids µj ). Then, out of all the different clusterings found, pick
the one that gives the lowest distortion J(c, µ).
 CS229 Lecture notes
Andrew Ng

Mixtures of Gaussians and the EM algorithm

In this set of notes, we discuss the EM (Expectation-Maximization) for den-
sity estimation.
Suppose that we are given a training set {x(1) , . . . , x(m) } as usual. Since
we are in the unsupervised learning setting, these points do not come with
any labels.
We wish to model the data by specifying a joint distribution P p(x(i) , z (i) ) =
p(x |z )p(z ). Here, z ∼ Multinomial(φ) (where φj ≥ 0, kj=1 φj = 1,
(i) (i) (i) (i)

and the parameter φj gives p(z (i) = j),), and x(i) |z (i) = j ∼ N (µj , Σj ). We
let k denote the number of values that the z (i) ’s can take on. Thus, our
model posits that each x(i) was generated by randomly choosing z (i) from
{1, . . . , k}, and then x(i) was drawn from one of k Gaussians depeneding on
z (i) . This is called the mixture of Gaussians model. Also, note that the
z (i) ’s are latent random variables, meaning that they’re hidden/unobserved.
This is what will make our estimation problem difficult.
The parameters of our model are thus φ, φ and Σ. To estimate them, we
can write down the likelihood of our data:
X
m
ℓ(φ, µ, Σ) = log p(x(i) ; φ, µ, Σ)
i=1
Xm X
k
= log p(x(i) |z (i) ; µ, Σ)p(z (i) ; φ).
i=1 z (i) =1

However, if we set to zero the derivatives of this formula with respect to

the parameters and try to solve, we’ll find that it is not possible to find the
maximum likelihood estimates of the parameters in closed form. (Try this
yourself at home.)
The random variables z (i) indicate which of the k Gaussians each x(i)
had come from. Note that if we knew what the z (i) ’s were, the maximum

1
 2

likelihood problem would have been easy. Specifically, we could then write
down the likelihood as
X
m
ℓ(φ, µ, Σ) = log p(x(i) |z (i) ; µ, Σ) + log p(z (i) ; φ).
i=1

Maximizing this with respect to φ, µ and Σ gives the parameters:

1 X
m
φj = 1{z (i) = j},
m i=1
Pm (i)
i=1 1{z = j}x(i)
µj = P m (i) = j}
,
i=1 1{z
Pm (i)
i=1 1{z P = j}(x(i) − µj )(x(i) − µj )T
Σj = m (i) = j}
.
i=1 1{z

Indeed, we see that if the z (i) ’s were known, then maximum likelihood
estimation becomes nearly identical to what we had when estimating the
parameters of the Gaussian discriminant analysis model, except that here
the z (i) ’s playing the role of the class labels.1
However, in our density estimation problem, the z (i) ’s are not known.
What can we do?
The EM algorithm is an iterative algorithm that has two main steps.
Applied to our problem, in the E-step, it tries to “guess” the values of the
z (i) ’s. In the M-step, it updates the parameters of our model based on our
guesses. Since in the M-step we are pretending that the guesses in the first
part were correct, the maximization becomes easy. Here’s the algorithm:

Repeat until convergence: {

(E-step) For each i, j, set

(i)
wj := p(z (i) = j|x(i) ; φ, µ, Σ)
1
There are other minor differences in the formulas here from what we’d obtained in
PS1 with Gaussian discriminant analysis, first because we’ve generalized the z (i) ’s to be
multinomial rather than Bernoulli, and second because here we are using a different Σj
for each Gaussian.
 3

(M-step) Update the parameters:

1 X (i)
m
φj := w ,
m i=1 j
Pm (i) (i)
i=1 wj x
µj := Pm (i) ,
i=1 wj
Pm (i) (i)
i=1 wj (x − µj )(x(i) − µj )T
Σj := Pm (i)
i=1 wj

In the E-step, we calculate the posterior probability of our parameters

the z (i) ’s, given the x(i) and using the current setting of our parameters. I.e.,
using Bayes rule, we obtain:

p(x(i) |z (i) = j; µ, Σ)p(z (i) = j; φ)

p(z (i) = j|x(i) ; φ, µ, Σ) = Pk
(i) (i) = l; µ, Σ)p(z (i) = l; φ)
l=1 p(x |z

Here, p(x(i) |z (i) = j; µ, Σ) is given by evaluating the density of a Gaussian

with mean µj and covariance Σj at x(i) ; p(z (i) = j; φ) is given by φj , and so
(i)
on. The values wj calculated in the E-step represent our “soft” guesses2 for
the values of z (i) .
Also, you should contrast the updates in the M-step with the formulas we
had when the z (i) ’s were known exactly. They are identical, except that in-
stead of the indicator functions “1{z (i) = j}” indicating from which Gaussian
(i)
each datapoint had come, we now instead have the wj ’s.
The EM-algorithm is also reminiscent of the K-means clustering algo-
rithm, except that instead of the “hard” cluster assignments c(i), we instead
(i)
have the “soft” assignments wj . Similar to K-means, it is also susceptible
to local optima, so reinitializing at several different initial parameters may
be a good idea.
It’s clear that the EM algorithm has a very natural interpretation of
repeatedly trying to guess the unknown z (i) ’s; but how did it come about,
and can we make any guarantees about it, such as regarding its convergence?
In the next set of notes, we will describe a more general view of EM, one
2
The term “soft” refers to our guesses being probabilities and taking values in [0, 1]; in
contrast, a “hard” guess is one that represents a single best guess (such as taking values
in {0, 1} or {1, . . . , k}).
 4

that will allow us to easily apply it to other estimation problems in which

there are also latent variables, and which will allow us to give a convergence
guarantee.
 CS229 Lecture notes
Tengyu Ma and Andrew Ng
May 13, 2019

Part IX
The EM algorithm
In the previous set of notes, we talked about the EM algorithm as applied to
fitting a mixture of Gaussians. In this set of notes, we give a broader view
of the EM algorithm, and show how it can be applied to a large family of
estimation problems with latent variables. We begin our discussion with a
very useful result called Jensen’s inequality

1 Jensen’s inequality
Let f be a function whose domain is the set of real numbers. Recall that
f is a convex function if f ′′ (x) ≥ 0 (for all x ∈ R). In the case of f taking
vector-valued inputs, this is generalized to the condition that its hessian H
is positive semi-definite (H ≥ 0). If f ′′ (x) > 0 for all x, then we say f is
strictly convex (in the vector-valued case, the corresponding statement is
that H must be positive definite, written H > 0). Jensen’s inequality can
then be stated as follows:
Theorem. Let f be a convex function, and let X be a random variable.
Then:
E[f (X)] ≥ f (EX).
Moreover, if f is strictly convex, then E[f (X)] = f (EX) holds true if and
only if X = E[X] with probability 1 (i.e., if X is a constant).
Recall our convention of occasionally dropping the parentheses when writ-
ing expectations, so in the theorem above, f (EX) = f (E[X]).
For an interpretation of the theorem, consider the figure below.

1
 2

f(a) f

E[f(X)]

f(b)
f(EX)

a E[X] b

Here, f is a convex function shown by the solid line. Also, X is a random

variable that has a 0.5 chance of taking the value a, and a 0.5 chance of
taking the value b (indicated on the x-axis). Thus, the expected value of X
is given by the midpoint between a and b.
We also see the values f (a), f (b) and f (E[X]) indicated on the y-axis.
Moreover, the value E[f (X)] is now the midpoint on the y-axis between f (a)
and f (b). From our example, we see that because f is convex, it must be the
case that E[f (X)] ≥ f (EX).
Incidentally, quite a lot of people have trouble remembering which way
the inequality goes, and remembering a picture like this is a good way to
quickly figure out the answer.
Remark. Recall that f is [strictly] concave if and only if −f is [strictly]
convex (i.e., f ′′ (x) ≤ 0 or H ≤ 0). Jensen’s inequality also holds for concave
functions f , but with the direction of all the inequalities reversed (E[f (X)] ≤
f (EX), etc.).

2 The EM algorithm
Suppose we have an estimation problem in which we have a training set
{x(1) , . . . , x(n) } consisting of n independent examples. We have a latent vari-
able model p(x, z; θ) with z being the latent variable (which for simplicity is
assumed to take finite number of values). The density for x can be obtained
by marginalized over the latent variable z:
X
p(x; θ) = p(x, z; θ) (1)
z
 3

We wish to fit the parameters θ by maximizing the log-likelihood of the

data, defined by
n
X
ℓ(θ) = log p(x(i) ; θ) (2)
i=1

We can rewrite the objective in terms of the joint density p(x, z; θ) by

n
X
ℓ(θ) = log p(x(i) ; θ) (3)
i=1
n
X X
= log p(x(i) , z (i) ; θ). (4)
i=1 z (i)

But, explicitly finding the maximum likelihood estimates of the parameters

θ may be hard since it will result in difficult non-convex optimization prob-
lems.1 Here, the z (i) ’s are the latent random variables; and it is often the case
that if the z (i) ’s were observed, then maximum likelihood estimation would
be easy.
In such a setting, the EM algorithm gives an efficient method for max-
imum likelihood estimation. Maximizing ℓ(θ) explicitly might be difficult,
and our strategy will be to instead repeatedly construct a lower-bound on ℓ
(E-step), and then optimize that lower-bound
Pn (M-step).2
It turns out that the summation i=1 is not essential here, and towards a
simpler exposition of the EM algorithm, we will first consider optimizing the
the likelihood log p(x) for a single example x. After we derive the algorithm
for optimizing log p(x), we will convert it to an algorithm that works for n
examples by adding back the sum to each of the relevant equations. Thus,
now we aim to optimize log p(x; θ) which can be rewritten as
X
log p(x; θ) = log p(x, z; θ) (5)
z

1
It’s mostly an empirical observation that the optimization problem is difficult to op-
timize.
2
Empirically, the E-step and M-step can often be computed more efficiently than op-
timizing the function ℓ(·) directly. However, it doesn’t necessarily mean that alternating
the two steps can always converge to the global optimum of ℓ(·). Even for mixture of
Gaussians, the EM algorithm can either converge to a global optimum or get stuck, de-
pending on the properties of the training data. Empirically, for real-world data, often EM
can converge to a solution with relatively high likelihood (if not the optimum), and the
theory behind it is still largely not understood.
 4

P
Let Q be a distribution over the possible values of z. That is, z Q(z) = 1,
Q(z) ≥ 0).
Consider the following:3
X
log p(x; θ) = log p(x, z; θ)
z
X p(x, z; θ)
= log Q(z) (6)
z
Q(z)
X p(x, z; θ)
≥ Q(z) log (7)
z
Q(z)

The last step of this derivation used Jensen’s inequality. Specifically,

f (x) = log x is a concave function, since f ′′ (x) = −1/x2 < 0 over its domain
x ∈ R+ . Also, the term
 
X p(x, z; θ)
Q(z)
z
Q(z)

in the summation is just an expectation of the quantity [p(x, z; θ)/Q(z)] with

respect to z drawn according to the distribution given by Q.4 By Jensen’s
inequality, we have
     
p(x, z; θ) p(x, z; θ)
f Ez∼Q ≥ Ez∼Q f ,
Q(z) Q(z)

where the “z ∼ Q” subscripts above indicate that the expectations are with
respect to z drawn from Q. This allowed us to go from Equation (6) to
Equation (7).
Now, for any distribution Q, the formula (7) gives a lower-bound on
log p(x; θ). There are many possible choices for the Q’s. Which should we
choose? Well, if we have some current guess θ of the parameters, it seems
natural to try to make the lower-bound tight at that value of θ. I.e., we will
make the inequality above hold with equality at our particular value of θ.
To make the bound tight for a particular value of θ, we need for the step
involving Jensen’s inequality in our derivation above to hold with equality.
3
If z were continuous, then Q would be a density, and the summations over z in our
discussion are replaced with integrals over z.
4
We note that the notion p(x,z;θ)
Q(z) only makes sense if Q(z) 6= 0 whenever p(x, z; θ) 6= 0.
Here we implicitly assume that we only consider those Q with such a property.
 5

For this to be true, we know it is sufficient that the expectation be taken

over a “constant”-valued random variable. I.e., we require that

p(x, z; θ)
=c
Q(z)

for some constant c that does not depend on z. This is easily accomplished
by choosing
Q(z) ∝ p(x, z; θ).
P
Actually, since we know z Q(z) = 1 (because it is a distribution), this
further tells us that
p(x, z; θ)
Q(z) = P
z p(x, z; θ)
p(x, z; θ)
=
p(x; θ)
= p(z|x; θ) (8)

Thus, we simply set the Q’s to be the posterior distribution of the z’s given
x and the setting of the parameters θ.
Indeed, we can directly verify that when Q(z) = p(z|x; θ), then equa-
tion (7) is an equality because
X p(x, z; θ) X p(x, z; θ)
Q(z) log = p(z|x; θ) log
z
Q(z) z
p(z|x; θ)
X p(z|x; θ)p(x; θ)
= p(z|x; θ) log
z
p(z|x; θ)
X
= p(z|x; θ) log p(x; θ)
z
X
= log p(x; θ) p(z|x; θ)
z
P
= log p(x; θ) (because z p(z|x; θ) = 1)

For convenience, we call the expression in Equation (7) the evidence

lower bound (ELBO) and we denote it by
X p(x, z; θ)
ELBO(x; Q, θ) = Q(z) log (9)
z
Q(z)
 6

With this equation, we can re-write equation (7) as

∀Q, θ, x, log p(x; θ) ≥ ELBO(x; Q, θ) (10)
Intuitively, the EM algorithm alternatively updates Q and θ by a) setting
Q(z) = p(z|x; θ) following Equation (8) so that ELBO(x; Q, θ) = log p(x; θ)
for x and the current θ, and b) maximizing ELBO(x; Q, θ) w.r.t θ while fixing
the choice of Q.
Recall that all the discussion above was under the assumption that we
aim to optimize the log-likelihood log p(x; θ) for a single example x. It turns
out that with multiple training examples, the basic idea is the same and we
only needs to take a sum over examples at relevant places. Next, we will
build the evidence lower bound for multiple training examples and make the
EM algorithm formal.
Recall we have a training set {x(1) , . . . , x(n) }. Note that the optimal choice
of Q is p(z|x; θ), and it depends on the particular example x. Therefore here
we will introduce n distributions Q1 , . . . , Qn , one for each example x(i) . For
each example x(i) , we can build the evidence lower bound
X p(x(i) , z (i) ; θ)
log p(x(i) ; θ) ≥ ELBO(x(i) ; Qi , θ) = Qi (z (i) ) log
Qi (z (i) )
z (i)

Taking sum over all the examples, we obtain a lower bound for the log-
likelihood
X
ℓ(θ) ≥ ELBO(x(i) ; Qi , θ) (11)
i
XX p(x(i) , z (i) ; θ)
= Qi (z (i) ) log
i
Qi (z (i) )
z (i)

For any set of distributions Q1 , . . . , Qn , the formula (11) gives a lower-

bound on ℓ(θ), and analogous to the argument around equation (8), the Qi
that attains equality satisfies
Qi (z (i) ) = p(z (i) |x(i) ; θ)
Thus, we simply set the Qi ’s to be the posterior distribution of the z (i) ’s
given x(i) with the current setting of the parameters θ.
Now, for this choice of the Qi ’s, Equation (11) gives a lower-bound on
the loglikelihood ℓ that we’re trying to maximize. This is the E-step. In
the M-step of the algorithm, we then maximize our formula in Equation (11)
with respect to the parameters to obtain a new setting of the θ’s. Repeatedly
carrying out these two steps gives us the EM algorithm, which is as follows:
 7

Repeat until convergence {

(E-step) For each i, set
Qi (z (i) ) := p(z (i) |x(i) ; θ).

(M-step) Set
n
X
θ := arg max ELBO(x(i) ; Qi , θ)
θ
i=1
XX p(x(i) , z (i) ; θ)
= arg max Qi (z (i) ) log . (12)
θ
i
Qi (z (i) )
z (i)

}
How do we know if this algorithm will converge? Well, suppose θ(t) and
θ(t+1) are the parameters from two successive iterations of EM. We will now
prove that ℓ(θ(t) ) ≤ ℓ(θ(t+1) ), which shows EM always monotonically im-
proves the log-likelihood. The key to showing this result lies in our choice of
the Qi ’s. Specifically, on the iteration of EM in which the parameters had
(t)
started out as θ(t) , we would have chosen Qi (z (i) ) := p(z (i) |x(i) ; θ(t) ). We
saw earlier that this choice ensures that Jensen’s inequality, as applied to get
Equation (11), holds with equality, and hence
n
(t)
X
ℓ(θ(t) ) = ELBO(x(i) ; Qi , θ(t) ) (13)
i=1

The parameters θ(t+1) are then obtained by maximizing the right hand side
of the equation above. Thus,
n
(t)
X
(t+1)
ℓ(θ )≥ ELBO(x(i) ; Qi , θ(t+1) )
i=1
(because ineqaulity (11) holds for all Q and θ)
n
(t)
X
≥ ELBO(x(i) ; Qi , θ(t) ) (see reason below)
i=1
(t)
= ℓ(θ ) (by equation (13))

where the last inequality follows from that θ(t+1) is chosen explicitly to be
n
(t)
X
arg max ELBO(x(i) ; Qi , θ)
θ
i=1
 8

Hence, EM causes the likelihood to converge monotonically. In our de-

scription of the EM algorithm, we said we’d run it until convergence. Given
the result that we just showed, one reasonable convergence test would be
to check if the increase in ℓ(θ) between successive iterations is smaller than
some tolerance parameter, and to declare convergence if EM is improving
ℓ(θ) too slowly.
Remark. If we define (by overloading ELBO(·))
n
X XX p(x(i) , z (i) ; θ)
ELBO(Q, θ) = ELBO(x(i) ; Qi , θ) = Qi (z (i) ) log
i=1 i
Qi (z (i) )
z (i)
(14)
then we know ℓ(θ) ≥ ELBO(Q, θ) from our previous derivation. The EM
can also be viewed an alternating maximization algorithm on ELBO(Q, θ),
in which the E-step maximizes it with respect to Q (check this yourself), and
the M-step maximizes it with respect to θ.

2.1 Other interpretation of ELBO

Let ELBO(x; Q, θ) = z Q(z) log p(x,z;θ)
P
Q(z)
be defined as in equation (9). There
are several other forms of ELBO. First, we can rewrite
ELBO(x; Q, θ) = Ez∼Q [log p(x, z; θ)] − Ez∼Q [log Q(z)]
= Ez∼Q [log p(x|z; θ)] − DKL (Qkpz ) (15)
where we use pz to denote the marginal distribution of z (under the distri-
bution p(x, z; θ)), and DKL () denotes the KL divergence
X Q(z)
DKL (Qkpz ) = Q(z) log (16)
z
p(z)
In many cases, the marginal distribution of z does not depend on the param-
eter θ. In this case, we can see that maximizing ELBO over θ is equivalent
to maximizing the first term in (15). This corresponds to maximizing the
conditional likelihood of x conditioned on z, which is often a simpler question
than the original question.
Another form of ELBO(·) is (please verify yourself)
ELBO(x; Q, θ) = log p(x) − DKL (Qkpz|x ) (17)
where pz|x is the conditional distribution of z given x under the parameter
θ. This forms shows that the maximizer of ELBO(Q, θ) over Q is obtained
when Q = pz|x , which was shown in equation (8) before.
 9

3 Mixture of Gaussians revisited

Armed with our general definition of the EM algorithm, let’s go back to our
old example of fitting the parameters φ, µ and Σ in a mixture of Gaussians.
For the sake of brevity, we carry out the derivations for the M-step updates
only for φ and µj , and leave the updates for Σj as an exercise for the reader.
The E-step is easy. Following our algorithm derivation above, we simply
calculate
(i)
wj = Qi (z (i) = j) = P (z (i) = j|x(i) ; φ, µ, Σ).
Here, “Qi (z (i) = j)” denotes the probability of z (i) taking the value j under
the distribution Qi .
Next, in the M-step, we need to maximize, with respect to our parameters
φ, µ, Σ, the quantity
n X
X p(x(i) , z (i) ; φ, µ, Σ)
Qi (z (i) ) log
i=1 z (i)
Qi (z (i) )
n X
k
X p(x(i) |z (i) = j; µ, Σ)p(z (i) = j; φ)
= Qi (z (i) = j) log
i=1 j=1
Qi (z (i) = j)
1
exp − 12 (x(i) − µj )T Σ−1 (i)


n X
X k
(i) (2π)d/2 |Σj |1/2 j (x − µj ) · φ j
= wj log (i)
i=1 j=1 wj
Let’s maximize this with respect to µl . If we take the derivative with respect
to µl , we find
1
exp − 12 (x(i) − µj )T Σ−1 (i)


X n X k
(i) (2π)d/2 |Σj |1/2 j (x − µj ) · φ j
∇ µl wj log (i)
i=1 j=1 wj
n X
k
(i) 1
X
= −∇µl wj (x(i) − µj )T Σ−1
j (x
(i)
− µj )
i=1 j=1
2
n
1 X (i)
= wl ∇µl 2µTl Σ−1
l x
(i)
− µTl Σ−1
l µl
2 i=1
n
(i)
X
Σ−1 (i)
− Σ−1


= wl l x l µl
i=1

Setting this to zero and solving for µl therefore yields the update rule
Pn (i) (i)
i=1 wl x
µl := Pn (i)
,
i=1 wl
 10

which was what we had in the previous set of notes.

Let’s do one more example, and derive the M-step update for the param-
eters φj . Grouping together only the terms that depend on φj , we find that
we need to maximize
n X k
(i)
X
wj log φj .
i=1 j=1

However, there is an additional constraint that the φj ’s sum to 1, since they

represent the probabilities φj = p(z (i) = j; φ). To deal with the constraint
that kj=1 φj = 1, we construct the Lagrangian
P

n X
k k
(i)
X X
L(φ) = wj log φj + β( φj − 1),
i=1 j=1 j=1

where β is the Lagrange multiplier.5 Taking derivatives, we find

n (i)
∂ X wj
L(φ) = +β
∂φj i=1
φj

Setting this to zero and solving, we get

Pn (i)
i=1wj
φj =
−β
Pn (i) P
I.e., φj ∝ i=1 wj . Using the constraint that j φj = 1, we easily find
Pn P k (i) Pn (i)
that −β = i=1 j=1 wj = (This used the fact that wj =
i=1 1 = n.
(i)
P (i)
Qi (z = j), and since probabilities sum to 1, j wj = 1.) We therefore
have our M-step updates for the parameters φj :
n
1 X (i)
φj := w .
n i=1 j

The derivation for the M-step updates to Σj are also entirely straightfor-
ward.
5
We don’t need to worry about the constraint that φj ≥ 0, because as we’ll shortly see,
the solution we’ll find from this derivation will automatically satisfy that anyway.
 11

4 Variational inference and variational auto-

encoder
Loosely speaking, variational auto-encoder [2] generally refers to a family of
algorithms that extend the EM algorithms to more complex models parame-
terized by neural networks. It extends the technique of variational inference
with the additional “re-parametrization trick” which will be introduced be-
low. Variational auto-encoder may not give the best performance for many
datasets, but it contains several central ideas about how to extend EM algo-
rithms to high-dimensional continuous latent variables with non-linear mod-
els. Understanding it will likely give you the language and backgrounds to
understand various recent papers related to it.
As a running example, we will consider the following parameterization of
p(x, z; θ) by a neural network. Let θ be the collection of the weights of a
neural network g(z; θ) that maps z ∈ Rk to Rd . Let
z ∼ N (0, Ik×k ) (18)
x|z ∼ N (g(z; θ), σ 2 Id×d ) (19)
Here Ik×k denotes identity matrix of dimension k by k, and σ is a scalar that
we assume to be known for simplicity.
For the Gaussian mixture models in Section 3, the optimal choice of
Q(z) = p(z|x; θ) for each fixed θ, that is the posterior distribution of z,
can be analytically computed. In many more complex models such as the
model (19), it’s intractable to compute the exact the posterior distribution
p(z|x; θ).
Recall that from equation (10), ELBO is always a lower bound for any
choice of Q, and therefore, we can also aim for finding an approximation of
the true posterior distribution. Often, one has to use some particular form
to approximate the true posterior distribution. Let Q be a family of Q’s that
we are considering, and we will aim to find a Q within the family of Q that is
closest to the true posterior distribution. To formalize, recall the definition
of the ELBO lower bound as a function of Q and θ defined in equation (14)
n
X XX p(x(i) , z (i) ; θ)
ELBO(Q, θ) = ELBO(x(i) ; Qi , θ) = Qi (z (i) ) log
i=1 i
Qi (z (i) )
z (i)

Recall that EM can be viewed as alternating maximization of ELBO(Q, θ).

Here instead, we optimize the EBLO over Q ∈ Q
max max ELBO(Q, θ) (20)
Q∈Q θ
 12

Now the next question is what form of Q (or what structural assumptions
to make about Q) allows us to efficiently maximize the objective above. When
the latent variable z are high-dimensional discrete variables, one popular as-
sumption is the mean field assumption, which assumes that Qi (z) gives a
distribution with independent coordinates, or in other words, Qi can be de-
composed into Qi (z) = Q1i (z1 ) · · · Qki (zk ). There are tremendous applications
of mean field assumptions to learning generative models with discrete latent
variables, and we refer to [1] for a survey of these models and their impact
to a wide range of applications including computational biology, computa-
tional neuroscience, social sciences. We will not get into the details about
the discrete latent variable cases, and our main focus is to deal with contin-
uous latent variables, which requires not only mean field assumptions, but
additional techniques.
When z ∈ Rk is a continuous latent variable, there are several decisions to
make towards successfully optimizing (20). First we need to give a succinct
representation of the distribution Qi because it is over an infinite number of
points. A natural choice is to assume Qi is a Gaussian distribution with some
mean and variance. We would also like to have more succinct representation
of the means of Qi of all the examples. Note that Qi (z (i) ) is supposed to
approximate p(z (i) |x(i) ; θ). It would make sense let all the means of the Qi ’s
be some function of x(i) . Concretely, let q(·; φ), v(·; φ) be two functions that
map from dimension d to k, which are parameterized by φ and ψ, we assume
that
Qi = N (q(x(i) ; φ), diag(v(x(i) ; ψ))2 ) (21)
Here diag(w) means the k × k matrix with the entries of w ∈ Rk on the
diagonal. In other words, the distribution Qi is assumed to be a Gaussian
distribution with independent coordinates, and the mean and standard de-
viations are governed by q and v. Often in variational auto-encoder, q and v
are chosen to be neural networks.6 In recent deep learning literature, often
q, v are called encoder (in the sense of encoding the data into latent code),
whereas g(z; θ) if often referred to as the decoder.
We remark that Qi of such form in many cases are very far from a good
approximation of the true posterior distribution. However, some approxima-
tion is necessary for feasible optimization. In fact, the form of Qi needs to
satisfy other requirements (which happened to be satisfied by the form (21))
Before optimizing the ELBO, let’s first verify whether we can efficiently
evaluate the value of the ELBO for fixed Q of the form (21) and θ. We
6
q and v can also share parameters. We sweep this level of details under the rug in this
note.
 13

rewrite the ELBO as a function of φ, ψ, θ by

n
p(x(i) , z (i) ; θ)
X  
ELBO(φ, ψ, θ) = Ez(i) ∼Qi log (i) )
, (22)
i=1
Q i (z
where Qi = N (q(x(i) ; φ), diag(v(x(i) ; ψ))2 )

Note that to evaluate Qi (z (i) ) inside the expectation, we should be able to

compute the density of Qi . To estimate the expectation Ez(i) ∼Qi , we
should be able to sample from distribution Qi so that we can build an
empirical estimator with samples. It happens that for Gaussian distribution
Qi = N (q(x(i) ; φ), diag(v(x(i) ; ψ))2 ), we are able to be both efficiently.
Now let’s optimize the ELBO. It turns out that we can run gradient
ascent over φ, ψ, θ instead of alternating maximization. There is no strong
need to compute the maximum over each variable at a much greater cost. (For
Gaussian mixture model in Section 3, computing the maximum is analytically
feasible and relatively cheap, and therefore we did alternating maximization.)
Mathematically, let η be the learning rate, the gradient ascent step is

θ := θ + η∇θ ELBO(φ, ψ, θ)
φ := φ + η∇φ ELBO(φ, ψ, θ)
ψ := ψ + η∇ψ ELBO(φ, ψ, θ)

Computing the gradient over θ is simple because

n
log p(x(i) , z (i) ; θ)
X  
∇θ ELBO(φ, ψ, θ) = ∇θ Ez(i) ∼Qi
i=1
Qi (z (i) )
n
X
Ez(i) ∼Qi log p(x(i) , z (i) ; θ)
 
= ∇θ
i=1
n
X
Ez(i) ∼Qi ∇θ log p(x(i) , z (i) ; θ) ,
 
= (23)
i=1

But computing the gradient over φ and ψ is tricky because the sampling
distribution Qi depends on φ and ψ. (Abstractly speaking, the issue we
face can be simplified as the problem of computing the gradient Ez∼Qφ [f (φ)]
with respect to variable φ. We know that in general, ∇Ez∼Qφ [f (φ)] 6=
Ez∼Qφ [∇f (φ)] because the dependency of Qφ on φ has to be taken into ac-
count as well. )
The idea that comes to rescue is the so-called re-parameterization
trick: we rewrite z (i) ∼ Qi = N (q(x(i) ; φ), diag(v(x(i) ; ψ))2 ) in an equivalent
 14

way:
z (i) = q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) where ξ (i) ∼ N (0, Ik×k ) (24)
Here x ⊙ y denotes the entry-wise product of two vectors of the same
dimension. Here we used the fact that x ∼ N (µ, σ 2 ) is equivalent to that
x = µ+ξσ with ξ ∼ N (0, 1). We mostly just used this fact in every dimension
simultaneously for the random variable z (i) ∼ Qi .
With this re-parameterization, we have that
p(x(i) , z (i) ; θ)
 
Ez(i) ∼Qi log (25)
Qi (z (i) )
p(x(i) , q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) ; θ)
 
= Eξ(i) ∼N (0,1) log
Qi (q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) )
It follows that
p(x(i) , z (i) ; θ)
 
∇φ Ez(i) ∼Qi log
Qi (z (i) )
p(x(i) , q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) ; θ)
 
= ∇φ Eξ(i) ∼N (0,1) log
Qi (q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) )
p(x(i) , q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) ; θ)
 
= Eξ(i) ∼N (0,1) ∇φ log
Qi (q(x(i) ; φ) + v(x(i) ; ψ) ⊙ ξ (i) )
We can now sample multiple copies of ξ (i) ’s to estimate the the expecta-
tion in the RHS of the equation above.7 We can estimate the gradient with
respect to ψ similarly, and with these, we can implement the gradient ascent
algorithm to optimize the ELBO over φ, ψ, θ.
There are not many high-dimensional distributions with analytically com-
putable density function are known to be re-parameterizable. We refer to [2]
for a few other choices that can replace Gaussian distribution.

References
[1] David M Blei, Alp Kucukelbir, and Jon D McAuliffe. Variational in-
ference: A review for statisticians. Journal of the American Statistical
Association, 112(518):859–877, 2017.
[2] Diederik P Kingma and Max Welling. Auto-encoding variational bayes.
arXiv preprint arXiv:1312.6114, 2013.
7
Empirically people sometimes just use one sample to estimate it for maximum com-
putational efficiency.
 CS229 Lecture notes
Andrew Ng

Part X
Factor analysis
When we have data x(i) ∈ Rn that comes from a mixture of several Gaussians,
the EM algorithm can be applied to fit a mixture model. In this setting,
we usually imagine problems were the we have sufficient data to be able
to discern the multiple-Gaussian structure in the data. For instance, this
would be the case if our training set size m was significantly larger than the
dimension n of the data.
Now, consider a setting in which n  m. In such a problem, it might be
difficult to model the data even with a single Gaussian, much less a mixture of
Gaussian. Specifically, since the m data points span only a low-dimensional
subspace of Rn , if we model the data as Gaussian, and estimate the mean
and covariance using the usual maximum likelihood estimators,
m
1 X (i)
µ = x
m i=1
m
1 X (i)
Σ = (x − µ)(x(i) − µ)T ,
m i=1

we would find that the matrix Σ is singular. This means that Σ−1 does not
exist, and 1/|Σ|1/2 = 1/0. But both of these terms are needed in computing
the usual density of a multivariate Gaussian distribution. Another way of
stating this difficulty is that maximum likelihood estimates of the parameters
result in a Gaussian that places all of its probability in the affine space
spanned by the data,1 and this corresponds to a singular covariance matrix.
1
Pm Pm
This is the set of points x satisfying x = i=1 αi x(i) , for some αi ’s so that i=1 α1 =
1.

1
 2

More generally, unless m exceeds n by some reasonable amount, the max-

imum likelihood estimates of the mean and covariance may be quite poor.
Nonetheless, we would still like to be able to fit a reasonable Gaussian model
to the data, and perhaps capture some interesting covariance structure in
the data. How can we do this?
In the next section, we begin by reviewing two possible restrictions on
Σ, ones that allow us to fit Σ with small amounts of data but neither of
which will give a satisfactory solution to our problem. We next discuss some
properties of Gaussians that will be needed later; specifically, how to find
marginal and conditonal distributions of Gaussians. Finally, we present the
factor analysis model, and EM for it.

1 Restrictions of Σ
If we do not have sufficient data to fit a full covariance matrix, we may
place some restrictions on the space of matrices Σ that we will consider. For
instance, we may choose to fit a covariance matrix Σ that is diagonal. In this
setting, the reader may easily verify that the maximum likelihood estimate
of the covariance matrix is given by the diagonal matrix Σ satisfying
m
1 X (i)
Σjj = (x − µj )2 .
m i=1 j

Thus, Σjj is just the empirical estimate of the variance of the j-th coordinate
of the data.
Recall that the contours of a Gaussian density are ellipses. A diagonal
Σ corresponds to a Gaussian where the major axes of these ellipses are axis-
aligned.
Sometimes, we may place a further restriction on the covariance matrix
that not only must it be diagonal, but its diagonal entries must all be equal.
In this setting, we have Σ = σ 2 I, where σ 2 is the parameter under our control.
The maximum likelihood estimate of σ 2 can be found to be:
n m
2 1 X X (i)
σ = (x − µj )2 .
mn j=1 i=1 j

This model corresponds to using Gaussians whose densities have contours

that are circles (in 2 dimesions; or spheres/hyperspheres in higher dimen-
sions).
 3

If we were fitting a full, unconstrained, covariance matrix Σ to data, it

was necessary that m ≥ n + 1 in order for the maximum likelihood estimate
of Σ not to be singular. Under either of the two restrictions above, we may
obtain non-singular Σ when m ≥ 2.
However, restricting Σ to be diagonal also means modeling the different
coordinates xi , xj of the data as being uncorrelated and independent. Often,
it would be nice to be able to capture some interesting correlation structure
in the data. If we were to use either of the restrictions on Σ described above,
we would therefore fail to do so. In this set of notes, we will describe the
factor analysis model, which uses more parameters than the diagonal Σ and
captures some correlations in the data, but also without having to fit a full
covariance matrix.

2 Marginals and conditionals of Gaussians

Before describing factor analysis, we digress to talk about how to find condi-
tional and marginal distributions of random variables with a joint multivari-
ate Gaussian distribution.
Suppose we have a vector-valued random variable
 
x1
x= ,
x2

where x1 ∈ Rr , x2 ∈ Rs , and x ∈ Rr+s . Suppose x ∼ N (µ, Σ), where

   
µ1 Σ11 Σ12
µ= , Σ= .
µ2 Σ21 Σ22

Here, µ1 ∈ Rr , µ2 ∈ Rs , Σ11 ∈ Rr×r , Σ12 ∈ Rr×s , and so on. Note that since
covariance matrices are symmetric, Σ12 = ΣT21 .
Under our assumptions, x1 and x2 are jointly multivariate Gaussian.
What is the marginal distribution of x1 ? It is not hard to see that E[x1 ] = µ1 ,
and that Cov(x1 ) = E[(x1 − µ1 )(x1 − µ1 )] = Σ11 . To see that the latter is
true, note that by definition of the joint covariance of x1 and x2 , we have
 4

that

Cov(x) = Σ
 
Σ11 Σ12
=
Σ21 Σ22
= E[(x − µ)(x − µ)T ]
"  T #
x1 − µ 1 x1 − µ 1
= E
x2 − µ 2 x2 − µ 2
(x1 − µ1 )(x1 − µ1 )T (x1 − µ1 )(x2 − µ2 )T
 
= E .
(x2 − µ2 )(x1 − µ1 )T (x2 − µ2 )(x2 − µ2 )T

Matching the upper-left subblocks in the matrices in the second and the last
lines above gives the result.
Since marginal distributions of Gaussians are themselves Gaussian, we
therefore have that the marginal distribution of x1 is given by x1 ∼ N (µ1 , Σ11 ).
Also, we can ask, what is the conditional distribution of x1 given x2 ? By
referring to the definition of the multivariate Gaussian distribution, it can
be shown that x1 |x2 ∼ N (µ1|2 , Σ1|2 ), where

µ1|2 = µ1 + Σ12 Σ−1

22 (x2 − µ2 ), (1)
Σ1|2 = Σ11 − Σ12 Σ−1
22 Σ21 . (2)

When working with the factor analysis model in the next section, these
formulas for finding conditional and marginal distributions of Gaussians will
be very useful.

3 The Factor analysis model

In the factor analysis model, we posit a joint distribution on (x, z) as follows,
where z ∈ Rk is a latent random variable:

z ∼ N (0, I)
x|z ∼ N (µ + Λz, Ψ).

Here, the parameters of our model are the vector µ ∈ Rn , the matrix
Λ ∈ Rn×k , and the diagonal matrix Ψ ∈ Rn×n . The value of k is usually
chosen to be smaller than n.
 5

Thus, we imagine that each datapoint x(i) is generated by sampling a k

dimension multivariate Gaussian z (i) . Then, it is mapped to a k-dimensional
affine space of Rn by computing µ + Λz (i) . Lastly, x(i) is generated by adding
covariance Ψ noise to µ + Λz (i) .
Equivalently (convince yourself that this is the case), we can therefore
also define the factor analysis model according to

z ∼ N (0, I)
 ∼ N (0, Ψ)
x = µ + Λz + .

where  and z are independent.

Lets work out exactly what distribution our model defines. Our random
variables z and x have a joint Gaussian distribution
 
z
∼ N (µzx , Σ).
x
We will now find µzx and Σ.
We know that E[z] = 0, from the fact that z ∼ N (0, I). Also, we have
that

E[x] = E[µ + Λz + ]
= µ + ΛE[z] + E[]
= µ.

Putting these together, we obtain

~0
 
µzx =
µ

Next, to find, Σ, we need to calculate Σzz = E[(z − E[z])(z − E[z])T ] (the

upper-left block of Σ), Σzx = E[(z − E[z])(x − E[x])T ] (upper-right block),
and Σxx = E[(x − E[x])(x − E[x])T ] (lower-right block).
Now, since z ∼ N (0, I), we easily find that Σzz = Cov(z) = I. Also,

E[(z − E[z])(x − E[x])T ] = E[z(µ + Λz +  − µ)T ]

= E[zz T ]ΛT + E[zT ]
= ΛT .

In the last step, we used the fact that E[zz T ] = Cov(z) (since z has zero
mean), and E[zT ] = E[z]E[T ] = 0 (since z and  are independent, and
 6

hence the expectation of their product is the product of their expectations).

Similarly, we can find Σxx as follows:

E[(x − E[x])(x − E[x])T ] = E[(µ + Λz +  − µ)(µ + Λz +  − µ)T ]

= E[Λzz T ΛT + z T ΛT + ΛzT + T ]
= ΛE[zz T ]ΛT + E[T ]
= ΛΛT + Ψ.

Putting everything together, we therefore have that

~0 ΛT
     
z I
∼N , . (3)
x µ Λ ΛΛT + Ψ

Hence, we also see that the marginal distribution of x is given by x ∼

N (µ, ΛΛT + Ψ). Thus, given a training set {x(i) ; i = 1, . . . , m}, we can write
down the log likelihood of the parameters:
m  
Y 1 1 (i) T T −1 (i)
`(µ, Λ, Ψ) = log exp − (x − µ) (ΛΛ + Ψ) (x − µ) .
i=1
(2π)n/2 |ΛΛT + Ψ| 2

To perform maximum likelihood estimation, we would like to maximize this

quantity with respect to the parameters. But maximizing this formula ex-
plicitly is hard (try it yourself), and we are aware of no algorithm that does
so in closed-form. So, we will instead use to the EM algorithm. In the next
section, we derive EM for factor analysis.

4 EM for factor analysis

The derivation for the E-step is easy. We need to compute Qi (z (i) ) =
p(z (i) |x(i) ; µ, Λ, Ψ). By substituting the distribution given in Equation (3)
into the formulas (1-2) used for finding the conditional distribution of a
Gaussian, we find that z (i) |x(i) ; µ, Λ, Ψ ∼ N (µz(i) |x(i) , Σz(i) |x(i) ), where

µz(i) |x(i) = ΛT (ΛΛT + Ψ)−1 (x(i) − µ),

Σz(i) |x(i) = I − ΛT (ΛΛT + Ψ)−1 Λ.

So, using these definitions for µz(i) |x(i) and Σz(i) |x(i) , we have
 
(i) 1 1 (i) T −1 (i)
Qi (z ) = exp − (z − µz(i) |x(i) ) Σz(i) |x(i) (z − µz(i) |x(i) ) .
(2π)k/2 |Σz(i) |x(i) |1/2 2
 7

Lets now work out the M-step. Here, we need to maximize

m Z
X p(x(i) , z (i) ; µ, Λ, Ψ) (i)
Qi (z (i) ) log (i) )
dz (4)
i=1 z (i) Q i (z
with respect to the parameters µ, Λ, Ψ. We will work out only the optimiza-
tion with respect to Λ, and leave the derivations of the updates for µ and Ψ
as an exercise to the reader.
We can simplify Equation (4) as follows:
Xm Z
Qi (z (i) ) log p(x(i) |z (i) ; µ, Λ, Ψ) + log p(z (i) ) − log Qi (z (i) ) dz (i) (5)
 
i=1 z (i)
m
X
Ez(i) ∼Qi log p(x(i) |z (i) ; µ, Λ, Ψ) + log p(z (i) ) − log Qi (z (i) )
 
= (6)
i=1

Here, the “z (i) ∼ Qi ” subscript indicates that the expectation is with respect
to z (i) drawn from Qi . In the subsequent development, we will omit this
subscript when there is no risk of ambiguity. Dropping terms that do not
depend on the parameters, we find that we need to maximize:
Xm
E log p(x(i) |z (i) ; µ, Λ, Ψ)
 
i=1
m   
X 1 1
= E log exp − (x(i) − µ − Λz (i) )T Ψ−1 (x(i) − µ − Λz (i) )
i=1
(2π)n/2 |Ψ|1/2 2
m  
X 1 n 1
= E − log |Ψ| − log(2π) − (x(i) − µ − Λz (i) )T Ψ−1 (x(i) − µ − Λz (i) )
i=1
2 2 2
Lets maximize this with respect to Λ. Only the last term above depends
on Λ. Taking derivatives, and using the facts that tr a = a (for a ∈ R),
trAB = trBA, and ∇A trABAT C = CAB + C T AB, we get:
m  
X 1 (i) (i) T −1 (i) (i)
∇Λ −E (x − µ − Λz ) Ψ (x − µ − Λz )
i=1
2
m  
X 1 (i) T T −1 (i) (i) T T −1 (i)
= ∇Λ E −tr z Λ Ψ Λz + trz Λ Ψ (x − µ)
i=1
2
m  
X 1 T −1 (i) (i) T T −1 (i) (i) T
= ∇Λ E −tr Λ Ψ Λz z + trΛ Ψ (x − µ)z
i=1
2
m h
(i) (i) T (i) T
X i
−1 −1 (i)
= E −Ψ Λz z + Ψ (x − µ)z
i=1
 8

Setting this to zero and simplifying, we get:

m h i m h Ti
(i) (i) T
X X
ΛEz(i) ∼Qi z z = (i)
(x − µ)Ez(i) ∼Qi z (i) .
i=1 i=1

Hence, solving for Λ, we obtain

m
! m
!−1
h i h i
(i) T (i) (i) T
X X
Λ= (x(i) − µ)Ez(i) ∼Qi z Ez(i) ∼Qi z z . (7)
i=1 i=1

It is interesting to note the close relationship between this equation and the
normal equation that we’d derived for least squares regression,

“θT = (y T X)(X T X)−1 .”

The analogy is that here, the x’s are a linear function of the z’s (plus noise).
Given the “guesses” for z that the E-step has found, we will now try to
estimate the unknown linearity Λ relating the x’s and z’s. It is therefore
no surprise that we obtain something similar to the normal equation. There
is, however, one important difference between this and an algorithm that
performs least squares using just the “best guesses” of the z’s; we will see
this difference shortly.
To complete our M-step update, lets work out the values of the expecta-
tions in Equation (7). From our definition of Qi being Gaussian with mean
µz(i) |x(i) and covariance Σz(i) |x(i) , we easily find
h
(i) T
i
Ez(i) ∼Qi z = µTz(i) |x(i)
h T
i
Ez(i) ∼Qi z (i) z (i) = µz(i) |x(i) µTz(i) |x(i) + Σz(i) |x(i) .

The latter comes from the fact that, for a random variable Y , Cov(Y ) =
E[Y Y T ] − E[Y ]E[Y ]T , and hence E[Y Y T ] = E[Y ]E[Y ]T + Cov(Y ). Substitut-
ing this back into Equation (7), we get the M-step update for Λ:
m
! m
!−1
X X
Λ= (x(i) − µ)µTz(i) |x(i) µz(i) |x(i) µTz(i) |x(i) + Σz(i) |x(i) . (8)
i=1 i=1

It is important to note the presence of the Σz(i) |x(i) on the right hand side of
this equation. This is the covariance in the posterior distribution p(z (i) |x(i) )
of z (i) give x(i) , and the M-step must take into account this uncertainty
 9

about z (i) in the posterior. A common mistake in deriving EM is to assume

that in the E-step, we need to calculate only expectation E[z] of the latent
random variable z, and then plug that into the optimization in the M-step
everywhere z occurs. While this worked for simple problems such as the
mixture of Gaussians, in our derivation for factor analysis, we needed E[zz T ]
as well E[z]; and as we saw, E[zz T ] and E[z]E[z]T differ by the quantity Σz|x .
Thus, the M-step update must take into account the covariance of z in the
posterior distribution p(z (i) |x(i) ).
Lastly, we can also find the M-step optimizations for the parameters µ
and Ψ. It is not hard to show that the first is given by
m
1 X (i)
µ= x .
m i=1

Since this doesn’t change as the parameters are varied (i.e., unlike the update
for Λ, the right hand side does not depend on Qi (z (i) ) = p(z (i) |x(i) ; µ, Λ, Ψ),
which in turn depends on the parameters), this can be calculated just once
and needs not be further updated as the algorithm is run. Similarly, the
diagonal Ψ can be found by calculating
m
1 X (i) (i) T (i) T T
Φ= x x −x µz(i) |x(i) ΛT −Λµz(i) |x(i) x(i) +Λ(µz(i) |x(i) µTz(i) |x(i) +Σz(i) |x(i) )ΛT ,
m i=1

and setting Ψii = Φii (i.e., letting Ψ be the diagonal matrix containing only
the diagonal entries of Φ).
 CS229 Lecture notes
Andrew Ng

Part XI
Principal components analysis
In our discussion of factor analysis, we gave a way to model data x ∈ Rn as
“approximately” lying in some k-dimension subspace, where k  n. Specif-
ically, we imagined that each point x(i) was created by first generating some
z (i) lying in the k-dimension affine space {Λz + µ; z ∈ Rk }, and then adding
Ψ-covariance noise. Factor analysis is based on a probabilistic model, and
parameter estimation used the iterative EM algorithm.
In this set of notes, we will develop a method, Principal Components
Analysis (PCA), that also tries to identify the subspace in which the data
approximately lies. However, PCA will do so more directly, and will require
only an eigenvector calculation (easily done with the eig function in Matlab),
and does not need to resort to EM.
Suppose we are given dataset {x(i) ; i = 1, . . . , m} of attributes of m dif-
ferent types of automobiles, such as their maximum speed, turn radius, and
so on. Lets x(i) ∈ Rn for each i (n  m). But unknown to us, two different
attributes—some xi and xj —respectively give a car’s maximum speed mea-
sured in miles per hour, and the maximum speed measured in kilometers per
hour. These two attributes are therefore almost linearly dependent, up to
only small differences introduced by rounding off to the nearest mph or kph.
Thus, the data really lies approximately on an n − 1 dimensional subspace.
How can we automatically detect, and perhaps remove, this redundancy?
For a less contrived example, consider a dataset resulting from a survey of
(i)
pilots for radio-controlled helicopters, where x1 is a measure of the piloting
(i)
skill of pilot i, and x2 captures how much he/she enjoys flying. Because
RC helicopters are very difficult to fly, only the most committed students,
ones that truly enjoy flying, become good pilots. So, the two attributes
x1 and x2 are strongly correlated. Indeed, we might posit that that the

1
 2

data actually likes along some diagonal axis (the u1 direction) capturing the
intrinsic piloting “karma” of a person, with only a small amount of noise
lying off this axis. (See figure.) How can we automatically compute this u1
direction?

u1
x2 (enjoyment)

u2

x1 (skill)

We will shortly develop the PCA algorithm. But prior to running PCA
per se, typically we first pre-process the data to normalize its mean and
variance, as follows:

1. Let µ = m1 m (i)
P
i=1 x .

2. Replace each x(i) with x(i) − µ.

P (i)
3. Let σj2 = m1 i (xj )2
(i) (i)
4. Replace each xj with xj /σj .

Steps (1-2) zero out the mean of the data, and may be omitted for data
known to have zero mean (for instance, time series corresponding to speech
or other acoustic signals). Steps (3-4) rescale each coordinate to have unit
variance, which ensures that different attributes are all treated on the same
“scale.” For instance, if x1 was cars’ maximum speed in mph (taking values
in the high tens or low hundreds) and x2 were the number of seats (taking
values around 2-4), then this renormalization rescales the different attributes
to make them more comparable. Steps (3-4) may be omitted if we had
apriori knowledge that the different attributes are all on the same scale. One
 3

example of this is if each data point represented a grayscale image, and each
(i)
xj took a value in {0, 1, . . . , 255} corresponding to the intensity value of
pixel j in image i.
Now, having carried out the normalization, how do we compute the “ma-
jor axis of variation” u—that is, the direction on which the data approxi-
mately lies? One way to pose this problem is as finding the unit vector u so
that when the data is projected onto the direction corresponding to u, the
variance of the projected data is maximized. Intuitively, the data starts off
with some amount of variance/information in it. We would like to choose a
direction u so that if we were to approximate the data as lying in the direc-
tion/subspace corresponding to u, as much as possible of this variance is still
retained.
Consider the following dataset, on which we have already carried out the
normalization steps:

Now, suppose we pick u to correspond the the direction shown in the

figure below. The circles denote the projections of the original data onto this
line.
 4



   
   
    

 


 


  

  


 



  
       




 




We see that the projected data still has a fairly large variance, and the
points tend to be far from zero. In contrast, suppose had instead picked the
following direction:

#$
$
#$ #$
$
#$ #$
$
#$ #$
$
#$ #$
$
#$ #$
$
#$ #$
$
#$ $#$#$#
#$# # # 
$ # # 
$ # # 
$ # # 
$ # # 
$ # # 
$ $#$#$#
$#
# 
$ # #$
$
# # #$
$
# # #$
$
# # #$
$
# # #$
$
# # #$
$
# #
'(
(
'(' '(
(
' 
$ $#$#$# ( '(
' 
$ $#$#$# ( '(
' 
$ $#$#$# ('('(' $ 
$ #$
# 
$ #$
$
# 
$ #$
$
# 
$ #$
$
# $#$#$#
(' ' 
( ' 
( # $ ' 
( ' 
( # $ ' 
( ' 
( # $ ' ( # 
$ # 
$ % 
& % 
& # 
$ # 
$ % 
& % 
& # 
$ # 
$ % 
& % 
& # 
$ # 
$ % 
& % 
& $#$#$# & %&
% %&
&
% %&
&
% &%&%%&
( ' ' 
( # $ ' 
( # $ ' 
( # $ ' ( # 
$ 
% & %& # 
$ 
% & %& # 
$ 
% & %& # 
$ 
% & %& 
% & %& 
% & %& 
% & %& &%&%&%
       
  
   
   
 
('
'   (
( '(
' $#$#  ( '(
' $#$# ( '(
' $#$# ('(' $ # &
#$ %&
% % 
&
#$# &
$ %&
% % 
&
#$# &
$ %&
% % 
&
#$# &
$ %&
% % 
&
$#$# & %&
% % 
&
&
%&
% % 
&
&
%&
% % 
& &%&%&%
      
 

 
  

  %&
&
% %&
&
% %&
&
% %&
&
% %&
&
% %&
&
% %&
&
%
   
  
 "  
 "  " % 
& %& % 
& %& % 
& %& % 
& %& % 
& %& % 
& %& % 
& %& &%&%&%
      
 

  
 ! 
" ! 
"
 ! ! 
"
 ! ! 
" ! ! 
" ! 
" ! 
"  ! " ! "     &
&
% % 
&
&
% % 
&
&
% % 
&
&
% % 
&
&
% % 
&
&
% % 
&
&
% % 
& % &%&%
        
 
 !"!! 
"  
 !"!! 

"  
 !"

" ! !"
 " ! ! 
" !" ! 
" ! 
" !" 
 ! " ! " % & % & % & % & % & % &
       
 

 ""
!"
!  "" !"
!  " !"
!  " !"
! "
!"
! "!"!"!
          
 ! 
"  !" ! 
"  !" ! 
"  !" ! 
" ! 
" "!
 
  "
 !"
!  "
!" !  "
!" !  "
!" ! "
!" !"
! "!"!"!
!"

Here, the projections have a significantly smaller variance, and are much
closer to the origin.
We would like to automatically select the direction u corresponding to
the first of the two figures shown above. To formalize this, note that given a
 5

unit vector u and a point x, the length of the projection of x onto u is given
by xT u. I.e., if x(i) is a point in our dataset (one of the crosses in the plot),
then its projection onto u (the corresponding circle in the figure) is distance
xT u from the origin. Hence, to maximize the variance of the projections, we
would like to choose a unit-length u so as to maximize:
m m
1 X (i) T 2 1 X T (i) (i) T
(x u) = u x x u
m i=1 m i=1
m
!
1 X (i) (i) T
= uT x x u.
m i=1

We easily recognize that the maximizing this subject to ||u||2 = 1 gives the
(i) (i) T
principal eigenvector of Σ = m1 m
P
i=1 x x , which is just the empirical
covariance matrix of the data (assuming it has zero mean).1
To summarize, we have found that if we wish to find a 1-dimensional
subspace with with to approximate the data, we should choose u to be the
principal eigenvector of Σ. More generally, if we wish to project our data
into a k-dimensional subspace (k < n), we should choose u1 , . . . , uk to be the
top k eigenvectors of Σ. The ui ’s now form a new, orthogonal basis for the
data.2
Then, to represent x(i) in this basis, we need only compute the corre-
sponding vector  
uT1 x(i)
 uT x(i) 
(i)  2
y = ..  ∈ Rk .

 . 
T (i)
uk x
Thus, whereas x(i) ∈ Rn , the vector y (i) now gives a lower, k-dimensional,
approximation/representation for x(i) . PCA is therefore also referred to as
a dimensionality reduction algorithm. The vectors u1 , . . . , uk are called
the first k principal components of the data.
Remark. Although we have shown it formally only for the case of k = 1,
using well-known properties of eigenvectors it is straightforward to show that
1
If you haven’t seen this before, try using the method of Lagrange multipliers to max-
imize uT Σu subject to that uT u = 1. You should be able to show that Σu = λu, for some
λ, which implies u is an eigenvector of Σ, with eigenvalue λ.
2
Because Σ is symmetric, the ui ’s will (or always can be chosen to be) orthogonal to
each other.
 6

of all possible orthogonal bases u1 , . . . , uk , the one that we have chosen max-
(i) 2
P
imizes i ||y ||2 . Thus, our choice of a basis preserves as much variability
as possible in the original data.
In problem set 4, you will see that PCA can also be derived by picking
the basis that minimizes the approximation error arising from projecting the
data onto the k-dimensional subspace spanned by them.
PCA has many applications, our discussion with a small number of exam-
ples. First, compression—representing x(i) ’s with lower dimension y (i) ’s—is
an obvious application. If we reduce high dimensional data to k = 2 or 3 di-
mensions, then we can also plot the y (i) ’s to visualize the data. For instance,
if we were to reduce our automobiles data to 2 dimensions, then we can plot
it (one point in our plot would correspond to one car type, say) to see what
cars are similar to each other and what groups of cars may cluster together.
Another standard application is to preprocess a dataset to reduce its
dimension before running a supervised learning learning algorithm with the
x(i) ’s as inputs. Apart from computational benefits, reducing the data’s
dimension can also reduce the complexity of the hypothesis class considered
and help avoid overfitting (e.g., linear classifiers over lower dimensional input
spaces will have smaller VC dimension).
Lastly, as in our RC pilot example, we can also view PCA as a noise
reduction algorithm. In our example it, estimates the intrinsic “piloting
karma” from the noisy measures of piloting skill and enjoyment. In class, we
also saw the application of this idea to face images, resulting in eigenfaces
method. Here, each point x(i) ∈ R100×100 was a 10000 dimensional vector,
with each coordinate corresponding to a pixel intensity value in a 100x100
image of a face. Using PCA, we represent each image x(i) with a much lower-
dimensional y (i) . In doing so, we hope that the principal components we
found retain the interesting, systematic variations between faces that capture
what a person really looks like, but not the “noise” in the images introduced
by minor lighting variations, slightly different imaging conditions, and so on.
We then measure distances between faces i and j by working in the reduced
dimension, and computing ||y (i) − y (j) ||2 . This resulted in a surprisingly good
face-matching and retrieval algorithm.
 CS229 Lecture notes
Andrew Ng

Part XII
Independent Components
Analysis
Our next topic is Independent Components Analysis (ICA). Similar to PCA,
this will find a new basis in which to represent our data. However, the goal
is very different.
As a motivating example, consider the “cocktail party problem.” Here, n
speakers are speaking simultaneously at a party, and any microphone placed
in the room records only an overlapping combination of the n speakers’ voices.
But lets say we have n different microphones placed in the room, and because
each microphone is a different distance from each of the speakers, it records a
different combination of the speakers’ voices. Using these microphone record-
ings, can we separate out the original n speakers’ speech signals?
To formalize this problem, we imagine that there is some data s ∈ Rn
that is generated via n independent sources. What we observe is

x = As,

where A is an unknown square matrix called the mixing matrix. Repeated

observations gives us a dataset {x(i) ; i = 1, . . . , m}, and our goal is to recover
the sources s(i) that had generated our data (x(i) = As(i) ).
(i)
In our cocktail party problem, s(i) is an n-dimensional vector, and sj is
the sound that speaker j was uttering at time i. Also, x(i) in an n-dimensional
(i)
vector, and xj is the acoustic reading recorded by microphone j at time i.
Let W = A−1 be the unmixing matrix. Our goal is to find W , so
that given our microphone recordings x(i) , we can recover the sources by
computing s(i) = W x(i) . For notational convenience, we also let wiT denote

1
 2

the i-th row of W , so that

 
— w1T —
W = ..
.
 
.
— wnT —
(i)
Thus, wi ∈ Rn , and the j-th source can be recovered by computing sj =
wjT x(i) .

1 ICA ambiguities
To what degree can W = A−1 be recovered? If we have no prior knowledge
about the sources and the mixing matrix, it is not hard to see that there are
some inherent ambiguities in A that are impossible to recover, given only the
x(i) ’s.
Specifically, let P be any n-by-n permutation matrix. This means that
each row and each column of P has exactly one “1.” Here’re some examples
of permutation matrices:
 
0 1 0    
0 1 1 0
P =  1 0 0 ; P = ; P = .
1 0 0 1
0 0 1

If z is a vector, then P z is another vector that’s contains a permuted version

of z’s coordinates. Given only the x(i) ’s, there will be no way to distinguish
between W and P W . Specifically, the permutation of the original sources is
ambiguous, which should be no surprise. Fortunately, this does not matter
for most applications.
Further, there is no way to recover the correct scaling of the wi ’s. For in-
stance, if A were replaced with 2A, and every s(i) were replaced with (0.5)s(i) ,
then our observed x(i) = 2A · (0.5)s(i) would still be the same. More broadly,
if a single column of A were scaled by a factor of α, and the corresponding
source were scaled by a factor of 1/α, then there is again no way, given only
the x(i) ’s to determine that this had happened. Thus, we cannot recover the
“correct” scaling of the sources. However, for the applications that we are
concerned with—including the cocktail party problem—this ambiguity also
(i)
does not matter. Specifically, scaling a speaker’s speech signal sj by some
positive factor α affects only the volume of that speaker’s speech. Also, sign
(i) (i)
changes do not matter, and sj and −sj sound identical when played on a
speaker. Thus, if the wi found by an algorithm is scaled by any non-zero real
 3

number, the corresponding recovered source si = wiT x will be scaled by the

same factor; but this usually does not matter. (These comments also apply
to ICA for the brain/MEG data that we talked about in class.)
Are these the only sources of ambiguity in ICA? It turns out that they
are, so long as the sources si are non-Gaussian. To see what the difficulty is
with Gaussian data, consider an example in which n = 2, and s ∼ N (0, I).
Here, I is the 2x2 identity matrix. Note that the contours of the density of
the standard normal distribution N (0, I) are circles centered on the origin,
and the density is rotationally symmetric.
Now, suppose we observe some x = As, where A is our mixing matrix.
The distribution of x will also be Gaussian, with zero mean and covariance
E[xxT ] = E[AssT AT ] = AAT . Now, let R be an arbitrary orthogonal (less
formally, a rotation/reflection) matrix, so that RR T = RT R = I, and let
A0 = AR. Then if the data had been mixed according to A0 instead of
A, we would have instead observed x0 = A0 s. The distribution of x0 is
also Gaussian, with zero mean and covariance E[x0 (x0 )T ] = E[A0 ssT (A0 )T ] =
E[ARssT (AR)T ] = ARRT AT = AAT . Hence, whether the mixing matrix
is A or A0 , we would observe data from a N (0, AAT ) distribution. Thus,
there is no way to tell if the sources were mixed using A and A0 . So, there
is an arbitrary rotational component in the mixing matrix that cannot be
determined from the data, and we cannot recover the original sources.
Our argument above was based on the fact that the multivariate standard
normal distribution is rotationally symmetric. Despite the bleak picture that
this paints for ICA on Gaussian data, it turns out that, so long as the data is
not Gaussian, it is possible, given enough data, to recover the n independent
sources.

2 Densities and linear transformations

Before moving on to derive the ICA algorithm proper, we first digress briefly
to talk about the effect of linear transformations on densities.
Suppose we have a random variable s drawn according to some density
ps (s). For simplicity, let us say for now that s ∈ R is a real number. Now, let
the random variable x be defined according to x = As (here, x ∈ R, A ∈ R).
Let px be the density of x. What is px ?
Let W = A−1 . To calculate the “probability” of a particular value of x,
it is tempting to compute s = W x, then then evaluate ps at that point, and
conclude that “px (x) = ps (W x).” However, this is incorrect. For example,
let s ∼ Uniform[0, 1], so that s’s density is ps (s) = 1{0 ≤ s ≤ 1}. Now, let
 4

A = 2, so that x = 2s. Clearly, x is distributed uniformly in the interval

[0, 2]. Thus, its density is given by px (x) = (0.5)1{0 ≤ x ≤ 2}. This does
not equal ps (W x), where W = 0.5 = A−1 . Instead, the correct formula is
px (x) = ps (W x)|W |.
More generally, if s is a vector-valued distribution with density ps , and
x = As for a square, invertible matrix A, then the density of x is given by
px (x) = ps (W x) · |W |,
where W = A−1 .
Remark. If you’re seen the result that A maps [0, 1]n to a set of volume |A|,
then here’s another way to remember the formula for px given above, that also
generalizes our previous 1-dimensional example. Specifically, let A ∈ Rn×n be
given, and let W = A−1 as usual. Also let C1 = [0, 1]n be the n-dimensional
hypercube, and define C2 = {As : s ∈ C1 } ⊆ Rn to be the image of C1
under the mapping given by A. Then it is a standard result in linear algebra
(and, indeed, one of the ways of defining determinants) that the volume of
C2 is given by |A|. Now, suppose s is uniformly distributed in [0, 1]n , so its
density is ps (s) = 1{s ∈ C1 }. Then clearly x will be uniformly distributed
in C2 . Its density is therefore found to be px (x) = 1{x ∈ C2 }/vol(C2 ) (since
it must integrate over C2 to 1). But using the fact that the determinant
of the inverse of a matrix is just the inverse of the determinant, we have
1/vol(C2 ) = 1/|A| = |A−1 | = |W |. Thus, px (x) = 1{x ∈ C2 }|W | = 1{W x ∈
C1 }|W | = ps (W x)|W |.

3 ICA algorithm
We are now ready to derive an ICA algorithm. The algorithm we describe
is due to Bell and Sejnowski, and the interpretation we give will be of their
algorithm as a method for maximum likelihood estimation. (This is different
from their original interpretation, which involved a complicated idea called
the infomax principal, that is no longer necessary in the derivation given the
modern understanding of ICA.)
We suppose that the distribution of each source si is given by a density
ps , and that the joint distribution of the sources s is given by
n
Y
p(s) = ps (si ).
i=1

Note that by modeling the joint distribution as a product of the marginal,

we capture the assumption that the sources are independent. Using our
 5

formulas from the previous section, this implies the following density on
x = As = W −1 s: n
Y
p(x) = ps (wiT x) · |W |.
i=1

All that remains is to specify a density for the individual sources ps .

Recall that, given a real-valued random variable z, its cumulative R z0 distri-
bution function (cdf) F is defined by F (z0 ) = P (z ≤ z0 ) = −∞ pz (z)dz.
Also, the density of z can be found from the cdf by taking its derivative:
pz (z) = F 0 (z).
Thus, to specify a density for the si ’s, all we need to do is to specify some
cdf for it. A cdf has to be a monotonic function that increases from zero
to one. Following our previous discussion, we cannot choose the cdf to be
the cdf of the Gaussian, as ICA doesn’t work on Gaussian data. What we’ll
choose instead for the cdf, as a reasonable “default” function that slowly
increases from 0 to 1, is the sigmoid function g(s) = 1/(1 + e−s ). Hence,
ps (s) = g 0 (s).1
The square matrix W is the parameter in our model. Given a training
set {x(i) ; i = 1, . . . , m}, the log likelihood is given by
m n
!
X X
`(W ) = log g 0 (wjT x(i) ) + log |W | .
i=1 j=1

We would like to maximize this in terms W . By taking derivatives and using

the fact (from the first set of notes) that ∇W |W | = |W |(W −1 )T , we easily
derive a stochastic gradient ascent learning rule. For a training example x(i) ,
the update rule is:
  
1 − 2g(w1T x(i) )
 1 − 2g(wT x(i) )  
2  (i) T
W := W + α  .. x + (W T )−1  ,
 
 .  
T (i)
1 − 2g(wn x )
1
If you have prior knowledge that the sources’ densities take a certain form, then it
is a good idea to substitute that in here. But in the absence of such knowledge, the
sigmoid function can be thought of as a reasonable default that seems to work well for
many problems. Also, the presentation here assumes that either the data x (i) has been
preprocessed to have zero mean, or that it can naturally be expected to have zero mean
(such as acoustic signals). This is necessary because our assumption that p s (s) = g 0 (s)
implies E[s] = 0 (the derivative of the logistic function is a symmetric function, and
hence gives a density corresponding to a random variable with zero mean), which implies
E[x] = E[As] = 0.
 6

where α is the learning rate.

After the algorithm converges, we then compute s(i) = W x(i) to recover
the original sources.
Remark. When writing down the likelihood of the data, we implicity as-
sumed that the x(i) ’s were independent of each other (for different values
of i; note this issue is different from whether the different coordinates of
x(i) are independent), so that the likelihood of the training set was given by
(i)
Q
i p(x ; W ). This assumption is clearly incorrect for speech data and other
time series where the x(i) ’s are dependent, but it can be shown that having
correlated training examples will not hurt the performance of the algorithm
if we have sufficient data. But, for problems where successive training ex-
amples are correlated, when implementing stochastic gradient ascent, it also
sometimes helps accelerate convergence if we visit training examples in a ran-
domly permuted order. (I.e., run stochastic gradient ascent on a randomly
shuffled copy of the training set.)
 CS229 Lecture notes
Andrew Ng

Part XIII
Reinforcement Learning and
Control
We now begin our study of reinforcement learning and adaptive control.
In supervised learning, we saw algorithms that tried to make their outputs
mimic the labels y given in the training set. In that setting, the labels gave
an unambiguous “right answer” for each of the inputs x. In contrast, for
many sequential decision making and control problems, it is very difficult to
provide this type of explicit supervision to a learning algorithm. For example,
if we have just built a four-legged robot and are trying to program it to walk,
then initially we have no idea what the “correct” actions to take are to make
it walk, and so do not know how to provide explicit supervision for a learning
algorithm to try to mimic.
In the reinforcement learning framework, we will instead provide our al-
gorithms only a reward function, which indicates to the learning agent when
it is doing well, and when it is doing poorly. In the four-legged walking ex-
ample, the reward function might give the robot positive rewards for moving
forwards, and negative rewards for either moving backwards or falling over.
It will then be the learning algorithm’s job to figure out how to choose actions
over time so as to obtain large rewards.
Reinforcement learning has been successful in applications as diverse as
autonomous helicopter flight, robot legged locomotion, cell-phone network
routing, marketing strategy selection, factory control, and efficient web-page
indexing. Our study of reinforcement learning will begin with a definition of
the Markov decision processes (MDP), which provides the formalism in
which RL problems are usually posed.

1
 2

1 Markov decision processes

A Markov decision process is a tuple (S, A, {Psa }, γ, R), where:

• S is a set of states. (For example, in autonomous helicopter flight, S

might be the set of all possible positions and orientations of the heli-
copter.)

• A is a set of actions. (For example, the set of all possible directions in

which you can push the helicopter’s control sticks.)

• Psa are the state transition probabilities. For each state s ∈ S and
action a ∈ A, Psa is a distribution over the state space. We’ll say more
about this later, but briefly, Psa gives the distribution over what states
we will transition to if we take action a in state s.

• γ ∈ [0, 1) is called the discount factor.

• R : S × A 7→ R is the reward function. (Rewards are sometimes also

written as a function of a state S only, in which case we would have
R : S 7→ R).

The dynamics of an MDP proceeds as follows: We start in some state s0 ,

and get to choose some action a0 ∈ A to take in the MDP. As a result of our
choice, the state of the MDP randomly transitions to some successor state
s1 , drawn according to s1 ∼ Ps0 a0 . Then, we get to pick another action a1 .
As a result of this action, the state transitions again, now to some s2 ∼ Ps1 a1 .
We then pick a2 , and so on. . . . Pictorially, we can represent this process as
follows:
a0 a1 a2 a3
s0 −→ s1 −→ s2 −→ s3 −→ ...
Upon visiting the sequence of states s0 , s1 , . . . with actions a0 , a1 , . . ., our
total payoff is given by

R(s0 , a0 ) + γR(s1 , a1 ) + γ 2 R(s2 , a2 ) + · · · .

Or, when we are writing rewards as a function of the states only, this becomes

R(s0 ) + γR(s1 ) + γ 2 R(s2 ) + · · · .

For most of our development, we will use the simpler state-rewards R(s),
though the generalization to state-action rewards R(s, a) offers no special
difficulties.
 3

Our goal in reinforcement learning is to choose actions over time so as to

maximize the expected value of the total payoff:
E R(s0 ) + γR(s1 ) + γ 2 R(s2 ) + · · ·
 

Note that the reward at timestep t is discounted by a factor of γ t . Thus, to

make this expectation large, we would like to accrue positive rewards as soon
as possible (and postpone negative rewards as long as possible). In economic
applications where R(·) is the amount of money made, γ also has a natural
interpretation in terms of the interest rate (where a dollar today is worth
more than a dollar tomorrow).
A policy is any function π : S 7→ A mapping from the states to the
actions. We say that we are executing some policy π if, whenever we are
in state s, we take action a = π(s). We also define the value function for
a policy π according to
V π (s) = E R(s0 ) + γR(s1 ) + γ 2 R(s2 ) + · · · s0 = s, π].


V π (s) is simply the expected sum of discounted rewards upon starting in

state s, and taking actions according to π.1
Given a fixed policy π, its value function V π satisfies the Bellman equa-
tions: X
V π (s) = R(s) + γ Psπ(s) (s′ )V π (s′ ).
s′ ∈S

This says that the expected sum of discounted rewards V π (s) for starting
in s consists of two terms: First, the immediate reward R(s) that we get
right away simply for starting in state s, and second, the expected sum of
future discounted rewards. Examining the second term in more detail, we
see that the summation term above can be rewritten Es′ ∼Psπ(s) [V π (s′ )]. This
is the expected sum of discounted rewards for starting in state s′ , where s′
is distributed according Psπ(s) , which is the distribution over where we will
end up after taking the first action π(s) in the MDP from state s. Thus, the
second term above gives the expected sum of discounted rewards obtained
after the first step in the MDP.
Bellman’s equations can be used to efficiently solve for V π . Specifically,
in a finite-state MDP (|S| < ∞), we can write down one such equation for
V π (s) for every state s. This gives us a set of |S| linear equations in |S|
variables (the unknown V π (s)’s, one for each state), which can be efficiently
solved for the V π (s)’s.
1
This notation in which we condition on π isn’t technically correct because π isn’t a
random variable, but this is quite standard in the literature.
 4

We also define the optimal value function according to

V ∗ (s) = max V π (s). (1)

π

In other words, this is the best possible expected sum of discounted rewards
that can be attained using any policy. There is also a version of Bellman’s
equations for the optimal value function:
X
V ∗ (s) = R(s) + max γ Psa (s′ )V ∗ (s′ ). (2)
a∈A
s′ ∈S

The first term above is the immediate reward as before. The second term
is the maximum over all actions a of the expected future sum of discounted
rewards we’ll get upon after action a. You should make sure you understand
this equation and see why it makes sense.
We also define a policy π ∗ : S 7→ A as follows:
X
π ∗ (s) = arg max Psa (s′ )V ∗ (s′ ). (3)
a∈A
s′ ∈S

Note that π ∗ (s) gives the action a that attains the maximum in the “max”
in Equation (2).
It is a fact that for every state s and every policy π, we have
∗
V ∗ (s) = V π (s) ≥ V π (s).
∗
The first equality says that the V π , the value function for π ∗ , is equal to the
optimal value function V ∗ for every state s. Further, the inequality above
says that π ∗ ’s value is at least a large as the value of any other other policy.
In other words, π ∗ as defined in Equation (3) is the optimal policy.
Note that π ∗ has the interesting property that it is the optimal policy for
all states s. Specifically, it is not the case that if we were starting in some
state s then there’d be some optimal policy for that state, and if we were
starting in some other state s′ then there’d be some other policy that’s opti-
mal policy for s′ . The same policy π ∗ attains the maximum in Equation (1)
for all states s. This means that we can use the same policy π ∗ no matter
what the initial state of our MDP is.

2 Value iteration and policy iteration

We now describe two efficient algorithms for solving finite-state MDPs. For
now, we will consider only MDPs with finite state and action spaces (|S| <
 5

∞, |A| < ∞). In this section, we will also assume that we know the state
transition probabilities {Psa } and the reward function R.
The first algorithm, value iteration, is as follows:

1. For each state s, initialize V (s) := 0.

2. Repeat until convergence {

P
For every state, update V (s) := R(s) + maxa∈A γ s′ Psa (s′ )V (s′ ).

This algorithm can be thought of as repeatedly trying to update the esti-

mated value function using Bellman Equations (2).
There are two possible ways of performing the updates in the inner loop of
the algorithm. In the first, we can first compute the new values for V (s) for
every state s, and then overwrite all the old values with the new values. This
is called a synchronous update. In this case, the algorithm can be viewed as
implementing a “Bellman backup operator” that takes a current estimate of
the value function, and maps it to a new estimate. (See homework problem
for details.) Alternatively, we can also perform asynchronous updates.
Here, we would loop over the states (in some order), updating the values one
at a time.
Under either synchronous or asynchronous updates, it can be shown that
value iteration will cause V to converge to V ∗ . Having found V ∗ , we can
then use Equation (3) to find the optimal policy.
Apart from value iteration, there is a second standard algorithm for find-
ing an optimal policy for an MDP. The policy iteration algorithm proceeds
as follows:

1. Initialize π randomly.

2. Repeat until convergence {

(a) Let V := V π .
P
(b) For each state s, let π(s) := arg maxa∈A s′ Psa (s′ )V (s′ ).

Thus, the inner-loop repeatedly computes the value function for the current
policy, and then updates the policy using the current value function. (The
policy π found in step (b) is also called the policy that is greedy with re-
spect to V .) Note that step (a) can be done via solving Bellman’s equations
 6

as described earlier, which in the case of a fixed policy, is just a set of |S|
linear equations in |S| variables.
After at most a finite number of iterations of this algorithm, V will con-
verge to V ∗ , and π will converge to π ∗ .
Both value iteration and policy iteration are standard algorithms for solv-
ing MDPs, and there isn’t currently universal agreement over which algo-
rithm is better. For small MDPs, policy iteration is often very fast and
converges with very few iterations. However, for MDPs with large state
spaces, solving for V π explicitly would involve solving a large system of lin-
ear equations, and could be difficult. In these problems, value iteration may
be preferred. For this reason, in practice value iteration seems to be used
more often than policy iteration.

3 Learning a model for an MDP

So far, we have discussed MDPs and algorithms for MDPs assuming that the
state transition probabilities and rewards are known. In many realistic prob-
lems, we are not given state transition probabilities and rewards explicitly,
but must instead estimate them from data. (Usually, S, A and γ are known.)
For example, suppose that, for the inverted pendulum problem (see prob-
lem set 4), we had a number of trials in the MDP, that proceeded as follows:
(1) (1) (1) (1)
(1) a
0 1 (1) a
2 3 (1) a (1) a
s0 −→ s1 −→ s2 −→ s3 −→ ...
(2) (2) (2) (2)
(2) a
0 1 (2) a
2 3 (2) a (2) a
s0 −→ s1 −→ s2 −→ s3 −→ ...
...
(j) (j)
Here, si is the state we were at time i of trial j, and ai is the cor-
responding action that was taken from that state. In practice, each of the
trials above might be run until the MDP terminates (such as if the pole falls
over in the inverted pendulum problem), or it might be run for some large
but finite number of timesteps.
Given this “experience” in the MDP consisting of a number of trials,
we can then easily derive the maximum likelihood estimates for the state
transition probabilities:
#times took we action a in state s and got to s′
Psa (s′ ) = (4)
#times we took action a in state s
Or, if the ratio above is “0/0”—corresponding to the case of never having
 7

taken action a in state s before—the we might simply estimate Psa (s′ ) to be

1/|S|. (I.e., estimate Psa to be the uniform distribution over all states.)
Note that, if we gain more experience (observe more trials) in the MDP,
there is an efficient way to update our estimated state transition probabilities
using the new experience. Specifically, if we keep around the counts for both
the numerator and denominator terms of (4), then as we observe more trials,
we can simply keep accumulating those counts. Computing the ratio of these
counts then given our estimate of Psa .
Using a similar procedure, if R is unknown, we can also pick our estimate
of the expected immediate reward R(s) in state s to be the average reward
observed in state s.
Having learned a model for the MDP, we can then use either value it-
eration or policy iteration to solve the MDP using the estimated transition
probabilities and rewards. For example, putting together model learning and
value iteration, here is one possible algorithm for learning in an MDP with
unknown state transition probabilities:
1. Initialize π randomly.
2. Repeat {
(a) Execute π in the MDP for some number of trials.
(b) Using the accumulated experience in the MDP, update our esti-
mates for Psa (and R, if applicable).
(c) Apply value iteration with the estimated state transition probabil-
ities and rewards to get a new estimated value function V .
(d) Update π to be the greedy policy with respect to V .
}
We note that, for this particular algorithm, there is one simple optimiza-
tion that can make it run much more quickly. Specifically, in the inner loop
of the algorithm where we apply value iteration, if instead of initializing value
iteration with V = 0, we initialize it with the solution found during the pre-
vious iteration of our algorithm, then that will provide value iteration with
a much better initial starting point and make it converge more quickly.

4 Continuous state MDPs

So far, we’ve focused our attention on MDPs with a finite number of states.
We now discuss algorithms for MDPs that may have an infinite number of
 8

states. For example, for a car, we might represent the state as (x, y, θ, ẋ, ẏ, θ̇),
comprising its position (x, y); orientation θ; velocity in the x and y directions
ẋ and ẏ; and angular velocity θ̇. Hence, S = R6 is an infinite set of states,
because there is an infinite number of possible positions and orientations
for the car.2 Similarly, the inverted pendulum you saw in PS4 has states
(x, θ, ẋ, θ̇), where θ is the angle of the pole. And, a helicopter flying in 3d
space has states of the form (x, y, z, φ, θ, ψ, ẋ, ẏ, ż, φ̇, θ̇, ψ̇), where here the roll
φ, pitch θ, and yaw ψ angles specify the 3d orientation of the helicopter.
In this section, we will consider settings where the state space is S = Rd ,
and describe ways for solving such MDPs.

4.1 Discretization
Perhaps the simplest way to solve a continuous-state MDP is to discretize
the state space, and then to use an algorithm like value iteration or policy
iteration, as described previously.
For example, if we have 2d states (s1 , s2 ), we can use a grid to discretize
the state space:

Here, each grid cell represents a separate discrete state s̄. We can then ap-
proximate the continuous-state MDP via a discrete-state one (S̄, A, {Ps̄a }, γ, R),
where S̄ is the set of discrete states, {Ps̄a } are our state transition probabil-
ities over the discrete states, and so on. We can then use value iteration or
policy iteration to solve for the V ∗ (s̄) and π ∗ (s̄) in the discrete state MDP
(S̄, A, {Ps̄a }, γ, R). When our actual system is in some continuous-valued
2
Technically, θ is an orientation and so the range of θ is better written θ ∈ [−π, π) than
θ ∈ R; but for our purposes, this distinction is not important.
 9

state s ∈ S and we need to pick an action to execute, we compute the

corresponding discretized state s̄, and execute action π ∗ (s̄).
This discretization approach can work well for many problems. However,
there are two downsides. First, it uses a fairly naive representation for V ∗
(and π ∗ ). Specifically, it assumes that the value function is takes a constant
value over each of the discretization intervals (i.e., that the value function is
piecewise constant in each of the gridcells).
To better understand the limitations of such a representation, consider a
supervised learning problem of fitting a function to this dataset:
5.5

4.5

3.5
y

2.5

1.5
1 2 3 4 5 6 7 8
x

Clearly, linear regression would do fine on this problem. However, if we

instead discretize the x-axis, and then use a representation that is piecewise
constant in each of the discretization intervals, then our fit to the data would
look like this:
5.5

4.5

3.5
y

2.5

1.5
1 2 3 4 5 6 7 8
x
 10

This piecewise constant representation just isn’t a good representation for

many smooth functions. It results in little smoothing over the inputs, and no
generalization over the different grid cells. Using this sort of representation,
we would also need a very fine discretization (very small grid cells) to get a
good approximation.
A second downside of this representation is called the curse of dimen-
sionality. Suppose S = Rd , and we discretize each of the d dimensions of the
state into k values. Then the total number of discrete states we have is k d .
This grows exponentially quickly in the dimension of the state space d, and
thus does not scale well to large problems. For example, with a 10d state, if
we discretize each state variable into 100 values, we would have 10010 = 1020
discrete states, which is far too many to represent even on a modern desktop
computer.
As a rule of thumb, discretization usually works extremely well for 1d
and 2d problems (and has the advantage of being simple and quick to im-
plement). Perhaps with a little bit of cleverness and some care in choosing
the discretization method, it often works well for problems with up to 4d
states. If you’re extremely clever, and somewhat lucky, you may even get it
to work for some 6d problems. But it very rarely works for problems any
higher dimensional than that.

4.2 Value function approximation

We now describe an alternative method for finding policies in continuous-
state MDPs, in which we approximate V ∗ directly, without resorting to dis-
cretization. This approach, called value function approximation, has been
successfully applied to many RL problems.

4.2.1 Using a model or simulator

To develop a value function approximation algorithm, we will assume that
we have a model, or simulator, for the MDP. Informally, a simulator is
a black-box that takes as input any (continuous-valued) state st and action
at , and outputs a next-state st+1 sampled according to the state transition
probabilities Pst at :
 11

There are several ways that one can get such a model. One is to use
physics simulation. For example, the simulator for the inverted pendulum
in PS4 was obtained by using the laws of physics to calculate what position
and orientation the cart/pole will be in at time t + 1, given the current state
at time t and the action a taken, assuming that we know all the parameters
of the system such as the length of the pole, the mass of the pole, and so
on. Alternatively, one can also use an off-the-shelf physics simulation software
package which takes as input a complete physical description of a mechanical
system, the current state st and action at , and computes the state st+1 of the
system a small fraction of a second into the future.3
An alternative way to get a model is to learn one from data collected in
the MDP. For example, suppose we execute n trials in which we repeatedly
take actions in an MDP, each trial for T timesteps. This can be done picking
actions at random, executing some specific policy, or via some other way of
choosing actions. We would then observe n state sequences like the following:
(1) (1) (1) (1)
(1) a
0 1 (1) a
2 (1) a aT −1 (1)
s0 −→ s1 −→ s2 −→ · · · −→ sT
(2) (2) (2) (2)
(2) a0 (2) a1 (2) a2 aT −1 (2)
s0 −→ s1 −→ s2 −→ · · · −→ sT
···
(n) (n) (n) (n)
(n) a0 (n) a1 (n) a2 aT −1 (n)
s0 −→ s1 −→ s2 −→ · · · −→ sT

We can then apply a learning algorithm to predict st+1 as a function of st

and at .
For example, one may choose to learn a linear model of the form

st+1 = Ast + Bat , (5)

3
Open Dynamics Engine (http://www.ode.com) is one example of a free/open-source
physics simulator that can be used to simulate systems like the inverted pendulum, and
that has been a reasonably popular choice among RL researchers.
 12

using an algorithm similar to linear regression. Here, the parameters of the

model are the matrices A and B, and we can estimate them using the data
collected from our n trials, by picking
n X
T −1   2
(i) (i) (i)
X
arg min st+1 − Ast + Bat .
A,B 2
i=1 t=0

(This corresponds to the maximum likelihood estimate of the parameters.)

We could also potentially use other loss functions for learning the model.
For example, it has been found in recent work [?] that using k · k2 norm
(without the square) may be helpful in certain cases.
Having learned A and B, one option is to build a deterministic model,
in which given an input st and at , the output st+1 is exactly determined.
Specifically, we always compute st+1 according to Equation (5). Alterna-
tively, we may also build a stochastic model, in which st+1 is a random
function of the inputs, by modeling it as

st+1 = Ast + Bat + ǫt ,

where here ǫt is a noise term, usually modeled as ǫt ∼ N (0, Σ). (The covari-
ance matrix Σ can also be estimated from data in a straightforward way.)
Here, we’ve written the next-state st+1 as a linear function of the current
state and action; but of course, non-linear functions are also possible. Specif-
ically, one can learn a model st+1 = Aφs (st ) + Bφa (at ), where φs and φa are
some non-linear feature mappings of the states and actions. Alternatively,
one can also use non-linear learning algorithms, such as locally weighted lin-
ear regression, to learn to estimate st+1 as a function of st and at . These
approaches can also be used to build either deterministic or stochastic sim-
ulators of an MDP.

4.2.2 Fitted value iteration

We now describe the fitted value iteration algorithm for approximating
the value function of a continuous state MDP. In the sequel, we will assume
that the problem has a continuous state space S = Rd , but that the action
space A is small and discrete.4
4
In practice, most MDPs have much smaller action spaces than state spaces. E.g., a car
has a 6d state space, and a 2d action space (steering and velocity controls); the inverted
pendulum has a 4d state space, and a 1d action space; a helicopter has a 12d state space,
and a 4d action space. So, discretizing this set of actions is usually less of a problem than
discretizing the state space would have been.
 13

Recall that in value iteration, we would like to perform the update

Z
V (s) := R(s) + γ max Psa (s′ )V (s′ )ds′ (6)
a s′
= R(s) + γ max Es′ ∼Psa [V (s′ )] (7)
a

(In Section 2, we had written

P the value iteration update with a summation
V (s) := R(s) + γ maxa s′ Psa (s′ )V (s′ ) rather than an integral over states;
the new notation reflects that we are now working in continuous states rather
than discrete states.)
The main idea of fitted value iteration is that we are going to approxi-
mately carry out this step, over a finite sample of states s(1) , . . . , s(n) . Specif-
ically, we will use a supervised learning algorithm—linear regression in our
description below—to approximate the value function as a linear or non-linear
function of the states:
V (s) = θT φ(s).
Here, φ is some appropriate feature mapping of the states.
For each state s in our finite sample of n states, fitted value itera-
tion will first compute a quantity y (i) , which will be our approximation
to R(s) + γ maxa Es′ ∼Psa [V (s′ )] (the right hand side of Equation 7). Then,
it will apply a supervised learning algorithm to try to get V (s) close to
R(s) + γ maxa Es′ ∼Psa [V (s′ )] (or, in other words, to try to get V (s) close to
y (i) ).
In detail, the algorithm is as follows:

1. Randomly sample n states s(1) , s(2) , . . . s(n) ∈ S.

2. Initialize θ := 0.

3. Repeat {

For i = 1, . . . , n {
For each action a ∈ A {
Sample s′1 , . . . , s′k ∼ Ps(i) a (using a model of the MDP).
Set q(a) = k1 kj=1 R(s(i) ) + γV (s′j )
P

// Hence, q(a) is an estimate of R(s(i) )+γEs′ ∼Ps(i) a [V (s′ )].

}
Set y (i) = maxa q(a).
// Hence, y (i) is an estimate of R(s(i) )+γ maxa Es′ ∼Ps(i) a [V (s′ )].
 14

}
// In the original value iteration algorithm (over discrete states)
// we updated the value function according to V (s(i) ) := y (i) .
// In this algorithm, we want V (s(i) ) ≈ y (i) , which we’ll achieve
// using supervised learning (linear regression).

2
Set θ := arg minθ 21 ni=1 θT φ(s(i) ) − y (i)
P

Above, we had written out fitted value iteration using linear regression
as the algorithm to try to make V (s(i) ) close to y (i) . That step of the algo-
rithm is completely analogous to a standard supervised learning (regression)
problem in which we have a training set (x(1) , y (1) ), (x(2) , y (2) ), . . . , (x(n) , y (n) ),
and want to learn a function mapping from x to y; the only difference is that
here s plays the role of x. Even though our description above used linear re-
gression, clearly other regression algorithms (such as locally weighted linear
regression) can also be used.
Unlike value iteration over a discrete set of states, fitted value iteration
cannot be proved to always to converge. However, in practice, it often does
converge (or approximately converge), and works well for many problems.
Note also that if we are using a deterministic simulator/model of the MDP,
then fitted value iteration can be simplified by setting k = 1 in the algorithm.
This is because the expectation in Equation (7) becomes an expectation over
a deterministic distribution, and so a single example is sufficient to exactly
compute that expectation. Otherwise, in the algorithm above, we had to
draw k samples, and average to try to approximate that expectation (see the
definition of q(a), in the algorithm pseudo-code).
Finally, fitted value iteration outputs V , which is an approximation to
V . This implicitly defines our policy. Specifically, when our system is in
∗

some state s, and we need to choose an action, we would like to choose the
action
arg max Es′ ∼Psa [V (s′ )] (8)
a

The process for computing/approximating this is similar to the inner-loop of

fitted value iteration, where for each action, we sample s′1 , . . . , s′k ∼ Psa to
approximate the expectation. (And again, if the simulator is deterministic,
we can set k = 1.)
In practice, there are often other ways to approximate this step as well.
For example, one very common case is if the simulator is of the form st+1 =
 15

f (st , at ) + ǫt , where f is some deterministic function of the states (such as

f (st , at ) = Ast + Bat ), and ǫ is zero-mean Gaussian noise. In this case, we
can pick the action given by

arg max V (f (s, a)).

a

In other words, here we are just setting ǫt = 0 (i.e., ignoring the noise in
the simulator), and setting k = 1. Equivalent, this can be derived from
Equation (8) using the approximation

Es′ [V (s′ )] ≈ V (Es′ [s′ ]) (9)

= V (f (s, a)), (10)

where here the expectation is over the random s′ ∼ Psa . So long as the noise
terms ǫt are small, this will usually be a reasonable approximation.
However, for problems that don’t lend themselves to such approximations,
having to sample k|A| states using the model, in order to approximate the
expectation above, can be computationally expensive.
 CS229 Lecture notes
Tengyu Ma

Part XV
Policy Gradient
(REINFORCE)
We will present a model-free algorithm called REINFORCE that does not
require the notion of value functions and Q functions. It turns out to be more
convenient to introduce REINFORCE in the finite horizon case, which will
be assumed throughout this note: we use τ = (s0 , a0 , . . . , sT −1 , aT −1 , sT ) to
denote a trajectory, where T < ∞ is the length of the trajectory. Moreover,
REINFORCE only applies to learning a randomized policy. We use πθ (a|s)
to denote the probability of the policy πθ outputting the action a at state s.
The other notations will be the same as in previous lecture notes.
The advantage of applying REINFORCE is that we only need to assume
that we can sample from the transition probabilities {Psa } and can query the
reward function R(s, a) at state s and action a,1 but we do not need to know
the analytical form of the transition probabilities or the reward function.
We do not explicitly learn the transition probabilities or the reward function
either.
Let s0 be sampled from some distribution µ. We consider optimizing the
expected total payoff of the policy πθ over the parameter θ defined as.
"T −1 #
X
η(θ) , E γ t R(st , at ) (1)
t=0

Recall that st ∼ Pst−1 at−1 and at ∼ πθ (·|st ). Also note that η(θ) =
Es0 ∼P [V πθ (s0 )] if we ignore the difference between finite and infinite hori-
zon.
1
In this notes we will work with the general setting where the reward depends on both
the state and the action.

1
 2

We aim to use gradient ascent to maximize η(θ). The main challenge

we face here is to compute (or estimate) the gradient of η(θ) without the
knowledge of the form of the reward function and the transition probabilities.
Let Pθ (τP) denote the distribution of τ (generated by the policy πθ ), and
−1 t
let f (τ ) = Tt=0 γ R(st , at ). We can rewrite η(θ) as

η(θ) = Eτ ∼Pθ [f (τ )] (2)

We face a similar situations in the variational auto-encoder (VAE) setting

covered in the previous lectures, where the we need to take the gradient w.r.t
to a variable that shows up under the expectation — the distribution Pθ
depends on θ. Recall that in VAE, we used the re-parametrization techniques
to address this problem. However it does not apply here because we do
know not how to compute the gradient of the function f . (We only have
an efficient way to evaluate the function f by taking a weighted sum of the
observed rewards, but we do not necessarily know the reward function itself
to compute the gradient.)
The REINFORCE algorithm uses an another approach to estimate the
gradient of η(θ). We start with the following derivation:
Z
∇θ Eτ ∼Pθ [f (τ )] = ∇θ Pθ (τ )f (τ )dτ
Z
= ∇θ (Pθ (τ )f (τ ))dτ (swap integration with gradient)
Z
= (∇θ Pθ (τ ))f (τ )dτ (becaue f does not depend on θ)
Z
= Pθ (τ )(∇θ log Pθ (τ ))f (τ )dτ
∇Pθ (τ )
(because ∇ log Pθ (τ ) = Pθ (τ )
)
= Eτ ∼Pθ [(∇θ log Pθ (τ ))f (τ )] (3)

Now we have a sample-based estimator for ∇θ Eτ ∼Pθ [f (τ )]. Let τ (1) , . . . , τ (n)
be n empirical samples from Pθ (which are obtained by running the policy
πθ for n times, with T steps for each run). We can estimate the gradient of
η(θ) by

∇θ Eτ ∼Pθ [f (τ )] = Eτ ∼Pθ [(∇θ log Pθ (τ ))f (τ )] (4)

n
1X
≈ (∇θ log Pθ (τ (i) ))f (τ (i) ) (5)
n i=1
 3

The next question is how to compute log Pθ (τ ). We derive an analyt-

ical formula for log Pθ (τ ) and compute its gradient w.r.t θ (using auto-
differentiation). Using the definition of τ , we have

Pθ (τ ) = µ(s0 )πθ (a0 |s0 )Ps0 a0 (s1 )πθ (a1 |s1 )Ps1 a1 (s2 ) · · · PsT −1 aT −1 (sT ) (6)

Here recall that µ to used to denote the density of the distribution of s0 . It

follows that

log Pθ (τ ) = log µ(s0 ) + log πθ (a0 |s0 ) + log Ps0 a0 (s1 ) + log πθ (a1 |s1 )
+ log Ps1 a1 (s2 ) + · · · + log PsT −1 aT −1 (sT ) (7)

Taking gradient w.r.t to θ, we obtain

∇θ log Pθ (τ ) = ∇θ log πθ (a0 |s0 ) + ∇θ log πθ (a1 |s1 ) + · · · + ∇θ log πθ (aT −1 |sT −1 )

Note that many of the terms disappear because they don’t depend on θ and
thus have zero gradients. (This is somewhat important — we don’t know how
to evaluate those terms such as log Ps0 a0 (s1 ) because we don’t have access to
the transition probabilities, but luckily those terms have zero gradients!)
Plugging the equation above into equation (4), we conclude that
" T −1 ! #
X
∇θ η(θ) = ∇θ Eτ ∼Pθ [f (τ )] = Eτ ∼Pθ ∇θ log πθ (at |st ) · f (τ )
t=0
" T −1
! T −1
!#
X X
= Eτ ∼Pθ ∇θ log πθ (at |st ) · γ t R(st , at )
t=0 t=0
(8)

We estimate the RHS of the equation above by empirical sample trajectories,

and the estimate is unbiased. The vanilla REINFORCE algorithm iteratively
updates the parameter by gradient ascent using the estimated gradients.

Interpretation
PT −1of the policy gradient formula (8). The quantity
∇θ Pθ (τ ) = t=0 ∇ θ log π (a |s
θ t t ) is intuitively the direction of the change
of θ that will make the trajectory τ more likely to occur (or increase the
probability of choosing action a0 , . . . , at−1 ), and f (τ ) is the total payoff of
this trajectory. Thus, by taking a gradient step, intuitively we are trying to
improve the likelihood of all the trajectories, but with a different emphasis
or weight for each τ (or for each set of actions a0 , a1 , . . . , at−1 ). If τ is very
rewarding (that is, f (τ ) is large), we try very hard to move in the direction
 4

that can increase the probability of the trajectory τ (or the direction that
increases the probability of choosing a0 , . . . , at−1 ), and if τ has low payoff,
we try less hard with a smaller weight.
An interesting fact that follows from formula (3) is that
"T −1 #
X
Eτ ∼Pθ ∇θ log πθ (at |st ) = 0 (9)
t=0

To see this, we take f (τ ) = 1 (that is, the reward is always a constant), PT then
the LHS of (8) is zero because the payoff is always a fixed constant t=0 γ t .
Thus the RHS of (8) is also zero, which implies (9).
In fact, one can verify that Eat ∼πθ (·|st ) ∇θ log πθ (at |st ) = 0 for any fixed t
and st .2 This fact has two consequences. First, we can simplify formula (8)
to
T −1
" T −1
!#
X X
j
∇θ η(θ) = Eτ ∼Pθ ∇θ log πθ (at |st ) · γ R(sj , aj )
t=0 j=0
T −1
" T −1
!#
X X
= Eτ ∼Pθ ∇θ log πθ (at |st ) · γ j R(sj , aj ) (10)
t=0 j≥t

where the second equality follows from

" !#
X
Eτ ∼Pθ ∇θ log πθ (at |st ) · γ j R(sj , aj )
0≤j<t
" !#
X
= E E [∇θ log πθ (at |st )|s0 , a0 , . . . , st−1 , at−1 , st ] · γ j R(sj , aj )
0≤j<t

=0 (because E [∇θ log πθ (at |st )|s0 , a0 , . . . , st−1 , at−1 , st ] = 0)

Note that here we used the law of total expectation. The outer expecta-
tion in the second line above is over the randomness of s0 , a0 , . . . , at−1 , st ,
whereas the inner expectation is over the randomness of at (conditioned on
s0 , a0 , . . . , at−1 , st .) We see that we’ve made the estimator slightly simpler.
The second consequence of Eat ∼πθ (·|st ) ∇θ log πθ (at |st ) = 0 is the following: for
any value B(st ) that only depends on st , it holds that
Eτ ∼Pθ [∇θ log πθ (at |st ) · B(st )]
= E [E [∇θ log πθ (at |st )|s0 , a0 , . . . , st−1 , at−1 , st ] B(st )]
=0 (because E [∇θ log πθ (at |st )|s0 , a0 , . . . , st−1 , at−1 , st ] = 0)
2
In general, it’s true that Ex∼pθ [∇ log pθ (x)] = 0.
 5

Again here we used the law of total expectation. The outer expecta-
tion in the second line above is over the randomness of s0 , a0 , . . . , at−1 , st ,
whereas the inner expectation is over the randomness of at (conditioned on
s0 , a0 , . . . , at−1 , st .) It follows from equation (10) and the equation above that
T −1
" T −1
!#
X X
∇θ η(θ) = Eτ ∼Pθ ∇θ log πθ (at |st ) · γ j R(sj , aj ) − γ t B(st )
t=0 j≥t
T −1
" T −1
!#
X X
t j−t
= Eτ ∼Pθ ∇θ log πθ (at |st ) · γ γ R(sj , aj ) − B(st )
t=0 j≥t
(11)

Therefore, we will get a different estimator for estimating the ∇η(θ) with a
difference choice of B(·). The benefit of introducing a proper B(·) — which
is often referred to as a baseline — is that it helps reduce the variance of the
estimator.3 It turns
hP out that a near optimal i estimator would be the expected
T −1 j−t
future payoff E j≥t γ R(sj , aj )|st , which is pretty much the same as the
value function V πθ (st ) (if we ignore the difference between finite and infinite
horizon.) Here one could estimate the value function V πθ (·) in a crude way,
because its precise value doesn’t influence the mean of the estimator but only
the variance. This leads to a policy gradient algorithm with baselines stated
in Algorithm 1.4

3
PT −1 As a heuristic but illustrating example, suppose for a fixed t, the future reward
j−t
j≥t γ R(sj , aj ) randomly takes two values 1000 + 1 and 1000 − 2 with equal proba-
bility, and the corresponding values for ∇θ log πθ (at |st ) are vector z and −z. (Note that
because E [∇θ log πθ (at |st )] = 0, if ∇θ log πθ (at |st ) can only take two values uniformly,
then the two values have to two vectors in an opposite direction.) In this case, without
subtracting the baseline, the estimators take two values (1000 + 1)z and −(1000 − 2)z,
whereas after subtracting a baseline of 1000, the estimator has two values z and 2z. The
latter estimator has much lower variance compared to the original estimator.
4
We note that the estimator of the gradient in the algorithm does not exactly match the
equation 11. If we multiply γ t in the summand of equation (13), then they will exactly
match. Removing such discount factors empirically works well because it gives a large
update.
 6

Algorithm 1 Vanilla policy gradient with baseline

for i = 1, · · · do
Collect a set of P
trajectories by executing the current policy. Use R≥t
−1 j−t
as a shorthand for Tj≥t γ R(sj , aj )
Fit the baseline by finding a function B that minimizes
XX
(R≥t − B(st ))2 (12)
τ t

Update the policy parameter θ with the gradient estimator

XX
∇θ log πθ (at |st ) · (R≥t − B(st )) (13)
τ t
 CS229 Supplemental Lecture notes
John Duchi

1 Boosting
We have seen so far how to solve classification (and other) problems when we
have a data representation already chosen. We now talk about a procedure,
known as boosting, which was originally discovered by Rob Schapire, and
further developed by Schapire and Yoav Freund, that automatically chooses
feature representations. We take an optimization-based perspective, which
is somewhat different from the original interpretation and justification of
Freund and Schapire, but which lends itself to our approach of (1) choose a
representation, (2) choose a loss, and (3) minimize the loss.
Before formulating the problem, we give a little intuition for what we
are going to do. Roughly, the idea of boosting is to take a weak learning
algorithm—any learning algorithm that gives a classifier that is slightly bet-
ter than random—and transforms it into a strong classifier, which does much
much better than random. To build a bit of intuition for what this means,
consider a hypothetical digit recognition experiment, where we wish to dis-
tinguish 0s from 1s, and we receive images we must classify. Then a natural
weak learner might be to take the middle pixel of the image, and if it is
colored, call the image a 1, and if it is blank, call the image a 0. This clas-
sifier may be far from perfect, but it is likely better than random. Boosting
procedures proceed by taking a collection of such weak classifiers, and then
reweighting their contributions to form a classifier with much better accuracy
than any individual classifier.
With that in mind, let us formulate the problem. Our interpretation of
boosting is as a coordinate descent method in an infinite dimensional space,
which—while it sounds complex—is not so bad as it seems. First, we assume
we have raw input examples x ∈ Rn with labels y ∈ {−1, 1}, as is usual in
binary classification. We also assume we have an infinite collection of feature
functions φj : Rn → {−1, 1} and an infinite vector θ = [θ1 θ2 · · · ]T , but

1
which we assume always has only a finite number of non-zero entries. For
our classifier we use
X
∞ 
hθ (x) = sign θj φj (x) .
j=1

P
We will abuse notation, and define θT φ(x) = ∞ j=1 θj φj (x).
In boosting, one usually calls the features φj weak hypotheses. Given a
training set (x(1) , y (1) ), . . . , (x(m) , y (m) ), we call a vector p = (p(1) , . . . , p(m) ) a
distribution on the examples if p(i) ≥ 0 for all i and
X
m
p(i) = 1.
i=1

Then we say that there is a weak learner with margin γ > 0 if for any
distribution p on the m training examples there exists one weak hypothesis
φj such that
X m
 1
p(i) 1 y (i) 6= φj (x(i) ) ≤ − γ. (1)
i=1
2
That is, we assume that there is some classifier that does slightly better than
random guessing on the dataset. The existence of a weak learning algorithm
is an assumption, but the surprising thing is that we can transform any weak
learning algorithm into one with perfect accuracy.
In more generality, we assume we have access to a weak learner, which is
an algorithm that takes as input a distribution (weights) p on the training
examples and returns a classifier doing slightly better than random. We will

(i) Input:
PmA distribution p(1) , . . . , p(m) and training set {(x(i) , y (i) )}m
i=1
with i=1 p(i) = 1 and p(i) ≥ 0

(ii) Return: A weak classifier φj : Rn → {−1, 1} such that

X
m
 1
p(i) 1 y (i) 6= φj (x(i) ) ≤ − γ.
i=1
2

Figure 1: Weak learning algorithm

2
show how, given access to a weak learning algorithm, boosting can return a
classifier with perfect accuracy on the training data. (Admittedly, we would
like the classifer to generalize well to unseen data, but for now, we ignore
this issue.)

1.1 The boosting algorithm

Roughly, boosting begins by assigning each training example equal weight
in the dataset. It then receives a weak-hypothesis that does well according
to the current weights on training examples, which it incorporates into its
current classification model. It then reweights the training examples so that
examples on which it makes mistakes receive higher weight—so that the weak
learning algorithm focuses on a classifier doing well on those examples—while
examples with no mistakes receive lower weight. This repeated reweighting
of the training data coupled with a weak learner doing well on examples for
which the classifier currently does poorly yields classifiers with good perfor-
mance.
The boosting algorithm specifically performs coordinate descent on the
exponential loss for classification problems, where the objective is
1 X
m
J(θ) = exp(−y (i) θT φ(x(i) )).
m i=1

We first show how to compute the exact form of the coordinate descent
update for the risk J(θ). Coordinate descent iterates as follows:
(i) Choose a coordinate j ∈ N
(ii) Update θj to
θj = arg min J(θ)
θj

while leaving θk identical for all k 6= j.

We iterate the above procedure until convergence.
In the case of boosting, the coordinate updates are not too challenging to
derive because of the analytic convenience of the exp function. We now show
how to derive the update. Suppose we wish to update coordinate k. Define
!
X
w(i) = exp −y (i) θj φj (x(i) )
j6=k

3
to be a weight, and note that optimizing coordinate k corresponds to mini-
mizing
Xm
w(i) exp(−y (i) φk (x(i) )α)
i=1

in α = θk . Now, define
X X
W + := w(i) and W − := w(i)
i:y (i) φk (x(i) )=1 i:y (i) φk (x(i) )=−1

to be the sums of the weights of examples that φk classifies correctly and

incorrectly, respectively. Then finding θk is the same as choosing
 1 W+
α = arg min W + e−α + W − eα = log − .
α 2 W
To see the final equality, take derivatives and set the resulting equation to
zero, so we have −W + e−α + W − eα = 0. That is, W − e2α = W + , or α =
1 W+
2
log W −.

What remains is to choose the particular coordinate to perform coordinate

descent on. We assume we have access to a weak-learning algorithm as in
Figure 1, which at iteration t takes as input a distribution p on the training
set and returns a weak hypothesis φt satisfying the margin condition (1).
We present the full boosting algorithm in Figure 2. It proceeds in iterations
t = 1, 2, 3, . . .. We represent the set of hypotheses returned by the weak
learning algorithm at time t by {φ1 , . . . , φt }.

2 The convergence of Boosting

We now argue that the boosting procedure achieves 0 training error, and we
also provide a rate of convergence to zero. To do so, we present a lemma
that guarantees progress is made.
Lemma 2.1. Let
 
1 X X
m t
(t) (i) (i)
J(θ ) = exp −y θτ φτ (x ) .
m i=1 τ =1

Then p
J(θ(t) ) ≤ 1 − 4γ 2 J(θ(t−1) ).

4
 For each iteration t = 1, 2, . . .:

(i) Define weights

 X
t−1 
(i) (i) (i)
w = exp −y θτ φτ (x )
τ =1
Pm
and distribution p(i) = w(i) / j=1 w(j)

(ii) Construct a weak hypothesis φt : Rn → {−1, 1} from the distribu-

tion p = (p(1) , . . . , p(m) ) on the training set
P P
(iii) Compute Wt+ = i:y(i) φt (x(i) )=1 w(i) and Wt− = i:y(i) φt (x(i) )=−1 w(i)
and set
1 W+
θt = log t− .
2 Wt

Figure 2: Boosting algorithm

As the proof of the lemma is somewhat involved and not the central focus of
these notes—though it is important to know one’s algorithm will converge!—
we defer the proof to Appendix A.1. Let us describe how it guarantees
convergence of the boosting procedure to a classifier with zero training error.
We initialize the procedure at θ(0) = ~0, so that the initial empirical risk
J(θ(0) ) = 1. Now, we note that for any θ, the misclassification error satisfies
 

1 sign(θT φ(x)) 6= y = 1 yθT φ(x) ≤ 0 ≤ exp −yθT φ(x)

because ez ≥ 1 for all z ≥ 0. Thus, we have that the misclassification error

rate has upper bound

1 X 
m
1 sign(θT φ(x(i) )) 6= y (i) ≤ J(θ),
m i=1

and so if J(θ) < m1 then the vector θ makes no mistakes on the training data.
After t iterations of boosting, we find that the empirical risk satisfies
t t
J(θ(t) ) ≤ (1 − 4γ 2 ) 2 J(θ(0) ) = (1 − 4γ 2 ) 2 .

5
 1
To find how many iterations are required to guarantee J(θ(t) ) < m
, we take
logarithms to find that J(θ(t) ) < 1/m if
t 1 2 log m
log(1 − 4γ 2 ) < log , or t > .
2 m − log(1 − 4γ 2 )

Using a first order Taylor expansion, that is, that log(1 − 4γ 2 ) ≤ −4γ 2 , we
see that if the number of rounds of boosting—the number of weak classifiers
we use—satisfies
log m 2 log m
t> 2
≥ ,
2γ − log(1 − 4γ 2 )
1
then J(θ(t) ) < m
.

3 Implementing weak-learners
One of the major advantages of boosting algorithms is that they automat-
ically generate features from raw data for us. Moreover, because the weak
hypotheses always return values in {−1, 1}, there is no need to normalize fea-
tures to have similar scales when using learning algorithms, which in practice
can make a large difference. Additionally, and while this is not theoret-
ically well-understood, many types of weak-learning procedures introduce
non-linearities intelligently into our classifiers, which can yield much more
expressive models than the simpler linear models of the form θT x that we
have seen so far.

3.1 Decision stumps

There are a number of strategies for weak learners, and here we focus on
one, known as decision stumps. For concreteness in this description, let
us suppose that the input variables x ∈ Rn are real-valued. A decision
stump is a function f , which is parameterized by a threshold s and index
j ∈ {1, 2, . . . , n}, and returns
(
1 if xj ≥ s
φj,s (x) = sign(xj − s) = (2)
−1 otherwise.

These classifiers are simple enough that we can fit them efficiently even to a
weighted dataset, as we now describe.

6
 Indeed, a decision stump weak learner proceeds as follows. We begin with
a distribution—set of weights p(1) , . . . , p(m) summing to 1—on the training
set, and we wish to choose a decision stump of the form (2) to minimize the
error on the training set. That is, we wish to find a threshold s ∈ R and
index j such that
X
m
 X
m n o
c j,s , p) =
Err(φ (i)
p 1 φj,s (x ) 6= y (i) (i)
=
(i)
p(i) 1 y (i) (xj − s) ≤ 0 (3)
i=1 i=1

is minimized. Naively, this could be an inefficient calculation, but a more

intelligent procedure allows us to solve this problem in roughly O(nm log m)
time. For each feature j = 1, 2, . . . , n, we sort the raw input features so that
(i ) (i ) (i )
xj 1 ≥ xj 2 ≥ · · · ≥ xj m .

As the only values s for which the error of the decision stump can change
(i)
are the values xj , a bit of clever book-keeping allows us to compute

X
m n o Xm n o
(i) (i )
p(i) 1 y (i) (xj − s) ≤ 0 = p(ik ) 1 y (ik ) (xj k − s) ≤ 0
i=1 k=1

efficiently by incrementally modifying the sum in sorted order, which takes

(i)
time O(m) after we have already sorted the values xj . (We do not describe
the algorithm in detail here, leaving that to the interested reader.) Thus,
performing this calcuation for each of the n input features takes total time
O(nm log m), and we may choose the index j and threshold s that give the
best decision stump for the error (3).
One very important issue to note is that by flipping the sign of the thresh-
c j,s , p), that is, the error
olded decision stump φj,s , we achieve error 1 − Err(φ
of
c
Err(−φ c
j,s , p) = 1 − Err(φj,s , p).

(You should convince yourself that this is true.) Thus, it is important to also
c j,s , p) over all thresholds, because this
track the smallest value of 1 − Err(φ
c j,s , p), which gives a better weak learner. Using
may be smaller than Err(φ
this procedure for our weak learner (Fig. 1) gives the basic, but extremely
useful, boosting classifier.

7
 1
0.9
0.8
0.7
0.6
0.5
0.4
0.3
0.2
0.1
0
0 0.2 0.4 0.6 0.8 1

Figure 3: Best logistic regression classifier using the raw features x ∈ R2

(and a bias term x0 = 1) for the example considered here.

3.2 Example
We now give an example showing the behavior of boosting on a simple
dataset. In particular, we consider a problem with data points x ∈ R2 ,
where the optimal classifier is
(
1 if x1 < .6 and x2 < .6
y= (4)
−1 otherwise.

This is a simple non-linear decision rule, but it is impossible for standard

linear classifiers, such as logistic regression, to learn. In Figure 3, we show
the best decision line that logistic regression learns, where positive examples
are circles and negative examples are x’s. It is clear that logistic regression
is not fitting the data particularly well.
With boosted decision stumps, however, we can achieve a much better
fit for the simple nonlinear classification problem (4). Figure 4 shows the
boosted classifiers we have learned after different numbers of iterations of
boosting, using a training set of size m = 150. From the figure, we see that
the first decision stump is to threshold the feature x1 at the value s ≈ .23,
that is, φ(x) = sign(x1 − s) for s ≈ .23.

8
 1
2 Iterations 1
4 Iterations
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

1
5 Iterations 1
10 Iterations
0.9 0.9
0.8 0.8
0.7 0.7
0.6 0.6
0.5 0.5
0.4 0.4
0.3 0.3
0.2 0.2
0.1 0.1
0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1

Figure 4: Boosted decision stumps after t = 2, 4, 5, and 10 iterations of

boosting, respectively.

9
3.3 Other strategies
There are a huge number of variations on the basic boosted decision stumps
idea. First, we do not require that the input features xj be real-valued. Some
of them may be categorical, meaning that xj ∈ {1, 2, . . . , k} for some k, in
which case natural decision stumps are of the form
(
1 if xj = l
φj (x) =
−1 otherwise,

as well as variants setting φj (x) = 1 if xj ∈ C for some set C ⊂ {1, . . . , k} of

categories.
Another natural variation is the boosted decision tree, in which instead of a
single level decision for the weak learners, we consider conjuctions of features
or trees of decisions. Google can help you find examples and information on
these types of problems.

A Appendices
A.1 Proof of Lemma 2.1
We now return to prove the progress lemma. We prove this result by directly
showing the relationship of the weights at time t to those at time t − 1. In
particular, we note by inspection that
q
J(θ ) = min{Wt e + Wt e } = 2 Wt+ Wt−
(t) + −α − α
α

while
 
1 X X
m t−1
(t−1) (i) (i)
J(θ )= exp −y θτ φτ (x ) = Wt+ + Wt− .
m i=1 τ =1

We know by the weak-learning assumption that

X
m
 1 1 X 1
p(i) 1 y (i) 6= φt (x(i) ) ≤ − γ, or + w(i) ≤ − γ.
i=1
2 Wt + Wt− 2
i:y (i) φt (x(i) )=−1
| {z }
=Wt−

10
Rewriting this expression by noting that the sum on the right is nothing but
Wt− , we have
 
1 1 + 2γ −
−
Wt ≤ − γ (Wt+ + Wt− ), or Wt+ ≥ W .
2 1 − 2γ t

By substituting α = 21 log 1+2γ

1−2γ
in the minimum defining J(θ(t) ), we obtain
r r
1 − 2γ 1 + 2γ
J(θ(t) ) ≤ Wt+ + Wt−
1 + 2γ 1 − 2γ
r r
+ 1 − 2γ − 1 + 2γ
= Wt + Wt (1 − 2γ + 2γ)
1 + 2γ 1 − 2γ
r r r
+ 1 − 2γ − 1 + 2γ 1 − 2γ 1 + 2γ +
≤ Wt + Wt (1 − 2γ) + 2γ W
1 + 2γ 1 − 2γ 1 + 2γ 1 − 2γ t
r r  p
+ 1 − 2γ 1 − 2γ
= Wt + 2γ + Wt− 1 − 4γ 2 ,
1 + 2γ 1 + 2γ
1−2γ
where we used that Wt− ≤ 1+2γ Wt+ . Performing a few algebraic manipula-
tions, we see that the final expression is equal to
p
1 − 4γ 2 (Wt+ + Wt− ).
p
That is, J(θ(t) ) ≤ 1 − 4γ 2 J(θ(t−1) ).

11