Lecture 1: Introduction, Entropy and ML Estimation

10-704: Information Processing and Learning
Spring 2012
Lecture 1: Introduction, Entropy and ML estimation

Lecturer: Aarti Singh
Scribes: Min Xu
Disclaimer: These notes have not been subjected to the usual scrutiny reserved for formal publications.
They may be distributed outside this class only with the permission of the Instructor.
1.1
About the Class
This class focuses on information theory, signal processing, machine learning, and the connections between
these fields.
signals can be audio, speech, music. data can be images, files. They overlap a lot of times.
Both signal processing and machine learning are about how to extract useful information from signals/data.
signals as used in the EE community can be different from data in that (1) they often have temporal
aspect, (2) they are often designed and (3) they are often transmitted through a medium (known as a
channel ).
Information theory studies 2 main questions:
1. How much information is contained in the signal/data?
Example: Consider the data compression (source coding) problem.
Source Compressor Decompressor Receiver
What is the fewest number of bits needed to describe the output of a source (also called message) while
preserving all the information, in the sense that a receiver can reconstruct the message from the bits
with arbitrarily low probability of error?
2. How much information can be reliably transmitted through a noisy channel?
Example: Consider the data transmission (channel coding) problem.
Source Encoder Channel Decoder Receiver
What is the maximum number of bits per channel use that can be reliably sent through a noisy channel,
in the sense that the receiver can reconstruct the source message with arbitrarily low probability of
error?
Remark: The data compression or source coding problem can be thought of as a noiseless version of the data
tranmission or channel coding problem.
1-1
1-2
Connection to Machine Learning:

1. Source Coding in ML: In ML, the source is essentially a model (e.g. p(X1 , ..., Xn )) that generates data
points X1 , ..., Xn , and the least number of bits needed to encode these data reflect the complexity
of the source or model. Thus, source coding can be used to pick a descriptive model with the least
complexity. This is the principle of Occams Razor.
2. Channel Coding in ML: The channel specifies a distribution p(y | x) where x is the input to the channel
and y is the output. For instance, we can view the output yi = m(xi ) + in regression as the output
of a noisy channel that takes m(xi ) as input. Similarly, in density estimation, x can be a parameter
and y is a sample generated according to p(y | x).
We will formalize these notions later in this course.
1.2
Information Content of Outcomes of Random Experiments
We will usually specify information content in bits, where a bit is defined to be of value either 0 or 1. We can
think of it as the output of a yes/no question, and the information content can be specified as the minimum
number of yes/no questions needed to lean the outcome of a random experiment.
1. We have an integer chosen randomly from 0 to 63. What is the smallest number of yes/no questions
needed to identify that integer? Answer: log2 (64) = 6 bits.
1
of being the correct integer. Thus,
Note: all integers from 0 to 63 have equal probability p = 64
1
log2 (64) = log2 ( p ). This is the Shannon Information Content.
2. Now lets consider an experiment where questions that do not lead to equi-probable outcomes.
An enemy ship is somewhere in an 8 8 grid (64 possible locations). We can launch a missile that hits
one location. Since the ship can be hidden in any of the 64 possible locations, we expect that we will
still gain 6 bits of information when we find the ship. However, each question (firing of a missile) now
may not provide the same amount of information.
1
The probability of hitting on first launch is p1 (h) = 64
, so the Shannon Information Content of hitting
1
on first launch is log2 ( p1 (h) ) = 6 bits. Since this was a low probability event, we gained a lot of
information (in fact all the infomation we hoped to gain on discovering the ship). However, we will
not gain the same amount of information on more probably events. For example:
The information gained from missing on the first launch is log2 ( 64

63 ) = 0.0227 bits
The information gained from missing on the first 32 launches is
log2 (p1 (m)) + log2 (p2 (m)) + ... log2 (p32 (m)) = log2 (
32
Y
pi (m))
i=1
64 63 33
... )
63 62 32
= log2 (2) = 1 bit
= log2 (
Which is what we intuitively expect, since ruling out 32 locations is equivalent to asking 1 question in
the previous experiment 1.
1-3
1
If we hit on the next try, we will gain log2 ( p3 3(h)
) = log2 (32) = 5 bits of information. Simple calculation
will show that, regardless of how many launches we needed, we gain a total of 6 bits of information
whenever we hit the ship.
3. What if the questions are allowed to have more than 2 answers?

Suppose we have a number of identically looking balls one of which is either heavier or lighter. We
have a balance and we want to find the odd ball with fewest number of weighings. There are now three
possible outputs of an experiment: left lighter, left heavier, equal weight. The minimum number of
experiments we need is then log3 (number of balls). Note that we may need more; information bounds
are often not achievable.
1.3
Information Content of Random Variables
A random variable is simply an assignment of probability to outcomes of a random experiment. We then

define the information content of a random variable as just the average Shannon Information Content. We
can also think of it as a measure of the uncertainty of the random variable.
Definition 1.1 The entropy of a random variable X with probability distribution p(x) is
X
1
= Ep [log p(x)]
H(X) =
p(x) log2
p(x)
xX
Where X is set of all possible values of the random variable X. We often also write H(X) as H(p) since
entropy is a property of the distribution.
Example
For a Bernoulli random variable with distribution Ber(p). The entropy H(p) = p log2 p (1 p) log2 (1 p)
Definition 1.2 Suppose we have random variables X, Y , then the Joint Entropy between them is
X
H(X, Y ) =
p(x, y) log2 p(x, y)
x,y
This is a measure of the total uncertainty of X, Y .

Note: If X, Y are independent, then it is easy to show that H(X, Y ) = H(X)+H(Y ). If X, Y are dependent,
then H(X, Y ) < H(X) + H(Y ) in general.
Definition 1.3 Suppose we have random variables X, Y , then the Conditional Entropy of Y conditioned
on X is defined as
X
H(Y | X) =
p(x)H(Y | X = x)
x
p(x)
XX
x
p(y | x) log2 p(y | x)
p(x, y) log2 p(y | x)
= EX,Y [log
1
]
p(y | x)
1-4
The conditional entropy is the average uncertainty in Y after we had observed X.

Theorem 1.4 (Chain Rule)
H(X, Y ) = H(Y | X) + H(X)
The proof of the chain rule will be covered in recitation on in the homework.
Definition 1.5 Given two distributions p, q for a random variable X. The Relative Entropy between p
and q is defined as
1
1
D(p || q) = Ep [log ] Ep [log ]
q
p
p
= Ep [log ]
q
X
p(x)
=
p(x) log
q(x)
x
The relative entropy is also known as Information divergence or KL (Kullback-Leibler) divergence.
The relative entropy is the cost incurred if we used distribution q to encode X when the true underlying
distribution is p. We will make this intuition more precise later in the course. We note that relative entropy
has some important properties:
D(p || q) 0, and equals 0 iff p = q
Relative entropy is often not symmetric, i.e., D(p || q) 6= D(q || p)
Definition 1.6 Let X, Y be two random variables. The Mutual Information between X and Y is the
following:
I(X, Y ) = D(p(x, y) || p(x)p(y))
where p(x, y) is the actual joint distribution of X, Y and p(x) and p(y) are the corresponding marginal
distributions. Thus, p(x)p(y) denotes the joint distribution that would result if X, Y were independent.
Note: we can give a preview of the answers of the two fundamental questions of information theory stated
before.
1. How much can we compress data? If the data is generated as a random variable X, then the answer
is H(X).
2. How much data can we transmit over a noisy channel? If X is the input and Y is the output, then the
answer is maxp(X) I(X, Y ).
We will prove these later.
1.4
1-5
Connection to Maximum Likelihood Estimation
Suppose X = (X1 , ..., Xn ) are data generated from a distribution p(X). In maximum likelihood estimation,
we want to find a distribution q in some family of distributions Q such that the likelihood q(X) is maximized:
max q(X) = min log q(X)
qQ
qQ
In machine learning, we often define a loss function. In this case, the loss function is the negative log loss:
1
]. We want
loss(q, X) = log q(X). The expected value of this loss function is the risk: Risk(q) = Ep [log q(x)
to find a distribution q that minimizes the risk. However, notice that minimizing the risk with respect to a
distribution q is exactly minimizing the relative entropy between p and q. This is because
Risk(q) = Ep [log
p(x)
1
1
] = Ep [log
] + Ep [log
] = D(p||q) + Risk(p)
q(x)
q(x)
p(x)
Because we know that relative entropy is always non-negative, we know here that the risk is minimized by
setting q equal to p. Thus the minimum risk R = Risk(p) = H(p), the entropy of distribution p. The
excess risk, Risk(q) R is precisely the relative entropy between p and q.

Lecture 1: Introduction, Entropy and ML Estimation

Uploaded by

Copyright:

Available Formats

Lecture 1: Introduction, Entropy and ML Estimation

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Lecture 1: Introduction, Entropy and ML Estimation

Uploaded by

Copyright:

Available Formats

10-704: Information Processing and Learning

Lecture 1: Introduction, Entropy and ML estimation

About the Class

Source Compressor Decompressor Receiver

Source Encoder Channel Decoder Receiver

Lecture 1: Introduction, Entropy and ML estimation

Connection to Machine Learning:

Information Content of Outcomes of Random Experiments

The information gained from missing on the first launch is log2 ( 64

Lecture 1: Introduction, Entropy and ML estimation

3. What if the questions are allowed to have more than 2 answers?

Information Content of Random Variables

A random variable is simply an assignment of probability to outcomes of a random experiment. We then

This is a measure of the total uncertainty of X, Y .

p(y | x) log2 p(y | x)

p(x, y) log2 p(y | x)

Lecture 1: Introduction, Entropy and ML estimation

The conditional entropy is the average uncertainty in Y after we had observed X.

Lecture 1: Introduction, Entropy and ML estimation

Connection to Maximum Likelihood Estimation

You might also like