Accepted Notation

xi
yj
wi,j
bj
sj

input signal of activation unit Xi .

yj = F (sj ) for an activation unit Yj .
Weight on connection from unit Xi to unit Yj .
Bias on unit Yj .
X
Net input to unit Yj : sj = bj +
xi wij

W
wj
uj
x
wij

weight matrix {wij }.

Vector of weights wj = (w1j , w2j , . . . , wnj )T .
Threshold for activation of neuron Yj .
Input vector for classification or response x = (x1 , . . . , xi , . . . , xn ).
Change in wij . wij = [wij (new) wij (old)].

Dr. E.C. Kulasekere ()

Neural Networks

1 / 36

Mathematical Models of Activation Functions

Typically the same activation function is used for all neurons in any
particular layer (this is not a requirement).
In a multi-layer network if the neurons have linear activation
functions the capabilities are no better than a single layer network
with a linear activation function.
Hence in most cases nonlinear activation functions are used.
Linear Activation Function
f (x) = x, x.

Dr. E.C. Kulasekere ()

Neural Networks

2 / 36

Mathematical Models of Activation Functions ...

Binary Step Function with Threshold

1 if x
f (x) =
0 if x <

f (x)

Dr. E.C. Kulasekere ()

Neural Networks

3 / 36

Mathematical Models of Activation Functions ...

Binary Sigmoid Function
f (x) =

1
.
1 + exp (x)

f (x) = f (x)[1 f (x)].

f (x)
1

This is especially useful when training the backpropagation neural

network where the derivatives are used to obtain the output value.
Dr. E.C. Kulasekere ()

Neural Networks

4 / 36

Mathematical Models of Activation Functions ...

Bipolar Sigmoid Function
g(x) = 2f (x) 1 =
0

g (x) =

1 exp (x)
.
1 + exp (x)

[1 + g(x)][1 g(x)].
2

The above is actually the binary sigmoid scaled to have a set of

values ranging from -1 to +1.
The bipolar sigmoid is also closely related to the hyperbolic
tangent function.
exp (x) exp (x)
exp (x) + exp (x)
1 exp (2x)
=
.
1 + exp (2x)

h(x) = tanh (x) =

Dr. E.C. Kulasekere ()

Neural Networks

5 / 36

Matrix Manipulations
1
X1

Xi

bj
w1j
wij

Yj

wnj
Xn

sj = wj xt =

Pn

i=1 xi wij ,

if the bias is neglected

If the bias is not neglected, the x includes the bias value.

Here row vectors for the input and weights are assumed.
Ex 2.2.1 is a direct application of the material learned thus far.
Dr. E.C. Kulasekere ()

Neural Networks

6 / 36

McCulloch-Pitts Neuron
Characteristics

The McCulloch-Pitt neuron is considered to be the first neural network.

It is a fixed weight network that can be used to implement boolean
functions.
Its characteristics are:
Binary activation (1 ON, 0 OFF). i.e. it either fires with an
activation 1 or does not fire with an activation of 0.
Neurons are connected by directed weighted paths.
If w > 0, excitatory, else inhibitory.
Excitatory weights are identical; inhibitory weights too.
Each neuron has a fixed threshold for firing. That is if the net input
to the neuron is greater than the threshold it fires.
The threshold is set such that the inhibition is absolute. Any
non-zero inhibitory input will prevent it from firing.
It takes one time step for the signal to pass over one link.
Dr. E.C. Kulasekere ()

Neural Networks

7 / 36

McCulloch-Pitts Neuron
Architecture

The inhibition should be

absolute. Hence the threshold
satisfies
> nw p

X1
w

Xn
Xn+1

w
p

The output neuron will fire if it

receives k or more excitatory
inputs with no inhibitory inputs,
where
kw > (k 1)w

p
Xn+m

The activation function

is

1 if s
f (s) =
0 if s <
Dr. E.C. Kulasekere ()

Neural Networks

8 / 36

Neural net to perform the AND function

X1

1
Y

X2

Write the truth table for the above network.

Threshold = 2
Given another configuration of the AND function with Threshold 1.

Dr. E.C. Kulasekere ()

Neural Networks

9 / 36

Neural net to perform the OR function

X1

2
Y

X2

Write the truth table for the above network.

Threshold=2
Give another configuration of the OR function with Threshold 1.

Dr. E.C. Kulasekere ()

Neural Networks

10 / 36

Neural net to perform the AND NOT function

X1

2
Y

X2

Write the truth table for the AND NOT function.

Threshold=2
Can you find another configuration for the weights to implement
the AND NOT function?

Dr. E.C. Kulasekere ()

Neural Networks

11 / 36

Neural net to perform the XOR function

X1

Z1

2
1
Y
1
2
X2

Z2

Write the truth table for the XOR function.

Threshold=2
Write the layer equations for the XOR function and show that it is
AND NOT combined with an OR function.
Dr. E.C. Kulasekere ()

Neural Networks

12 / 36

Geometric View of Neural Architectures

Single output neuron

Note that the threshold function can be reduced to the discrete

activation function when the threshold is considered as a bias.
Then the synaptic input can be written as s = wxt .
For two input neurons with inputs x1 and x2 , s = 0 is a straight line.
On the x1 x2 -plane, s > 0 gives one half-plane and s < 0 gives
another half-plane.
If the weights and threshold are multiplied by a constant c, then
the only difference is that the half-planes in which the earlier
classification occurred will be switched. (pp.27)
If you have more that two inputs you will be plotting on a
x1 x2 . . . xn -hyper plane. For example if n = 3 you will end up with a
plane.
Dr. E.C. Kulasekere ()

Neural Networks

13 / 36

Geometric View of Neural Architectures

Multiple output neurons

Each output neuron will have a summation block of sj = wxt .

For j = 2 two input neurons the above will give two straight lines in
the x1 , x2 -plane.
For j = 2 with more than two input neurons, the above will give
two n-dimensional hyper-planes.
For j > 2 and several inputs will give higher order planes.

Dr. E.C. Kulasekere ()

Neural Networks

14 / 36

Example 3.3.5
Note that the threshold function can be converted to the discrete
activation function by converting s > 0 to s u > 0. By doing this
we are converting a threshold to a bias.
We have to mark the area in which the output turns out to be 1.
That is by looking at the vertices, we determine it is
x1 > 0, x1 < 1, x2 > 0, and, x2 < 1.
In order to obtain the required output the threshold function has to
satisfy s u > 0. That is w1 x1 + w2 x2 u > 0. Now compare the
above two items to obtain the required w1 , w2 and u
Even for Exercise 3.3.2 we have a region bonded by a triangle in
which the same computation can be carried out.
Even when you are building a two layer network the arguments
used in MCP networks can be used for example the question in
pp.33.
All of these techniques produce fixed weight ANNs.
Dr. E.C. Kulasekere ()

Neural Networks

15 / 36

Pattern Classification Using ANNs

This is the simplest task a neural network can be trained to

perform.
Each input vector pattern either belongs or does not belong to a
class.
The correct input data is assumed to be known.
The activation function is either binary or bipolar.
The output of the trained network will be 1 when presented with a
pattern from the class.
If the pattern is not in the class a -1 (or 0 for a binary output) is
output.
Since it is either one set of outputs the activation is a hard step.

Dr. E.C. Kulasekere ()

Neural Networks

16 / 36

Biases and Thresholds

The bias acts exactly as a weight on a connection from a unit
whose activation is always 1.
Increasing the bias increases the net input to to the unit.
The Activation with a bias is

1 if s 0;
f (s) =
1 if s < 0;

where s = b +

The Activation with a threshold is


1 if s ;
f (s) =
1 if s < ;

xi wi,.

where s =

xi wi,.

Essentially the above are equivalent for most cases. However the
bias becomes essential for certain problems associated with linear
separability.
Dr. E.C. Kulasekere ()

Neural Networks

17 / 36

Requirement of Linear Separability

In pattern classification, the network is trained to output an
indicator to show the presented input is either a member of one
class or another class.
This will depend on the decision boundary that will either fire the
neuron or not fire it. This can be determined by setting y _in = 0
for the case with a bias.
The decision boundary is
X
b+
xi wi = 0.
i

If the system is linearly separable, the two classes should lie on

either side of the decision boundary (y _in > 0 or y _in < 0).
The boundary is non unique
x2 =
Dr. E.C. Kulasekere ()

w1
b
x1
.
w2
w2

Neural Networks

18 / 36

Requirement of Linear Separability ...

For bipolar signals the outputs for the two classes are -1 and +1.
For unipolar signals it is 0 and 1.
Depending on the number of inputs the decision boundary can be
a line, plane or a hyperplane. Eg. For two inputs its a line and for
three inputs its a plane.
If all of the training input vectors for which the correct response is
+1 lie on one side of the decision boundary we say that this
system is linearly separable.
It has been shown that a single layer network can only learn
linearly separable problems.
The trained weights are not unique.

Dr. E.C. Kulasekere ()

Neural Networks

19 / 36

Response Regions for the AND Function

The AND function for Bipolar inputs

and targets is
Input (x1 , x2 ) Output (t)
(1,1)
+1
(1,-1)
-1
(-1,1)
-1
(-1,-1)
-1
x2 = x1 + 1 is the decision
boundary where w1 = w2 = 1 and
b = 1.
This boundary is not unique.

Dr. E.C. Kulasekere ()

Neural Networks

x2

x1

20 / 36

Response Regions for the OR function

The OR function for Bipolar inputs

and targets is
Input (x1 , x2 ) Output (t)
(1,1)
+1
(1,-1)
+1
One
(-1,1)
+1
(-1,-1)
-1
possible decision boundary is
given by x2 = x1 1 where
b = 1, w1 = 1 = w2 .
Draw the above graphically.

Dr. E.C. Kulasekere ()

Importance of Bias

Neural Networks

If the bias weight was not

included in the previous
example the decision
boundary would have been
forced to go through the origin.
This changes the problem
from a solvable one to an
unsolvable problem.

21 / 36

Response Regions for the XOR Function

The XOR function for bipolar inputs

and targets is given by
Input (x1 , x2 ) Output (t)
(1,1)
-1
(1,-1)
+1
This
(-1,1)
+1
(-1,-1)
-1
is not linearly separable.

x2

x1

Dr. E.C. Kulasekere ()

Neural Networks

22 / 36

Binary/Bipolar Data Representation

In most cases the binary input data can be modified to bipolar data.
However the form of the data can change the problem from one that is
solvable to a problem that cannot be solved.
Binary representation is also not as good as the bipolar if we want the
net to generalize. i.e. to respond to input data that is similar but not
the same as the training data.
Bipolar inputs and targets are also better during training where the
adaptive weight change is computed. For example
w = xy
is the weight change in the Hebb learning algorithm. If either the
training input vector or the target vector is binary (unipolar) the update
becomes zero and the learning stops.
Using bipolar missing data can be distinguished from the mistaken
data by assigning 0 to missing data and a mistake by changing -1 to
+1 or vise versa.
Dr. E.C. Kulasekere ()

Neural Networks

23 / 36

The HEBB Net

This is the earliest and the most simplest learning rule for the
neural networks.
Hebb proposed that if two interconnected neurons are both on at
the same time, then the weight between them should be increased.
However the original statement did not discuss neurons that are
connected but do not fire together. Later this was also included in
the Hebb rule so that the weight increased of not firing together
and the original rule was made a strongest learning algorithm.
The weight update can be represented as
wi (new) = wi (old) + xi y .
Note that if the signals are binary, the update rule cannot
distinguish between a pair of inputs outputs for the following
conditions.
Input is on and the target is off.
Both the input and the target units are off.
Dr. E.C. Kulasekere ()

Neural Networks

24 / 36

Hebb Algorithm
Step 0 Initialize all weights: wi = 0, i = 1, . . . , n.
Step 1 For each input training vector and target output pair, s : t,
do steps 2-4.
Step 2 Set activations for input units: xi = si , i = 1, . . . , n.
Step 3 Set activation for output unit: y = t.
Step 4 Adjust the weight and bias:
wi (new) = wi (old) + xi y for i = 1, . . . , n
b(new) = b(old) + y .
If the bias is considered to be an input signal that is always 1, the
weight change can be written as
w(new) = w(old) + w
where w = xy .
Dr. E.C. Kulasekere ()

Neural Networks

25 / 36

Hebb Net for AND Function

Binary inputs and targets

Input
Target
(x1 , x2 , 1)
(1,1,1)
1
The truth table for the AND function is
(1,0,1)
0
(0,1,1)
0
(0,0,1)
0
The initial values are w(old) = (0, 0) and b(old) = 0.
The first step of the algorithm is
Input
Target Weight Changes Weights
(x1 , x2 , 1)
(w1 , w2 , b) (w1 , w2 , b)
(0,0,0)
(1,1,1)
1
(1,1,1)
(1,1,1)
The separating line after the first step is
x2 = x1 1
Dr. E.C. Kulasekere ()

Neural Networks

26 / 36

Hebb Net for AND Function ...

Binary inputs and targets

Now if we present the second, third and fourth training vectors the
weight change is given by
Input
Target Weight Changes Weights
(x1 , x2 , 1)
(w1 , w2 , b) (w1 , w2 , b)
(1,1,1)
(1,0,1)
0
(0,0,0)
(1,1,1)
(0,1,1)
0
(0,0,0)
(1,1,1)
(0,0,1)
0
(0,0,0)
(1,1,1)
We note that in the above problem when the target value is zero,
no learning occurs and hence no weight change.
With this we can determine that the bipolar inputs have resulted in
a short coming in the learning method.

Dr. E.C. Kulasekere ()

Neural Networks

27 / 36

Hebb Net for the AND function

Binary inputs, bipolar targets

Input
(x1 , x2 , 1)
(1,1,1)
The truth table in this case is
(1,0,1)
(0,1,1)
(0,0,1)
Presenting the first input
Input
Target Weight Changes
(x1 , x2 , 1)
(w1 , w2 , b)

Target
+1
-1
-1
-1

Weights
(w1 , w2 , b)
(0,0,0)
(1,1,1)
1
(1,1,1)
(1,1,1)
The separating line becomes x2 = x1 1. This is the correct
classification for the first input.
Dr. E.C. Kulasekere ()

Neural Networks

28 / 36

Hebb Net for the AND function ...

Binary inputs, bipolar targets
x2

Dr. E.C. Kulasekere ()

Neural Networks

x1

29 / 36

Hebb Net for the AND function ...

Binary inputs, bipolar targets

Presenting the other inputs we have the following

Input
Target Weight Changes Weights
(x1 , x2 , 1)
(w1 , w2 , b) (w1 , w2 , b)
(1,0,1)
-1
(-1,0,-1)
(0,1,0)
(0,1,1)
-1
(0,-1,-1)
(0,0,-1)
(0,0,1)
-1
(0,0,-1)
(0,0,-2)
Again we see that the problem has not correctly classified the
output. Hence alternate training data should be used.

Dr. E.C. Kulasekere ()

Neural Networks

30 / 36

Hebb Net for the AND function

Bipolar inputs and targets

The truth table is given by

Input
Target
(x1 , x2 , 1)
(1,1,1)
+1
(1,-1,1)
-1
(-1,1,1)
-1
(-1,-1,1)
-1
Presenting the first input:target
pair yields the same result as
before with the separation line
being x2 = x1 1.
Now the classification is
correct for the point (1,1) and
also for (-1,-1).
Dr. E.C. Kulasekere ()

Neural Networks

x2

x1

31 / 36

Hebb Net for the AND function ...

Bipolar inputs and targets

For the second input:target pair we obtain the separating line as

x2 = 0.
Input
Target Weight Changes Weights
(x1 , x2 , 1)
(w1 , w2 , b) (w1 , w2 , b)
(1,1,1)
(1,-1,1)
-1
(-1,1,-1)
(0,2,0)
With the third pair we obtain the separating line as x2 = x1 + 1.
Input
Target Weight Changes Weights
(x1 , x2 , 1)
(w1 , w2 , b) (w1 , w2 , b)
(0,2,0)
(-1,1,1)
-1
(1,-1,-1)
(1,1,-1)
Presenting the fourth pair, the separating line does not change from
what we obtained previously even though the weights have changed.
Input
Target Weight Changes Weights
(x1 , x2 , 1)
(w1 , w2 , b) (w1 , w2 , b)
(1,1,-1)
(-1,-1,1)
-1
(1,1,-1)
(2,2,-2)
Dr. E.C. Kulasekere ()

Neural Networks

32 / 36

Hebb Net for the AND function ...

Bipolar inputs and targets
x2

x1

Figure: Decision boundary for bipolar AND function using Hebb rule after
third/fourth training.

Draw the diagram for the intermediate stages as well.

Dr. E.C. Kulasekere ()

Neural Networks

33 / 36

Character Recognition Example Using Hebb Net

The Problem: Distinguish between the two patterns.

The steps involved are as follows:

We use the bipolar representation to convert the patterns into the
input values that can be used for training. We assign # to 1 and to
-1.
The correct target for the first pattern is +1 and for the second
pattern it is -1. This is used as the classification.
Then use the Hebb rule to find the weights by repeatedly
presenting the input:target pairs ti find the weights.

Check the system with training patterns that are similar but not
identical to see if the system will still react with the correct
classification.
Dr. E.C. Kulasekere ()

Neural Networks

34 / 36

Missing and Mistaken Data

Binary representation of inputs and targets (0 and 1 levels)

Missing Data: Cannot be represented.
Mistaken Data: A mistake of 1 is represented as 0 and vise versa.

Bipolar representation of inputs and targets (-1 and 1 levels)

Missing Data: this is represented by a 0.
Mistaken Data: A mistake in +1 is represented as -1 and vise versa.

In general a net can handle more missing components than wrong

components.
For input data the above translates to Its better not to guess.
The reason is that if the value is zero the weight change during
training may not get affected.

Dr. E.C. Kulasekere ()

Neural Networks

35 / 36

Additional Comments

ANN Training is completely ad hoc. Sometimes the result will not

converge.
The weight vector ca be multiplied by a positive number without
changing the actions of the neurons. This normalization can
eliminate rounding errors account for input variations etc.
The value of is significant for convergence as well as the rate of
convergence.
The order in which the training vectors are presented is also
important. Randomizing is generally preferable.
In multilayer networks, it is accepted that more than two layers are
rarely used.

Dr. E.C. Kulasekere ()

Neural Networks

36 / 36