Lecture 87

Artificial neural networks:

Supervised learning
I Introduction, or how the brain works
I The neuron as a simple computing element
I The perceptron
I Multilayer neural networks
I Accelerated learning in multilayer neural networks
I The Hopfield network
I Bidirectional associative memories (BAM)
I Summary
 Negnevitsky, Pearson Education, 2011 1
 Introduction, or how the brain works
Machine learning involves adaptive mechanisms
that enable computers to learn from experience,
learn by example and learn by analogy. Learning
capabilities can improve the performance of an
intelligent system over time. The most popular
approaches to machine learning are artificial
neural networks and genetic algorithms. This
lecture is dedicated to neural networks.

 Negnevitsky, Pearson Education, 2011 2

 I A neural network can be defined as a model of
reasoning based on the human brain. The brain
consists of a densely interconnected set of nerve
cells, or basic information-processing units, called
neurons.
I The human brain incorporates nearly 10 billion
neurons and 60 trillion connections, synapses,
between them. By using multiple neurons
simultaneously, the brain can perform its functions
much faster than the fastest computers in existence
today.

 Negnevitsky, Pearson Education, 2011 3

 I Each neuron has a very simple structure, but an
army of such elements constitutes a tremendous
processing power.
I A neuron consists of a cell body, soma, a number of
fibers called dendrites, and a single long fiber
called the axon.

 Negnevitsky, Pearson Education, 2011 4

 Biological neural network
Synapse

Synapse Dendrites
Axon
Axon

Soma Soma
Dendrites
Synapse

 Negnevitsky, Pearson Education, 2011 5

 I Our brain can be considered as a highly complex,
non-linear and parallel information-processing
system.
I Information is stored and processed in a neural
network simultaneously throughout the whole
network, rather than at specific locations. In other
words, in neural networks, both data and its
processing are global rather than local.
I Learning is a fundamental and essential
characteristic of biological neural networks. The
ease with which they can learn led to attempts to
emulate a biological neural network in a computer.
 Negnevitsky, Pearson Education, 2011 6
 I An artificial neural network consists of a number of
very simple processors, also called neurons, which
are analogous to the biological neurons in the brain.
I The neurons are connected by weighted links
passing signals from one neuron to another.
I The output signal is transmitted through the
neuron’s outgoing connection. The outgoing
connection splits into a number of branches that
transmit the same signal. The outgoing branches
terminate at the incoming connections of other
neurons in the network.

 Negnevitsky, Pearson Education, 2011 7

 Architecture of a typical artificial neural network

Out put Signals

Input Signals

Middle Layer
Input Layer Output Layer

 Negnevitsky, Pearson Education, 2011 8

 Analogy between biological and
artificial neural networks

Biological Neural Network Artificial Neural Network

Soma Neuron
Dendrite Input
Axon Output
Synapse Weight

 Negnevitsky, Pearson Education, 2011 9

 The neuron as a simple computing element
Diagram of a neuron

Input Signals Weights Output Signals

x1
Y
w1
x2
w2
Neuron Y Y

wn
Y
xn

 Negnevitsky, Pearson Education, 2011 10

 I The neuron computes the weighted sum of the input
signals and compares the result with a threshold
value, θ. If the net input is less than the threshold,
the neuron output is –1. But if the net input is greater
than or equal to the threshold, the neuron becomes
activated and its output attains a value +1.
I The neuron uses the following transfer or activation
function:
n + 1, if X ≥ θ
X = ∑ xi wi Y =
i =1 − 1, if X < θ
I This type of activation function is called a sign
function.
 Negnevitsky, Pearson Education, 2011 11
 Activation functions of a neuron

Step function Sign function Sigmoid function Linear function

Y Y Y Y
+1 +1 +1 +1

0 X 0 X 0 X 0 X
-1 -1 -1 -1

step 1, if X ≥ 0 sign +1, if X ≥ 0 sigmoid 1

Y = Y = Y = Y linear= X
0, if X < 0 −1, if X < 0 1+ e− X

 Negnevitsky, Pearson Education, 2011 12

 Can a single neuron learn a task?

I In 1958, Frank Rosenblatt introduced a training

algorithm that provided the first procedure for
training a simple ANN: a perceptron.
I The perceptron is the simplest form of a neural
network. It consists of a single neuron with
adjustable synaptic weights and a hard limiter.

 Negnevitsky, Pearson Education, 2011 13

 Single-layer two-input perceptron

Inputs
x1 Linear Hard
w1 Combiner Limiter
Output
∑ Y
w2
θ
x2
Threshold

 Negnevitsky, Pearson Education, 2011 14

 The Perceptron
I The operation of Rosenblatt’s perceptron is based
on the McCulloch and Pitts neuron model. The
model consists of a linear combiner followed by a
hard limiter.
I The weighted sum of the inputs is applied to the
hard limiter, which produces an output equal to +1
if its input is positive and −1 if it is negative.

 Negnevitsky, Pearson Education, 2011 15

 I The aim of the perceptron is to classify inputs,
x1, x2, . . ., xn, into one of two classes, say
A1 and A2.
I In the case of an elementary perceptron, the n-
dimensional space is divided by a hyperplane into
two decision regions. The hyperplane is defined by
the linearly separable function:
n
∑ xi wi − θ = 0
i =1

 Negnevitsky, Pearson Education, 2011 16

 Linear separability in the perceptrons
x2 x2

Class A 1

1
2
1
x1
Class A 2 x1

x1w 1 + x2w 2 − θ = 0 x1 w 1 + x 2 w 2 + x 3 w 3 − θ = 0
x3
(a) Two-input perceptron. (b) Three-input perceptron.

 Negnevitsky, Pearson Education, 2011 17

 How does the perceptron learn its classification
tasks?
This is done by making small adjustments in the
weights to reduce the difference between the actual
and desired outputs of the perceptron. The initial
weights are randomly assigned, usually in the range
[−0.5, 0.5], and then updated to obtain the output
consistent with the training examples.

 Negnevitsky, Pearson Education, 2011 18

 I If at iteration p, the actual output is Y(p) and the
desired output is Yd (p), then the error is given by:

e( p) = Yd ( p) −Y ( p) where p = 1, 2, 3, . . .

Iteration p here refers to the pth training example

presented to the perceptron.
I If the error, e(p), is positive, we need to increase
perceptron output Y(p), but if it is negative, we
need to decrease Y(p).

 Negnevitsky, Pearson Education, 2011 19

 The perceptron learning rule
wi ( p + 1) = wi ( p ) + α ⋅ xi ( p ) ⋅ e( p)
where p = 1, 2, 3, . . .
α is the learning rate, a positive constant less than
unity.

The perceptron learning rule was first proposed by

Rosenblatt in 1960. Using this rule we can derive
the perceptron training algorithm for classification
tasks.

 Negnevitsky, Pearson Education, 2011 20

 Perceptron’s training algorithm
Step 1: Initialisation
Set initial weights w1, w2,…, wn and threshold θ
to random numbers in the range [−0.5, 0.5].
Step 2: Activation
Activate the perceptron by applying inputs x1(p),
x2(p),…, xn(p) and desired output Yd (p).
Calculate the actual output at iteration p = 1
 n 
Y ( p ) = step  ∑ x i ( p ) w i ( p ) − θ 
 i = 1 
where n is the number of the perceptron inputs,
and step is a step activation function.
 Negnevitsky, Pearson Education, 2011 21
 Perceptron’s training algorithm (continued)
Step 3: Weight training
Update the weights of the perceptron
wi ( p + 1) = wi ( p ) + ∆wi ( p )
where ∆wi(p) is the weight correction at iteration p.
The weight correction is computed by the delta
rule:
∆wi ( p ) = α ⋅ xi ( p) ⋅ e( p)
Step 4: Iteration
Increase iteration p by one, go back to Step 2 and
repeat the process until convergence.
 Negnevitsky, Pearson Education, 2011 22
Example of perceptron learning: the logical operation AND
Inputs Desired Initial Actual Error Final
Epoch output weights output weights
x1 x2 Yd w1 w2 Y e w1 w2
1 0 0 0 0.3 −0.1 0 0 0.3 −0.1
0 1 0 0.3 −0.1 0 0 0.3 −0.1
1 0 0 0.3 −0.1 1 −1 0.2 −0.1
1 1 1 0.2 −0.1 0 1 0.3 0.0
2 0 0 0 0.3 0.0 0 0 0.3 0.0
0 1 0 0.3 0.0 0 0 0.3 0.0
1 0 0 0.3 0.0 1 −1 0.2 0.0
1 1 1 0.2 0.0 1 0 0.2 0.0
3 0 0 0 0.2 0.0 0 0 0.2 0.0
0 1 0 0.2 0.0 0 0 0.2 0.0
1 0 0 0.2 0.0 1 −1 0.1 0.0
1 1 1 0.1 0.0 0 1 0.2 0.1
4 0 0 0 0.2 0.1 0 0 0.2 0.1
0 1 0 0.2 0.1 0 0 0.2 0.1
1 0 0 0.2 0.1 1 −1 0.1 0.1
1 1 1 0.1 0.1 1 0 0.1 0.1
5 0 0 0 0.1 0.1 0 0 0.1 0.1
0 1 0 0.1 0.1 0 0 0.1 0.1
1 0 0 0.1 0.1 0 0 0.1 0.1
1 1 1 0.1 0.1 1 0 0.1 0.1
Threshold: θ = 0.2; learning rate: α = 0.1
 Negnevitsky, Pearson Education, 2011 23
 Two-dimensional plots of basic logical operations
x2 x2 x2

1 1 1

x1 x1 x1
0 1 0 1 0 1

(a) AND (x 1 ∩ x 2 ) (b) OR (x 1 ∪ x 2 ) (c) Exclusive-OR

(x 1 ⊕ x 2 )

A perceptron can learn the operations AND and OR,

but not Exclusive-OR.

 Negnevitsky, Pearson Education, 2011 24

 Multilayer neural networks
I A multilayer perceptron is a feedforward neural
network with one or more hidden layers.
I The network consists of an input layer of source
neurons, at least one middle or hidden layer of
computational neurons, and an output layer of
computational neurons.
I The input signals are propagated in a forward
direction on a layer-by-layer basis.

 Negnevitsky, Pearson Education, 2011 25

 Multilayer perceptron with two hidden layers

Out put Signals

Input Signals

First Second
Input hidden hidden Output
layer layer layer layer

 Negnevitsky, Pearson Education, 2011 26

 What does the middle layer hide?
I A hidden layer “hides” its desired output.
Neurons in the hidden layer cannot be observed
through the input/output behaviour of the network.
There is no obvious way to know what the desired
output of the hidden layer should be.
I Commercial ANNs incorporate three and
sometimes four layers, including one or two
hidden layers. Each layer can contain from 10 to
1000 neurons. Experimental neural networks may
have five or even six layers, including three or
four hidden layers, and utilise millions of neurons.

 Negnevitsky, Pearson Education, 2011 27

 Back-propagation neural network
I Learning in a multilayer network proceeds the
same way as for a perceptron.
I A training set of input patterns is presented to the
network.
I The network computes its output pattern, and if
there is an error − or in other words a difference
between actual and desired output patterns − the
weights are adjusted to reduce this error.

 Negnevitsky, Pearson Education, 2011 28

 I In a back-propagation neural network, the learning
algorithm has two phases.
I First, a training input pattern is presented to the
network input layer. The network propagates the
input pattern from layer to layer until the output
pattern is generated by the output layer.
I If this pattern is different from the desired output,
an error is calculated and then propagated
backwards through the network from the output
layer to the input layer. The weights are modified
as the error is propagated.

 Negnevitsky, Pearson Education, 2011 29

 Three-layer back-propagation neural network
Input signals
1
x1 1 y1
1
2
x2 2 y2
2

i wij j wjk
xi k yk

m
n l yl
xn
Input Hidden Output
layer layer layer

Error signals

 Negnevitsky, Pearson Education, 2011 30

 The back-propagation training algorithm
Step 1: Initialisation
Set all the weights and threshold levels of the
network to random numbers uniformly
distributed inside a small range:
 2.4 2.4 
 − , + 
 Fi Fi 

where Fi is the total number of inputs of neuron i

in the network. The weight initialisation is done
on a neuron-by-neuron basis.

 Negnevitsky, Pearson Education, 2011 31

 Step 2: Activation
Activate the back-propagation neural network by
applying inputs x1(p), x2(p),…, xn(p) and desired
outputs yd,1(p), yd,2(p),…, yd,n(p).
(a) Calculate the actual outputs of the neurons in
the hidden layer:
 n 
y j ( p ) = sigmoid  ∑ x i ( p ) ⋅ wij ( p ) − θ j 
 i =1 
where n is the number of inputs of neuron j in the
hidden layer, and sigmoid is the sigmoid activation
function.
 Negnevitsky, Pearson Education, 2011 32
 Step 2: Activation (continued)
(b) Calculate the actual outputs of the neurons in
the output layer:
m 
y k ( p ) = sigmoid  ∑ x jk ( p ) ⋅ w jk ( p ) − θ k 
 j =1 
where m is the number of inputs of neuron k in the
output layer.

 Negnevitsky, Pearson Education, 2011 33

 Step 3: Weight training
Update the weights in the back-propagation network
propagating backward the errors associated with
output neurons.
(a) Calculate the error gradient for the neurons in the
output layer:
δ k ( p) = yk ( p ) ⋅ [1 − yk ( p )]⋅ ek ( p)
where ek ( p ) = y d , k ( p ) − y k ( p )

Calculate the weight corrections:

∆w jk ( p ) = α ⋅ y j ( p ) ⋅ δ k ( p )
Update the weights at the output neurons:
w jk ( p + 1) = w jk ( p ) + ∆w jk ( p )
 Negnevitsky, Pearson Education, 2011 34
 Step 3: Weight training (continued)
(b) Calculate the error gradient for the neurons in
the hidden layer:
l
δ j ( p ) = y j ( p ) ⋅ [1 − y j ( p )] ⋅ ∑ δ k ( p ) w jk ( p )
k =1
Calculate the weight corrections:
∆wij ( p ) = α ⋅ xi ( p ) ⋅ δ j ( p )
Update the weights at the hidden neurons:
wij ( p + 1) = wij ( p ) + ∆wij ( p )

 Negnevitsky, Pearson Education, 2011 35

 Step 4: Iteration
Increase iteration p by one, go back to Step 2 and
repeat the process until the selected error criterion
is satisfied.

As an example, we may consider the three-layer

back-propagation network. Suppose that the
network is required to perform logical operation
Exclusive-OR. Recall that a single-layer perceptron
could not do this operation. Now we will apply the
three-layer net.

 Negnevitsky, Pearson Education, 2011 36

 Three-layer network for solving the
Exclusive-OR operation
−1

θ3
w13 −1
x1 1 3 w35
w23 θ5

5 y5
w24
x2 2 4 w45
w24
Input θ4 Output
layer layer
−1
Hidden layer
 Negnevitsky, Pearson Education, 2011 37
 I The effect of the threshold applied to a neuron in the
hidden or output layer is represented by its weight, θ,
connected to a fixed input equal to −1.
I The initial weights and threshold levels are set
randomly as follows:
w13 = 0.5, w14 = 0.9, w23 = 0.4, w24 = 1.0, w35 = −1.2,
w45 = 1.1, θ3 = 0.8, θ4 = −0.1 and θ5 = 0.3.

 Negnevitsky, Pearson Education, 2011 38

 I We consider a training set where inputs x1 and x2 are
equal to 1 and desired output yd,5 is 0. The actual
outputs of neurons 3 and 4 in the hidden layer are
calculated as
[
y3 = sigmoid ( x1w13 + x2 w23 − θ3 ) = 1 / 1 + e − (1⋅0.5+1⋅0.4 −1⋅0.8) = 0.5250 ]
y4 = sigmoid ( x1w14 + x2 w24 − θ ) = 1 / [1 + e
4
− (1⋅0.9 +1⋅1.0 +1⋅0.1)
] = 0.8808
I Now the actual output of neuron 5 in the output layer
is determined as:
[ ]
y5 = sigmoid( y3w35 + y4w45 − θ5 ) = 1/ 1+ e−(−0.5250⋅1.2+0.8808⋅1.1−1⋅0.3) = 0.5097

I Thus, the following error is obtained:

e = yd ,5 − y5 = 0 − 0.5097 = −0.5097
 Negnevitsky, Pearson Education, 2011 39
 I The next step is weight training. To update the
weights and threshold levels in our network, we
propagate the error, e, from the output layer
backward to the input layer.
I First, we calculate the error gradient for neuron 5 in
the output layer:
δ 5 = y5 (1 − y5 ) e = 0.5097 ⋅ (1 − 0.5097) ⋅ (−0.5097) = −0.1274

I Then we determine the weight corrections assuming

that the learning rate parameter, α, is equal to 0.1:
∆w35 = α ⋅ y3 ⋅ δ 5 = 0.1⋅ 0.5250 ⋅ (−0.1274) = −0.0067
∆w45 = α ⋅ y4 ⋅ δ 5 = 0.1 ⋅ 0.8808 ⋅ (−0.1274) = −0.0112
∆θ5 = α ⋅ (−1) ⋅ δ 5 = 0.1 ⋅ (−1) ⋅ (−0.1274) = −0.0127
 Negnevitsky, Pearson Education, 2011 40
 I Next we calculate the error gradients for neurons 3
and 4 in the hidden layer:
δ 3 = y3 (1 − y3 ) ⋅ δ 5 ⋅ w35 = 0.5250 ⋅ (1 − 0.5250) ⋅ ( − 0.1274) ⋅ ( − 1.2) = 0.0381
δ 4 = y4 (1 − y4 ) ⋅ δ 5 ⋅ w45 = 0.8808 ⋅ (1 − 0.8808) ⋅ ( − 0.127 4) ⋅ 1.1 = −0.0147

I We then determine the weight corrections:

∆w13 = α ⋅ x1 ⋅ δ 3 = 0.1 ⋅1 ⋅ 0.0381 = 0.0038
∆w23 = α ⋅ x2 ⋅ δ 3 = 0.1 ⋅1 ⋅ 0.0381 = 0.0038
∆θ3 = α ⋅ ( −1) ⋅ δ 3 = 0.1 ⋅ ( −1) ⋅ 0.0381 = −0.0038
∆w14 = α ⋅ x1 ⋅ δ 4 = 0.1 ⋅1 ⋅ ( −0.0147 ) = −0.0015
∆w24 = α ⋅ x2 ⋅ δ 4 = 0.1 ⋅1 ⋅ (−0.0147 ) = −0.0015
∆θ 4 = α ⋅ ( −1) ⋅ δ 4 = 0.1 ⋅ ( −1) ⋅ ( −0.0147 ) = 0.0015

 Negnevitsky, Pearson Education, 2011 41

 I At last, we update all weights and threshold:
w13 = w13 + ∆ w13 = 0 .5 + 0 .0038 = 0 .5038
w14 = w14 + ∆ w14 = 0 .9 − 0 .0015 = 0 .8985
w 23 = w 23 + ∆ w 23 = 0 .4 + 0 .0038 = 0 .4038
w 24 = w 24 + ∆ w 24 = 1 .0 − 0 .0015 = 0 .9985
w35 = w35 + ∆ w35 = − 1 .2 − 0 .0067 = − 1 .2067
w 45 = w 45 + ∆ w 45 = 1 .1 − 0 .0112 = 1 .0888
θ 3 = θ 3 + ∆ θ 3 = 0 .8 − 0 .0038 = 0 .7962
θ 4 = θ 4 + ∆ θ 4 = − 0 .1 + 0 .0015 = − 0 .0985

θ 5 = θ 5 + ∆ θ 5 = 0 .3 + 0 .0127 = 0 .3127

I The training process is repeated until the sum of

squared errors is less than 0.001.
 Negnevitsky, Pearson Education, 2011 42
 Learning curve for operation Exclusive-OR
Sum-Squared Network Error for 224 Epochs
1
10

0
10
Sum-Squared Error

-1
10

-2
10

-3
10

-4
10
0 50 100 150 200
Epoch

 Negnevitsky, Pearson Education, 2011 43

 Final results of three-layer network learning

Inputs Desired Actual Error Sum of

output output squared
x1 x2 yd y5 e errors
1 1 0 Y
0.0155 −0.0155 0.0010
0 1 1 0.9849 0.0151
1 0 1 0.9849 0.0151
0 0 0 0.0175 −0.0175

 Negnevitsky, Pearson Education, 2011 44

 Network represented by McCulloch-Pitts model
for solving the Exclusive-OR operation
−1

+1.5
−1
+1.0
x1 1 3 −2.0 +0.5
+1.0

5 y5

+1.0
x2 2 +1.0
4
+1.0
+0.5

−1
 Negnevitsky, Pearson Education, 2011 45
 Decision boundaries
x2 x2 x2
x 1 + x 2 – 1.5 = 0 x 1 + x 2 – 0.5 = 0

1 1 1

x1 x1 x1
0 1 0 1 0 1

(a) (b) (c)

(a) Decision boundary constructed by hidden neuron 3;

(b) Decision boundary constructed by hidden neuron 4;
(c) Decision boundaries constructed by the complete
three-layer network
 Negnevitsky, Pearson Education, 2011 46
 Accelerated learning in multilayer
neural networks
I A multilayer network learns much faster when the
sigmoidal activation function is represented by a
hyperbolic tangent:

tan h 2a
Y = −bX
−a
1+ e
where a and b are constants.
Suitable values for a and b are:
a = 1.716 and b = 0.667
 Negnevitsky, Pearson Education, 2011 47
 I We also can accelerate training by including a
momentum term in the delta rule:

∆w jk ( p ) = β ⋅ ∆w jk ( p − 1) + α ⋅ y j ( p ) ⋅ δ k ( p )

where β is a positive number (0 ≤ β < 1) called the

momentum constant. Typically, the momentum
constant is set to 0.95.

This equation is called the generalised delta rule.

 Negnevitsky, Pearson Education, 2011 48

Learning with momentum for operation Exclusive-OR
Training for 126 Epochs
2
10
Sum-Squared Error
101
100
10-1
10-2
10-3
10-4
0 20 40 60 80 100 120
Epoch
1.5

1
Learning Rate

0.5

-0.5

-1
0 20 40 60 80 100 120 140
Epoch

 Negnevitsky, Pearson Education, 2011 49

 Learning with adaptive learning rate
To accelerate the convergence and yet avoid the
danger of instability, we can apply two heuristics:
Heuristic 1
If the change of the sum of squared errors has the same
algebraic sign for several consequent epochs, then the
learning rate parameter, α, should be increased.
Heuristic 2
If the algebraic sign of the change of the sum of
squared errors alternates for several consequent
epochs, then the learning rate parameter, α, should be
decreased.
 Negnevitsky, Pearson Education, 2011 50
 I Adapting the learning rate requires some changes
in the back-propagation algorithm.
I If the sum of squared errors at the current epoch
exceeds the previous value by more than a
predefined ratio (typically 1.04), the learning rate
parameter is decreased (typically by multiplying
by 0.7) and new weights and thresholds are
calculated.
I If the error is less than the previous one, the
learning rate is increased (typically by multiplying
by 1.05).

 Negnevitsky, Pearson Education, 2011 51

 Learning with adaptive learning rate
Training for 103 Epochs
2
10
Sum-Squared Error
101
100
10-1
10-2
10-3
10-4
0 10 20 30 40 50 60 70 80 90 100
Epoch
1

0.8
Learning Rate

0.6

0.4

0.2

0
0 20 40 60 80 100 120
Epoch

 Negnevitsky, Pearson Education, 2011 52

Learning with momentum and adaptive learning rate
Training for 85 Epochs
2
Sum-Squared Error 10
101
100
10-1
10-2
10-3
10-4
0 10 20 30 40 50 60 70 80
Epoch
2.5

2
Learning Rate

1.5

0.5

0
0 10 20 30 40 50 60 70 80 90
Epoch

 Negnevitsky, Pearson Education, 2011 53

 The Hopfield Network
I Neural networks were designed on analogy with
the brain. The brain’s memory, however, works
by association. For example, we can recognise a
familiar face even in an unfamiliar environment
within 100-200 ms. We can also recall a
complete sensory experience, including sounds
and scenes, when we hear only a few bars of
music. The brain routinely associates one thing
with another.

 Negnevitsky, Pearson Education, 2011 54

 I Multilayer neural networks trained with the back-
propagation algorithm are used for pattern
recognition problems. However, to emulate the
human memory’s associative characteristics we
need a different type of network: a recurrent
neural network.
I A recurrent neural network has feedback loops
from its outputs to its inputs. The presence of
such loops has a profound impact on the learning
capability of the network.

 Negnevitsky, Pearson Education, 2011 55

 I The stability of recurrent networks intrigued
several researchers in the 1960s and 1970s.
However, none was able to predict which network
would be stable, and some researchers were
pessimistic about finding a solution at all. The
problem was solved only in 1982, when John
Hopfield formulated the physical principle of
storing information in a dynamically stable
network.

 Negnevitsky, Pearson Education, 2011 56

 Single-layer n-neuron Hopfield network
x1 1 y1

Output Signals
Input Signals

x2 2 y2

xi i yi

xn n yn

 Negnevitsky, Pearson Education, 2011 57

 I The Hopfield network uses McCulloch and Pitts
neurons with the sign activation function as its
computing element:

+1, if X > 0
sign 
Y = −1, if X < 0
 Y , if X = 0


 Negnevitsky, Pearson Education, 2011 58

 I The current state of the Hopfield network is
determined by the current outputs of all neurons,
y1, y2, . . ., yn.
Thus, for a single-layer n-neuron network, the state
can be defined by the state vector as:

 y1 
 y 
Y =  2 
 M 
 
 y n 

 Negnevitsky, Pearson Education, 2011 59

 I In the Hopfield network, synaptic weights between
neurons are usually represented in matrix form as
follows:
M
W= ∑ m m −M I
Y YT

m=1

where M is the number of states to be memorised

by the network, Ym is the n-dimensional binary
vector, I is n × n identity matrix, and superscript T
denotes a matrix transposition.

 Negnevitsky, Pearson Education, 2011 60

 Possible states for the three-neuron
Hopfield network
y2

(−1, 1, −1) (1, 1, −1)

(−1, 1, 1) (1, 1, 1)
y1
0

(−1, −1, −1) (1, −1, −1)

(−1, −1, 1) (1, −1, 1)

y3

 Negnevitsky, Pearson Education, 2011 61

I The stable state-vertex is determined by the weight
matrix W, the current input vector X, and the
threshold matrix θ. If the input vector is partially
incorrect or incomplete, the initial state will converge
into the stable state-vertex after a few iterations.
I Suppose, for instance, that our network is required to
memorise two opposite states, (1, 1, 1) and (−1, −1, −1).
Thus,
1 −1
Y1 = 1 Y2 = −1 or 1 = [1 1 1] Y2T = [− 1 − 1 − 1]
Y T

1 −1

where Y1 and Y2 are the three-dimensional vectors.

 Negnevitsky, Pearson Education, 2011 62
I The 3 × 3 identity matrix I is
1 0 0
I = 0 1 0
0 0 1

I Thus, we can now determine the weight matrix as

follows:
1 −1 1 0 0 0 2 2
W = 1[1 1 1] + −1[−1 −1 −1] − 2 0 1 0 = 2 0 2
1 −1 0 0 1 2 2 0

I Next, the network is tested by the sequence of input

vectors, X1 and X2, which are equal to the output (or
target) vectors Y1 and Y2, respectively.
 Negnevitsky, Pearson Education, 2011 63
I First, we activate the Hopfield network by applying
the input vector X. Then, we calculate the actual
output vector Y, and finally, we compare the result
with the initial input vector X.

 0 2 2 1 0  1

       
Y1 = sign 2 0 2 1 − 0  = 1
2 2 0 1 0  1


0 2 2 −1 0 −1

        
Y2 = sign 2 0 2 −1 − 0 = −1
2 2 0 −1 0 −1


 Negnevitsky, Pearson Education, 2011 64

I The remaining six states are all unstable. However,
stable states (also called fundamental memories) are
capable of attracting states that are close to them.
I The fundamental memory (1, 1, 1) attracts unstable
states (−1, 1, 1), (1, −1, 1) and (1, 1, −1). Each of
these unstable states represents a single error,
compared to the fundamental memory (1, 1, 1).
I The fundamental memory (−1, −1, −1) attracts
unstable states (−1, −1, 1), (−1, 1, −1) and (1, −1, −1).
I Thus, the Hopfield network can act as an error
correction network.

 Negnevitsky, Pearson Education, 2011 65

 Storage capacity of the Hopfield network
I Storage capacity is or the largest number of
fundamental memories that can be stored and
retrieved correctly.
I The maximum number of fundamental memories
Mmax that can be stored in the n-neuron recurrent
network is limited by
M max = 0.15 n

 Negnevitsky, Pearson Education, 2011 66

 Bidirectional associative memory (BAM)
I The Hopfield network represents an autoassociative
type of memory − it can retrieve a corrupted or
incomplete memory but cannot associate this memory
with another different memory.
I Human memory is essentially associative. One thing
may remind us of another, and that of another, and so
on. We use a chain of mental associations to recover
a lost memory. If we forget where we left an
umbrella, we try to recall where we last had it, what
we were doing, and who we were talking to. We
attempt to establish a chain of associations, and
thereby to restore a lost memory.
 Negnevitsky, Pearson Education, 2011 67
 I To associate one memory with another, we need a
recurrent neural network capable of accepting an
input pattern on one set of neurons and producing
a related, but different, output pattern on another
set of neurons.
I Bidirectional associative memory (BAM), first
proposed by Bart Kosko, is a heteroassociative
network. It associates patterns from one set, set A,
to patterns from another set, set B, and vice versa.
Like a Hopfield network, the BAM can generalise
and also produce correct outputs despite corrupted
or incomplete inputs.

 Negnevitsky, Pearson Education, 2011 68

 BAM operation

x1(p) 1 x1(p+1) 1

1 y1(p) 1 y1(p)
x2(p) 2 x2(p+1) 2
2 y2(p) 2 y2(p)

xi(p)
j yj(p) j yj(p)
i xi(p+1) i

m ym(p) m ym(p)
xn(p) n xn(p+1) n
Input Output Input Output
layer layer layer layer

(a) Forward direction. (b) Backward direction.

 Negnevitsky, Pearson Education, 2011 69

 The basic idea behind the BAM is to store
pattern pairs so that when n-dimensional vector
X from set A is presented as input, the BAM
recalls m-dimensional vector Y from set B, but
when Y is presented as input, the BAM recalls X.

 Negnevitsky, Pearson Education, 2011 70

 I To develop the BAM, we need to create a
correlation matrix for each pattern pair we want to
store. The correlation matrix is the matrix product
of the input vector X, and the transpose of the
output vector YT. The BAM weight matrix is the
sum of all correlation matrices, that is,
M
W= ∑ T
X m Ym
m=1
where M is the number of pattern pairs to be stored
in the BAM.

 Negnevitsky, Pearson Education, 2011 71

 Stability and storage capacity of the BAM
I The BAM is unconditionally stable. This means that
any set of associations can be learned without risk of
instability.
I The maximum number of associations to be stored
in the BAM should not exceed the number of
neurons in the smaller layer.
I The more serious problem with the BAM is
incorrect convergence. The BAM may not
always produce the closest association. In fact, a
stable association may be only slightly related to
the initial input vector.
 Negnevitsky, Pearson Education, 2011 72