Unsupervised learning

•Introduction
•Winner–take–all learning
•Max net
•Hamming Net
•counter propagation network
•feature mapping
•Self-organizing feature map
•Applications of neural algorithms
•Neural Network control

1
 Introduction
The main property of a neural network is an ability to learn from its
environment, and to improve its performance through learning. So far we
have considered supervised or active learning learning with an external
“teacher” or a supervisor who presents a training set to the network. But
another type of learning also exists: unsupervised learning.
Learning without a teacher.
No feedback to indicate the desired outputs.
The network must by itself discover the relationship of interest
from the input data.
– E.g., patterns, features, regularities, correlations, or
categories.
Translate the discovered relationship into output

2
• Unsupervised learning
– Training samples contain only input patterns
• No desired output is given (teacher-less)
– Learn to form classes/clusters of sample patterns according to
similarities among them
• Patterns in a cluster would have similar features
• No prior knowledge as what features are important for
classification, and how many classes are there.
• In contrast to supervised learning, unsupervised or self-
organised learning does not require an external teacher.
During the training session, the neural network receives a
number of different input patterns, discovers significant
features in these patterns and learns how to classify input
data into appropriate categories. Unsupervised learning
tends to follow the neuro-biological organisation of the brain.
• Unsupervised learning algorithms aim to learn rapidly and
can be used in real-time 3
 Hebbian learning
In 1949, Donald Hebb proposed one of the key
ideas in biological learning, commonly known as
Hebb’s Law. Hebb’s Law states that if neuron i is
near enough to excite neuron j and repeatedly
participates in its activation, the synaptic connection
between these two neurons is strengthened and
neuron j becomes more sensitive to stimuli from
neuron i.

4
 Hebb’s Law can be represented in the form of two
rules:
1. If two neurons on either side of a connection
are activated synchronously, then the weight of
that connection is increased.
2. If two neurons on either side of a connection
are activated asynchronously, then the weight
of that connection is decreased.
Hebb’s Law provides the basis for learning
without a teacher. Learning here is a local
phenomenon occurring without feedback from
the environment.
5
 Hebbian learning in a neural network

Output Signals
Input Signals

i j

6
 Using Hebb’s Law we can express the adjustment
applied to the weight wij at iteration p in the
following form:
 w ij ( p )  F [ y j ( p ), x i ( p ) ]

 As a special case, we can represent Hebb’s Law as

follows:
wij ( p)  a y j ( p) xi ( p )

where  is the learning rate parameter.

This equation is referred to as the activity
product rule.
7
 Hebbian learning implies that weights can only
increase. To resolve this problem, we might
impose a limit on the growth of synaptic weights.
It can be done by introducing a non-linear
forgetting factor into Hebb’s Law:
wij ( p )   y j ( p ) xi ( p )   y j ( p ) wij ( p )

where is the forgetting factor.

Forgetting factor usually falls in the interval
between 0 and 1, typically between 0.01 and 0.1,
to allow only a little “forgetting” while limiting
the weight growth.
8
 Hebbian learning algorithm
Step 1: Initialisation.
Set initial synaptic weights and thresholds to small
random values, say in an interval [0, 1 ].

Step 2: Activation.

Compute the n neuron output at iteration p

y j ( p )   xi ( p ) wij ( p )   j
i1

where n is the number of neuron inputs, and j is the

9

threshold value of neuron j.

Step 3: Learning.
Update the weights in the network:
wij ( p  1)  wij ( p)  wij ( p)
where wij(p) is the weight correction at iteration p.
The weight correction is determined by the
generalised activity product rule:
wij ( p)   y j ( p)[ xi ( p)  wij ( p)]
Step 4: Iteration.
Increase iteration p by one, go back to Step 2.

10
 Hebbian learning example
To illustrate Hebbian learning, consider a fully
connected feedforward network with a
single layer of five computation neurons. Each
neuron is represented by a McCulloch and Pitts model
with the sign activation function. The network is
trained on the following set of input vectors:
0 0 0  0  0 
0 1 0  0  1 
         
X1  0 X 2  0 X3  0 X 4  1 X5  0
         
0 0 1  0  0 
0 1 0 0 1
11
 Initial and final states of the network
1 1 y1 1 0 y1
x1 1 1 x 1 1

0 0 y2 0 1 y2
x2 2 2 x 2 2

x3 0 3 3
0 y3 x 0 3 3
0 y3

0 0 y4 0 0 y4
x4 4 4 x 4 4

x5 1 5 5 1 y5 x 1 5 5 1 y5
Input layer Output layer Input layer Output layer

12
 Initial and final weight matrices
Ou tp u t l a ye r Ou tp u t l a yer
1 2 3 4 5 1 2 3 4 5
1 1 0 0 0 0 1 0 0 0 0 0 
0 1 0 0 0  
2
 
2
0 2.0204 0 0 2.0204
3 0 0 1 0 0 3 0 0 1.0200 0 0 
   
4
0 0 0 1 0 4 0 0 00 .9996 0 
5 0 0 0 0 1 5 0 2.0204 0 0 2.0204 

13
 A test input vector, or probe, is defined as
1 
 0
 
X   0
 
 0
1 

 When this probe is presented to the network, we

obtain:
0 0 0 0 0  1 0.4940  0
 
0 2.0204 0 0 2.0204 0 0.2661  1
 
Y  sign 0 0 1.0200 0 0  0  0.0907   0
0 0
  
0 0.9996 0  0 0.9478  0
  
 
0 2.0204 0 0 2.0204 1 0.0737  1
14
 NN Based on Competition
• Competition is important for NN
– Competition between neurons has been observed in
biological nerve systems
– Competition is important in solving many problems

• To classify an input pattern x_1 C_1

into one of the m classes
– idea case: one class node
has output 1, all other 0 ; x_n C_m
– often more than one class
nodes have non-zero output INPUT CLASSIFICATION

– If these class nodes compete with each other, maybe only one
will win eventually and all others lose (winner-takes-all). The
winner represents the computed classification of the input
15
 Competitive learning
 In competitive learning, neurons compete among
themselves to be activated.
 While in Hebbian learning, several output neurons
can be activated simultaneously, in competitive
learning, only a single output neuron is active at
any time.
 The output neuron that wins the “competition” is
called the winner-takes-all neuron.

16
• Winner-takes-all (WTA):
– Among all competing nodes, only one will win and all
others will lose
– We mainly deal with single winner WTA, but multiple
winners WTA are possible (and useful in some
applications)
– Easiest way to realize WTA: have an external, central
arbitrator (a program) to decide the winner by
comparing the current outputs of the competitors (break
the tie arbitrarily)
– This is biologically unsound (no such external
arbitrator exists in biological nerve system).

17
• Ways to realize competition in NN
– Lateral inhibition (Maxnet, Mexican hat)
output of each node feeds wij , w ji  0
xi xj
to others through inhibitory
connections (with negative weights)

– Resource competition w ii  0 w jj  0
• output of node k is distributed to
node i and j proportional to wik xi xj
and wjk , as well as xi and xj
wik w jk
• self decay
• biologically sound xk

18
19
• This rule is an example of competitive learning, and it is used for
unsupervised network training.
• Typically, winners take all learning is used for learning statistical
properties of inputs
• The learning is based on the premise that one of the neurons in
the layer, say the mth, has the maximum response due to input x
• This neuron is declared the winner.
• The weight vectors are defined as below

The increment in weight for each iteration is defines as below

20
 The individual weight adjustment becomes

 > 0 is a small learning constant, typically decreasing as learning

progresses.
The winner selection is based on the following criterion of maximum
activation among all p neurons participating in the competetion

• In this method, the winning neighborhood is sometimes extended beyond the

single neuron winner so that it includes the neighboring neurons.
• Weights are typically initialized at random values and their lengths are
normalized during learning in this method.

21
 HAMMING NET AND MAXNET

Fig: Block Diagram of minimum HD(Hamming Distance) classifier

It is a minimum Hamming distance classifier, which selects the

stored classes that are at a Minimum HD value to the noisy or
incomplete argument vector presented at the input.
22
• The Hamming network is of the feed forward type and
constitutes the first layer of the classifier.
• The p-class Hamming network has p output neurons.
• The strongest response of a neuron is indicative of the minimum
HD value between the input and the category this neuron
represents.
• The second layer of the classifier is called MAXNET and it
operates as a recurrent recall network in an auxiliary mode.

• Its only function is to suppress values at MAXNET output nodes

other than the initially maximum output node of the first layer.
• The Hamming network responsible for matching of the input
vector with stored vectors

23
Fig: Hamming network for n bit bipolar binary vectors representing p classes:
(a) classifier network and (b) neurons' activation function.
24
 The above equation is equivalent to

The weight matrix WH of the Hamming network can be created by encoding the
class vector prototypes as rows in the form as below

where the 1/2 factor is convenient for scaling purposes.

25
 Adding the fixed bias value of n/2 to the input of each neuron results
in the total input net,

A perfect match of input vector to class m, which is equivalent to the condition

HD = 0, is signaled by f(netm) = 1.

An input vector that is the complement of the prototype of class m would result
in f(netm) = 0.

26
Fig. MAXNET for p classes: (a) network architecture and (b) neuron's activation function.

27
• MAXNET needs to be employed as a second layer only for cases in which an
enhancement of the initial dominant response of the m th node is required.
• As a result of MAXNET recurrent processing, the mth node responds
positively, as opposed to all remaining nodes whose responses should have
decayed to zero.
• MAXNET is a recurrent network involving both excitatory and inhibitory
connections.
• The excitatory connection within the network is implemented in the form of
a single positive self-feedback loop with a weighting coefficient of 1.
• All the remaining connections of this fully coupled feedback network are
inhibitory.
• They are represented as M - 1 cross-feedback synapses with coefficients -E
from each output.
• The second layer weight matrix W, of size p X p is thus of the form

28
• The quantity E can be called the lateral interaction coefficient.
• With the activation function as shown previous figure and the initializing
inputs fulfilling conditions

• As a result of a number of recurrences, the only unsuppressed node will be

the one with the largest initializing entry ym0.
• This means that the only nonzero output response node is the node closest
to the input vector argument in HD sense.
• The recurrent processing by MAXNET leading to this response is

where  is a nonlinear diagonal matrix operator with entries f(*) given below

29
• Assume that ym0 > yi0:, i = 1, 2, . . . , p and i  m.
• During the first recurrence, all entries of y1 are computed
on the linear portion of f(net).
• The smallest of all yo entries will first reach the level f(net)
= 0, assumed at the k'th step.
• The clipping of one output entry slows down the decrease
of ymk+1 all forthcoming steps.
• Then, the second smallest entry of yo reaches f(net) = 0.
• The process repeats itself until all values except for one, at
the output of the m'th node, remain at nonzero values.

30
 The basic idea of competitive learning was
introduced in the early 1970s.
 In the late 1980s, Teuvo Kohonen introduced a
special class of artificial neural networks called
self-organizing feature maps. These maps are
based on competitive learning.

31
What is a self-organizing feature map?
Our brain is dominated by the cerebral cortex, a
very complex structure of billions of neurons and
hundreds of billions of synapses. The cortex
includes areas that are responsible for different
human activities (motor, visual, auditory,
somatosensory, etc.), and associated with different
sensory inputs. We can say that each sensory
input is mapped into a corresponding area of the
cerebral cortex. The cortex is a self-organizing
computational map in the human brain.

32
 The Kohonen network
 The Kohonen model provides a topological
mapping. It places a fixed number of input
patterns from the input layer into a higher-
dimensional output or Kohonen layer.
 Training in the Kohonen network begins with the
winner’s neighborhood of a fairly large size. Then,
as training proceeds, the neighborhood size
gradually decreases.

33
Feature-mapping Kohonen model
Kohonen layer Kohonen layer

Input layer Input layer

1 0 0 1
(a) (b)
34
Architecture of the Kohonen Network

y1
x1
y2
x2
y3

Input Output
layer layer
35
 The lateral connections are used to create a
competition between neurons.
 The neuron with the largest activation level
among all neurons in the output layer becomes
the winner. This neuron is the only neuron that
produces an output signal. The activity of all
other neurons is suppressed in the competition.
 The lateral feedback connections produce
excitatory or inhibitory effects, depending on the
distance from the winning neuron. This is
achieved by the use of a Mexican hat function
which describes synaptic weights between
neurons in the Kohonen layer. 36
The Mexican hat function of lateral connection

Connection
1 strength
Excitatory
effect

0 Distance
Inhibitory Inhibitory
effect effect

37
 In the Kohonen network, a neuron learns by
shifting its weights from inactive connections to
active ones. Only the winning neuron and its
neighbourhood are allowed to learn. If a neuron
does not respond to a given input pattern, then
learning cannot occur in that particular neuron.
 The competitive learning rule defines the change
wij applied to synaptic weight wij as
  ( xi  wij ), if neuron j wins the competition
 wij  
 0, if neuron j loses the competition

where xi is the input signal and  is the learning

rate parameter.
38
 The overall effect of the competitive learning rule
resides in moving the synaptic weight vector Wj of
the winning neuron j towards the input pattern X.
The matching criterion is equivalent to the
minimum Euclidean distance between vectors.
 The Euclidean distance between a pair of n-by-1
vectors X and Wj is defined by
1/ 2
 n
2
d  X  W j   ( xi  wij ) 
 i 1 

where xi and wij are the ith elements of the vectors

X and Wj, respectively. 39
 To identify the winning neuron, jX, that best
matches the input vector X, we may apply the
following condition:

jX  min X  W j , j = 1, 2, . . ., m
j

where m is the number of neurons in the Kohonen

layer.

40
 Suppose, for instance, that the 2-dimensional input
vector X is presented to the three-neuron Kohonen
network,
0.52
X 
0. 12 

 The initial weight vectors, W , are given by

j

0.27 0.42 0.43

W1    W2    W3   
 0 .81   0. 70   0 .21

41
 We find the winning (best-matching) neuron j X

using the minimum-distance Euclidean criterion:

d1  ( x1  w11) 2  ( x2  w21 ) 2  (0.52  0.27) 2  (0.12  0.81) 2  0.73

d 2  ( x1  w12 ) 2  ( x2  w22 ) 2  (0.52  0.42) 2  (0.12  0.70) 2  0.59

d 3  ( x1  w13 ) 2  ( x2  w23 ) 2  (0.52  0.43) 2  (0.12  0.21) 2  0.13

 Neuron 3 is the winner and its weight vector W3 is

updated according to the competitive learning rule.
w13  ( x1  w13 )  0.1 (0.52  0.43)  0.01

w23  ( x2  w23 )  0.1 (0.12  0.21)   0.01

42
 The updated weight vector W at iteration (p + 1)
3

is determined as:

0.43  0.01 0.44 

W3 ( p  1)  W3 ( p )  W3 ( p )       
 0 .21    0. 01  0. 20 

 The weight vector W of the wining neuron 3

3

becomes closer to the input vector X with each

iteration.

43
 Competitive Learning Algorithm
Step 1: Initialization.
Set initial synaptic weights to small random
values, say in an interval [0, 1], and assign a small
positive value to the learning rate parameter .

44
Step 2: Activation and Similarity Matching.
Activate the Kohonen network by applying the
input vector X, and find the winner-takes-all (best
matching) neuron j at iteration p, using the
X

minimum-distance Euclidean criterion

1/ 2
n
2

jX ( p)  min X  W j ( p)  [ xi  wij ( p)]  ,
j  i1 
j = 1, 2, . . ., m

where n is the number of neurons in the input

layer, and m is the number of neurons in the
Kohonen layer.
45
Step 3: Learning.
Update the synaptic weights
wij ( p  1)  wij ( p )  wij ( p )
where w (p) is the weight correction at iteration p.
ij

The weight correction is determined by the

competitive learning rule:
 [ xi  wij ( p)] , j   j ( p)
wij ( p)  
 0, j   j ( p)

where  is the learning rate parameter, and  (p) is j

the neighborhood function centered around the

winner-takes-all neuron j at iteration p.
X
46
Step 4: Iteration.
Increase iteration p by one, go back to Step 2 and
continue until the minimum-distance Euclidean
criterion is satisfied, or no noticeable changes
occur in the feature map.

47
Competitive learning in the Kohonen network
 To illustrate competitive learning, consider the
Kohonen network with 100 neurons arranged in the
form of a two-dimensional lattice with 10 rows and
10 columns. The network is required to classify
two-dimensional input vectors - each neuron in the
network should respond only to the input vectors
occurring in its region.
 The network is trained with 1000 two-dimensional
input vectors generated randomly in a square
region in the interval between –1 and +1. The
learning rate parameter a is equal to 0.1.
48
 Initial random weights
1

0.8

0.6

0.4

0.2
W(2,j)

-0.2

-0.4

-0.6

-0.8

-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
W(1,j)
49
 Network after 100 iterations
1

0.8

0.6

0.4

0.2
W(2,j)

-0.2

-0.4

-0.6

-0.8

-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
W(1,j)
50
 Network after 1000 iterations
1

0.8

0.6

0.4

0.2
W(2,j)

-0.2

-0.4

-0.6

-0.8

-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
W(1,j)
51
 Network after 10,000 iterations
1

0.8

0.6

0.4

0.2
W(2,j)

-0.2

-0.4

-0.6

-0.8

-1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
W(1,j)
52
 Competitive learning (Kohonen 1982) is a special case of
SOM (Kohonen 1989)
 In competitive learning,
 the network is trained to organize input vector space into
subspaces/classes/clusters
 each output node corresponds to one class
 the output nodes are not ordered: random map
cluster_1

• The topological order of the

cluster_2
three clusters is 1, 2, 3
w_2 • The order of their maps at
output nodes are 2, 3, 1
w_3 cluster_3 • The map does not preserve
the topological order of the
w_1
training vectors 53
 A
Height

C
IQ 54
 l a s s
s i
i f
fiiccaattiioonn
C las
TTrryy C

A
Height

C
IQ 55
 A
Height

C
IQ 56
Height

A B

C
IQ 57
 l a s s
s i
i f
fiiccaattiioonn
C las
TTrryy C
Height

A B

C
IQ 58
 Height

IQ
59
 Height

IQ
60
 Categorize the input patterns into several
classes based on the similarity among patterns.

Height

IQ 61
 Categorize the input patterns into several
classes based on the similarity among patterns.

m a y
y h
h a
a v
v e
e ?
?
l a s s
s e
e s
s w
w e
e m a
o w m
m a
a n
n y
y c
c l a s
Ho w
H
Height

IQ 62
 Categorize the input patterns into several
classes based on the similarity among patterns.

22 clusters
clusters
Height

IQ 63
 Categorize the input patterns into several
classes based on the similarity among patterns.

33 clusters
clusters
Height

IQ 64
 Categorize the input patterns into several
classes based on the similarity among patterns.

44 clusters
clusters
Height

IQ 65
 Suppose that we have p prototypes centered at x(1),
x(2), …, x(p).
 Given pattern x, it is assigned to the class label of
the ith prototype if

i  arg min dist ( x, x ) (k )

k

 Examples of distance measures include the

Hamming distance and Euclidean distance. 66
 e d Prroottoottyyppeess
r
TThhee SSttoored P

11 22
x(1) x(2)

33 44
x(3) x(4)
67
 11 22
x(1) x(2)

33  ?Class
44
x(3) x(4)
68
 Applications of neural algorithms
• Neural networks represent a class of very
powerful, general-purpose tools that have been
successfully applied to prediction, classification
and clustering problems. They are used in a
variety of areas, from speech and character
recognition to detecting fraudulent transactions,
from medical diagnosis of heart attacks to process
control and robotics, from predicting foreign
exchange rates to detecting and identifying
radar targets.

69
We demonstrate an application of a multilayer
feedforward network for printed
character recognition.

For simplicity, we can limit our task to the

recognition of digits from 0 to 9. Each digit is
represented by a 5 9 bit map.

In commercial applications, where a better

resolution is required, at least 16 16 bit maps
are used.

70
 Bit maps for digit recognition
1 2 3 4 5
6 7 8 9 10
11 12 13 14 15
16 17 18 19 20
21 22 23 24 25
26 27 28 29 30
31 32 33 34 35
36 37 38 39 40
41 42 43 44 45

71
How do we choose the architecture of a
neural network?
 The number of neurons in the input layer is
decided by the number of pixels in the bit map.
The bit map in our example consists of 45
pixels, and thus we need 45 input neurons.
 The output layer has 10 neurons – one neuron

for each digit to be recognised.

72
How do we determine an optimal number
of hidden neurons?
 Complex patterns cannot be detected by a small
number of hidden neurons; however too many
of them can dramatically increase the
computational burden.
 Another problem is overfitting. The greater the
number of hidden neurons, the greater the
ability of the network to recognise existing
patterns. However, if the number of hidden
neurons is too big, the network might simply
memorise all training examples.

73
Neural network for printed digit recognition
0 1 1 0
1 2 2 1
1 3
3 0
1
1 4
4 0
0 5 2
5 0
3
6 0
1 41 4
7 0
1 42 5

1 43
8 0
1 44 9 0
1 45 10 0

74
What are the test examples for character
recognition?

 A test set has to be strictly independent from the

training examples.
 To test the character recognition network, we
present it with examples that include “noise” –
the distortion of the input patterns.
 We evaluate the performance of the printed digit
recognition networks with 1000 test
examples (100 for each digit to
be recognised).

75
Learning curves of the digit recognition
three-layer
neural networks
10 2

10 1
1

10

4 3 2
1 -1
1 – two hidden neurons;
2 – five hidden neurons;
1 -2 3 – ten hidden neurons;
4 – twenty hidden neurons
-3
1

-4
1
0 50 150 200 250
Epochs
76
Performance evaluation of the digit
recognition
25 neural networks

0
0 0.1 0.2 0.3 0.4 0.5

77
Can we improve the performance of the
character recognition neural
network?
A neural network is as good as the examples
used to train it.
Therefore, we can attempt to improve digit
recognition by feeding the
network with
“noisy” examples of digits from 0 to 9.

78
Performance evaluation of the digit recognition
network trained with “noisy”
examples
16

14 network trained with “perfect ” examples

network trained with “noisy ” examples
12

10

0
0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.
Noise Level
79
Characteristics of NNs
Learning from experience: Complex difficult to solve
problems, but with plenty of data that describe the problem
Generalizing from examples: Can interpolate from previous
learning and give the correct response to unseen data
Rapid applications development: NNs are generic machines
and quite independent from domain knowledge
Adaptability: Adapts to a changing environment, if is
properly designed
Computational efficiency: Although the training off a neural
network demands a lot of computer power, a trained network
demands almost nothing in recall mode
Non-linearity: Not based on linear assumptions about the real
word

80
Neural Networks Projects Are Different
• Projects are data driven: Therefore, there is a need to collect and analyse data
as part of the design process and to train the neural network. This task is often
time-consuming and the effort, resources and time required are frequently
underestimated
• It is not usually possible to specify fully the solution at the design stage:
Therefore, it is necessary to build prototypes and experiment with them in order
to resolve design issues. This iterative development process can be difficult to
control
• Performance, rather than speed of processing, is the key issue: More
attention must be paid to performance issues during the requirements analysis,
design and test phases. Furthermore, demonstrating that the performance meets
the requirements can be particularly difficult.
• These issues affect the following areas :
– Project planning
– Project management
– Project documentation
81
Project/Process life cycle:
Application Identification

Feasibility Study

Design Prototype

Data Collection
Development
and validation Build Train and Test
of prototype

Optimize prototype

Validate prototype

Implement System

82
Validate System
NNs in real problems

Rest of System
Raw data
Pre-processing
Feature vector
Input encode
Network inputs
Neural Network
Network outputs
Output encode
Decoded outputs
Post-processing

Rest of System

83
Pre-processing

Transform data to NN inputs

Applying a mathematical or statistical function
Encoding textual data from a database
Selection of the most relevant data and outlier
removal
Minimizing network inputs
Feature extraction
Principal components analysis
Waveform / Image analysis
Coding pre-processing data to network inputs

84