Associative Memory : Soft Computing Course Lecture 21 24, notes, slides

www.myreaders.info/ , RC Chakraborty, e-mail [email protected] , Aug. 10, 2010

http://www.myreaders.info/html/soft_computing.html

Associative Memory

Soft Computing

Associative Memory (AM), topics : Description, content address-

ability, working, classes of AM : auto and hetero, AM related
terms - encoding or memorization, retrieval or recollection, errors
and noise, performance measure - memory capacity and content-
addressability. Associative memory models : network architectures -
linear associator, Hopfield model and bi-directional model (BAM).
Auto-associative memory (auto-correlators) : how to store patterns ?
how to retrieve patterns? recognition of noisy patterns. Bi-directional
hetero-associative memory (hetero-correlators) : BAM operations -
retrieve the nearest pair, addition and deletion of pattern pairs,
energy function for BAM - working of Kosko's BAM, incorrect recall of
pattern, multiple training encoding strategy augmentation matrix,
generalized correlation matrix and algorithm.
 Associative Memory

Soft Computing

Topics
(Lectures 21, 22, 23, 24 4 hours) Slides
1. Associative Memory (AM) Description 03-12
Content addressability; Working of AM; AM Classes : auto and hetero;
AM related terms - encoding or memorization, retrieval or recollection,
errors and noise; Performance measure - memory capacity and
content-addressability.

2. Associative Memory Models 13-20

AM Classes auto and hetero; AM Models; Network architectures -
Linear associator, Hopfield model and Bi-directional model (BAM).

3. Auto-associative Memory (auto-correlators) 21-24

How to store patterns? How to retrieve patterns? Recognition of noisy
patterns.

4. Bi-directional Hetero-associative Memory (hetero-correlators) 25-41

BAM operations - retrieve the nearest pair, Addition and deletion of
pattern pairs; Energy function for BAM - working of Kosko's BAM,
incorrect recall of pattern; Multiple training encoding strategy
augmentation matrix, generalized correlation matrix and algorithm .

5. References 42

02
What is Associative Memory ?

An associative memory is a content-addressable structure that maps a

set of input patterns to a set of output patterns.

A content-addressable structure is a type of memory that allows the

recall of data based on the degree of similarity between the input pattern
and the patterns stored in memory.

There are two types of associative memory : auto-associative and hetero-

associative.

An auto-associative memory retrieves a previously stored pattern

that most closely resembles the current pattern.

In a hetero-associative memory, the retrieved pattern is in general,

different from the input pattern not only in content but possibly also
in type and format.

Neural networks are used to implement these associative memory

models called NAM (Neural associative memory).
03
 SC - AM description
1. Associative Memory

An associative memory is a content-addressable structure that maps a

set of input patterns to a set of output patterns. A content-addressable
structure refers to a memory organization where the memory is accessed
by its content as opposed to an explicit address in the traditional computer
memory system. The associative memory are of two types : auto-associative
and hetero-associative.

An auto-associative memory retrieves a previously stored pattern that

most closely resembles the current pattern.

In hetero-associative memory, the retrieved pattern is in general different

from the input pattern not only in content but possibly also in type
and format.

04
 1.1 Description of Associative Memory

An associative memory is a content-addressable structure that allows,

the recall of data, based on the degree of similarity between the
input pattern and the patterns stored in memory.

Example : Associative Memory

The figure below shows a memory containing names of several people.

If the given memory is content-addressable,

Then using the erroneous string "Crhistpher Columbos" as key is

sufficient to retrieve the correct name "Christopher Colombus."

In this sense, this type of memory is robust and fault-tolerant, because

this type of memory exhibits some form of error-correction capability.

Alex Graham Bell

Thomas Edison
Christopher Columbus
Albert Einstein
Crhistpher Columbos Charles Darwin Christopher Columbus

Blaise Pascal
Marco Polo
Neil Armstrong
Sigmund Freud

Fig. A content-addressable memory, Input and Output

Note : An associative memory is accessed by its content, opposed

to an explicit address in the traditional computer memory system.
The memory allows the recall of information based on partial knowledge
of its contents.

[Continued in next slide]

05
 [Continued from previous slide]

Associative memory is a system that associates two patterns (X, Y)

such that when one is encountered, the other can be recalled.
The associative memory are of two types : auto-associative memory
and hetero-associative memory.

Auto-associative memory
Consider, y[1], y[2], y[3], . . . . . y[M], be the number of stored
pattern vectors and let y(m) be the components of these vectors,
representing features extracted from the patterns. The auto-associative
memory will output a pattern vector y(m) when inputting a noisy or
incomplete version of y(m).

Hetero-associative memory
Here the memory function is more general. Consider, we have a
number of key-response pairs {c(1), y(1)}, {c(2), y(2)}, . . . . . . ,
{c(M), y(M)}. The hetero-associative memory will output a pattern
vector y(m) if a noisy or incomplete verson of the c(m) is given.

Neural networks are used to implement associative memory models.

The well-known neural associative memory models are :
Linear associater is the simplest artificial neural associative
memory.
Hopfield model and Bidirectional Associative Memory (BAM)

are the other popular ANN models used as associative memories.

These models follow different neural network architectures to

memorize information.
06
 1.2 Working of Associative Memory

Example

An associative memory is a storehouse of associated patterns which

are encoded in some form.
When the storehouse is triggered or excited with a pattern, then
the associated pattern pair is recalled or appears at the output.
The input could be an exact or distorted or partial representation of
a stored pattern.
Fig below illustrates the working of an associated memory.

The associated pattern pairs

Recalled ( , ), ( , +), (7 , 4).
Pattern
Input
Pattern The association is represented

7
by the symbol

The associated pattern pairs

are stored the memory.
+ 4
.

Fig. Working of an associated memory

When the memory is triggered with an input pattern say then

the associated pattern is retrieved automatically.

07
 1.3 Associative Memory - Classes

As stated before, there are two classes of associative memory:

auto-associative and
hetero-associative memory.

An auto-associative memory, also known as auto-associative correlator,

is used to retrieve a previously stored pattern that most closely
resembles the current pattern;

A hetero-associative memory, also known as hetero-associative correlator,

is used to retrieve pattern in general, different from the input pattern
not only in content but possibly also different in type and format.

Examples

Input Recall of Presented Recall of

pattern associated distorted perfect
presented pattern pattern pattern

Hetero-associative memory Auto-associative memory

Fig. Hetero and Auto Associative memory Correlators

08
 1.4 Related Terms

Here explained : Encoding or memorization, Retrieval or recollection,

Errors and Noise, Memory capacity and Content-addressability.

Encoding or memorization
Building an associative memory means, constructing a connection
weight matrix W such that when an input pattern is presented, and
the stored pattern associated with the input pattern is retrieved.

This process of constructing the connection weight matrix is called

encoding. During encoding, for an associated pattern pair (Xk, Yk) ,

the weight values of the correlation matrix Wk are computed as

(wij)k = (xi)k (yj)k , where
th
(xi)k represents the i component of pattern Xk , and
th
(yj)k represents the j component of pattern Yk
for i = 1, 2, . . . , m and j = 1, 2, . . . , n.

Constructing of the connection weight matrix W is accomplished by

summing up the individual correlation matrices Wk , i.e.,

p Wk where
W=
k=1

is the proportionality or normalizing constant.

09
 Retrieval or recollection
After memorization, the process of retrieving a stored pattern, given
an input pattern, is called decoding.

Given an input pattern X, the decoding or recollection is accomplished by:

first compute the net input to the output units using

m
input j = xi w ij
j=1

where input j is weighted sum of the input or activation value of

node j , for j = 1, 2, ..., n.

then determine the units output using a bipolar output function:

+1 if input j j

Y j=
-1 other wise

where j is the threshold value of output neuron j .

10
 Errors and noise
The input pattern may contain errors and noise, or may be an
incomplete version of some previously encoded pattern.

When a corrupted input pattern is presented, the network will

retrieve the stored pattern that is closest to actual input pattern.

The presence of noise or errors results only in a mere decrease

rather than total degradation in the performance of the network.

Thus, associative memories are robust and fault tolerant because

of many processing elements performing highly parallel and
distributed computations.

11
 Performance Measures
The memory capacity and content-addressability are the measures
of associative memory performance for correct retrieval. These two
performance measures are related to each other.

Memory capacity refers to the maximum number of associated

pattern pairs that can be stored and correctly retrieved.

Content-addressability is the ability of the network to retrieve the

correct stored pattern.

If input patterns are mutually orthogonal - perfect retrieval is possible.

If the stored input patterns are not mutually orthogonal - non-perfect

retrieval can happen due to crosstalk among the patterns.

12
 SC - AM models
2. Associative Memory Models

An associative memory is a system which stores mappings of specific input

representations to specific output representations.

An associative memory "associates" two patterns such that when one is

encountered, the other can be reliably recalled.

Most associative memory implementations are realized as connectionist

networks.

The simplest associative memory model is Linear associator, which is a feed-

forward type of network. It has very low memory capacity and therefore
not much used.

The popular models are Hopfield Model and Bi-directional Associative

Memory (BAM) model.

The Network Architecture of these models are presented in this section.

13
 2.1 Associative Memory Models

The simplest and among the first studied associative memory models
is Linear associator. It is a feed-forward type of network where the
output is produced in a single feed-forward computation. It can be
used as an auto-associator as well as a hetero-associator, but it
possesses a very low memory capacity and therefore not much used.

The popular associative memory models are Hopfield Model and

Bi-directional Associative Memory (BAM) model.

The Hopfield model is an auto-associative memory, proposed by

John Hopfield in 1982. It is an ensemble of simple processing units
that have a fairly complex collective computational abilities and
behavior. The Hopfield model computes its output recursively in
time until the system becomes stable. Hopfield networks are
designed using bipolar units and a learning procedure.

The Bi-directional associative memory (BAM) model is similar to

linear associator, but the connections are bi-directional and
therefore allows forward and backward flow of information between
the layers. The BAM model can perform both auto-associative
and hetero-associative recall of stored information.

The network architecture of these three models are described in

the next few slides.
14
 2.2 Network Architectures of AM Models

The neural associative memory models follow different neural network

architectures to memorize information. The network architectures
are either single layer or two layers .

The Linear associator model, is a feed forward type network,

consists, two layers of processing units, one serving as the input layer
while the other as the output layer.

The Hopfield model, is a single layer of processing elements where

each unit is connected to every other unit in the network other
than itself.

The Bi-directional associative memory (BAM) model is similar to

that of linear associator but the connections are bidirectional.

In this section, the neural network architectures of these models and

the construction of the corresponding connection weight matrix W of
the associative memory are illustrated.

15
 Linear Associator Model (two layers)

It is a feed-forward type network where the output is produced in a

single feed-forward computation. The model consists of two layers
of processing units, one serving as the input layer while the other as
the output layer. The inputs are directly connected to the outputs,
via a series of weights. The links carrying weights connect every input
to every output. The sum of the products of the weights and the
inputs is calculated in each neuron node. The network architecture
of the linear associator is as shown below.

neurons
weights wij

x1 w11 y1
w12 w21

w22
y2
x2
w2m outputs
inputs w1m
wn1

wn2
Ym
Xn
wnm

Fig. Linear associator model

all n input units are connected to all m output units via connection

weight matrix W = [wij]n x m where wij denotes the strength

th th
of the unidirectional connection from the i input unit to the j
output unit.

the connection weight matrix stores the p different associated

pattern pairs {(Xk, Yk) | k = 1, 2, ..., p} .

building an associative memory is constructing the connection

weight matrix W such that when an input pattern is presented,
then the stored pattern associated with the input pattern is retrieved.
[Continued in next slide]
16
 [Continued from previous slide]

Encoding : The process of constructing the connection weight matrix

is called encoding. During encoding the weight values of correlation
matrix Wk for an associated pattern pair (Xk, Yk) are computed as:
(wij)k = (xi)k (yj)k where
th
(xi)k is the i component of pattern Xk for i = 1, 2, ..., m, and
th
(yj)k is the j component of pattern Yk for j = 1, 2, ..., n.

Weight matrix : Construction of weight matrix W is accomplished

p
by summing those individual correlation matrices Wk, ie, W = Wk
k=1
where is the constant of proportionality, for normalizing, usually
set to 1/p to store p different associated pattern pairs.

Decoding : After memorization, the network can be used for retrieval;

the process of retrieving a stored pattern, is called decoding;
given an input pattern X, the decoding or retrieving is accomplished
m
by computing, first the net Input as input j = xi w i j where
j=1
input j stands for the weighted sum of the input or activation value of
th
node j , for j = 1, 2, . . , n. and xi is the i component of pattern Xk ,
and then determine the units Output using a bipolar output function:
+1 if input j j
Y j=
-1 other wise
where j is the threshold value of output neuron j .

Note: The output units behave like linear threshold units; that
compute a weighted sum of the input and produces a -1 or +1
depending whether the weighted sum is below or above a certain
threshold value.

Performance : The input pattern may contain errors and noise, or an

incomplete version of some previously encoded pattern. When such
corrupt input pattern is presented, the network will retrieve the stored
pattern that is closest to actual input pattern. Therefore, the linear
associator is robust and fault tolerant. The presence of noise or error
results in a mere decrease rather than total degradation in the
performance of the network.
17
 Auto-associative Memory Model - Hopfield model (single layer)

Auto-associative memory means patterns rather than associated

pattern pairs, are stored in memory. Hopfield model is one-layer
unidirectional auto-associative memory.
Hopfield network alternate view

W24 W34
1
W14
W13 W23 W43
W12 W32 W42
W21 W31
W41
4 2

connection
1 2 3 4 neurons

weights I1
wij I2 I3 I4
inputs I 3

V1 V2 V3 V4
outputs V

Fig. Hopfield model with four units

the model consists, a single layer of processing elements where each

unit is connected to every other unit in the network but not to itself.
connection weight between or from neuron j to i is given by a
number wij. The collection of all such numbers are represented
by the weight matrix W which is square and symmetric, ie , w i j = w j i

for i, j = 1, 2, . . . . . , m.
each unit has an external input I which leads to a modification
in the computation of the net input to the units as
m
input j = xi w i j + I j for j = 1, 2, . . ., m.
i=1

th
and xi is the i component of pattern Xk

each unit acts as both input and output unit. Like linear associator,
a single associated pattern pair is stored by computing the weight
T
matrix as Wk = X k Yk where XK = YK
[Continued in next slide]
18
 [Continued from previous slide]

Weight Matrix : Construction of weight matrix W is accomplished by

p
summing those individual correlation matrices, ie, W = Wk where

k=1
is the constant of proportionality, for normalizing, usually set to 1/p
to store p different associated pattern pairs. Since the Hopfield
model is an auto-associative memory model, it is the patterns
rather than associated pattern pairs, are stored in memory.
Decoding : After memorization, the network can be used for retrieval;
the process of retrieving a stored pattern, is called decoding; given an
input pattern X, the decoding or retrieving is accomplished by
m
computing, first the net Input as input j = xi w ij where input j
j=1
stands for the weighted sum of the input or activation value of node j ,
th
for j = 1, 2, ..., n. and xi is the i component of pattern Xk , and
then determine the units Output using a bipolar output function:
+1 if input j j
Y j=
-1 other wise
where j is the threshold value of output neuron j .
Note: The output units behave like linear threshold units; that compute a
weighted sum of the input and produces a -1 or +1 depending whether the
weighted sum is below or above a certain threshold value.

Decoding in the Hopfield model is achieved by a collective and

recursive relaxation search for a stored pattern given an initial
stimulus pattern. Given an input pattern X, decoding is accomplished
by computing the net input to the units and determining the output of
those units using the output function to produce the pattern X'. The
pattern X' is then fed back to the units as an input pattern to produce
the pattern X''. The pattern X'' is again fed back to the units to
produce the pattern X'''. The process is repeated until the network
stabilizes on a stored pattern where further computations do not
change the output of the units.
In the next section, the working of an auto-correlator : how to store
patterns, recall a pattern from the stored patterns and how to
recognize a noisy pattern are explained.
19
 Bidirectional Associative Memory (two-layer)

Kosko (1988) extended the Hopfield model, which is single layer,

by incorporating an additional layer to perform recurrent auto-
associations as well as hetero-associations on the stored memories.
The network structure of the bidirectional associative memory model is
similar to that of the linear associator but the connections are
bidirectional; i.e.,

wij = wji , for i = 1, 2, . . . , n and j = 1, 2, . . . , m.

neurons neurons
weights wij
x1 w11
y1
w21
w12
y2
w22
x2
outputs
w2m
inputs
wn1 w1m

wn2

Xn wnm
Ym

Fig. Bidirectional Associative Memory model

In the bidirectional associative memory, a single associated pattern

T
pair is stored by computing the weight matrix as Wk = X k Yk .

the construction of the connection weight matrix W, to store p

different associated pattern pairs simultaneously, is accomplished
by summing up the individual correlation matrices Wk ,
p
i.e., W = Wk
k=1

where is the proportionality or normalizing constant.

20
 SC - AM auto correlator
3. Auto-associative Memory (auto-correlators)

In the previous section, the structure of the Hopfield model has been
explained. It is an auto-associative memory model which means patterns,
rather than associated pattern pairs, are stored in memory. In this
section, the working of an auto-associative memory (auto-correlator) is
illustrated using some examples.

Working of an auto-correlator :

how to store the patterns,

how to retrieve / recall a pattern from the stored patterns, and
how to recognize a noisy pattern

21
 How to Store Patterns : Example
Consider the three bipolar patterns A1 , A2, A3 to be stored as
an auto-correlator.
A1 = (-1, 1 , -1 , 1 )
A2 = ( 1, 1 , 1 , -1 )

A3 = (-1, -1 , -1 , 1 )

Note that the outer product of two vectors U and V is

U1 U1V1 U1V2 U1V3

U2 V1 V2 V3
U V = UT V = U3 =
U2V1 U2V2 U2V3
U3V1 U3V2 U3V3
U4 U4V1 U4V2 U4V3

Thus, the outer products of each of these three A1 , A2, A3 bipolar

patterns are
j

T 1 -1 1 -1
-1 1 -1 1
[A1 ] 4x1 [A1 ] 1x4 =
1 -1 1 -1
-1 1 -1 1
i

T 1 1 1 -1
1 1 1 -1
[A2 ] 4x1 [A2 ] 1x4 =
1 1 1 -1
-1 -1 -1 1
i

T
1 1 1 -1
1 1 1 -1
[A3 ] 4x1 [A3 ] 1x4 =
1 1 1 -1
-1 -1 -1 1
i

Therefore the connection matrix is

j
3 1 3 -3
3 T
T = [t i j ] = 1 3 1 -1
[Ai ]
[Ai ] 4x1 1x4
3 1 3 -3
i=1 = i -3 -1 -3 3
This is how the patterns are stored .
22
 Retrieve a Pattern from the Stored Patterns (ref. previous slide)

The previous slide shows the connection matrix T of the three

bipolar patterns A1 , A2, A3 stored as
i
3
3 1 3 -3
T
1 3 1 -1
T = [t i j ] = i= [Ai ] 4x1 [Ai ] 1x4 =
1
3 1 3 -3
j -3 -1 -3 3

and one of the three stored pattern is A2 = ( 1, 1 , 1 , -1 )

ai
Retrieve or recall of this pattern A2 from the three stored patterns.

The recall equation is

old
= (ai t ij , ) for j = 1 , 2 , . . . , p
new aj
aj

Computation for the recall equation A2 yields = ai t i j and

then find

i= 1 2 3 4
= ai t i , j=1 1x3 + 1x1 + 1x3 + -1x-3 = 10 1
= ai t i , j=2 1x1 + 1x3 + 1x1 + -1x-1 = 6 1
= ai t i , j=3 1x3 + 1x1 + 1x3 + -1x-3 = 10 1
= ai t i , j=4 1x-3 + 1x-1 + 1x-3 + -1x3 = -1 -1

new old ) for j = 1 , 2 , . . . , p is ( , )

Therefore aj = (ai t ij , aj
new
a1
= (10 , 1)
new
a2
= (6 , 1)
new
a3
new
= (10 , 1)
a4
= (-1 , -1)

The values of is the vector pattern ( 1, 1 , 1 , -1 ) which is A2 .

This is how to retrieve or recall a pattern from the stored patterns.

Similarly, retrieval of vector pattern A3 as

new new new
new

( a1 , a , a3 , a 4 , ) = ( -1, -1 , -1 , 1 ) = A3
2

23
 Recognition of Noisy Patterns (ref. previous slide)

Consider a vector A' = ( 1, 1 , 1 , 1 ) which is a noisy presentation

of one among the stored patterns.
find the proximity of the noisy vector to the stored patterns
using Hamming distance measure.
note that the Hamming distance (HD) of a vector X from Y, where

X = (x1 , x2 , . . . , xn) and Y = ( y1, y2 , . . . , yn) is given by

m | (xi - yi ) |
HD (x , y)
= i=
1

The HDs of A' from each of the stored patterns A1 , A2, A3 are

HD (A' , A1) = |(x1 - y1 )|, |(x2 - y2)|, |(x3 - y3 )|, |(x4 - y4 )|

= |(1 - (-1))|, |(1 - 1)|, |(1 - (-1) )|, |(1 - 1)|
= 4
HD (A' , A2) = 2
HD (A' , A3) = 6
Therefore the vector A' is closest to A2 and so resembles it.
In other words the vector A' is a noisy version of vector A2.
Computation of recall equation using vector A' yields :

i= 1 2 3 4
= ai t i , j=1 1x3 + 1x1 + 1x3 + 1x-3 = 4 1
= ai t i , j=2 1x1 + 1x3 + 1x1 + 1x-1 = 4 1
= ai t i , j=3 1x3 + 1x1 + 1x3 + 1x-3 = 4 1
= ai t i , j=4 1x-3 + 1x-1 + 1x-3 + 1x3 = -4 -1

new = (ai t ij old ) for j = 1 , 2 , . . . , p is ( , )

Therefore aj , aj
new
a1 = (4 , 1)

new
= (4 , 1)
a2
new
= (4 , 1)
a3
new
a4 = (-4 , -1)

The values of is the vector pattern ( 1, 1 , 1 , -1 ) which is A2 .

Note : In presence of noise or in case of partial representation of vectors,
an autocorrelator results in the refinement of the pattern or removal of
noise to retrieve the closest matching stored pattern.
24
 SC - Bidirectional hetero AM
4. Bidirectional Hetero-associative Memory

The Hopfield one-layer unidirectional auto-associators have been discussed

in previous section. Kosko (1987) extended this network to two-layer
bidirectional structure called Bidirectional Associative Memory (BAM) which
can achieve hetero-association. The important performance attributes of the
BAM is its ability to recall stored pairs particularly in the presence of noise.

Definition : If the associated pattern pairs (X, Y) are different and if the
model recalls a pattern Y given a pattern X or vice-versa, then it is
termed as hetero-associative memory.

This section illustrates the bidirectional associative memory :

Operations (retrieval, addition and deletion) ,

Energy Function (Kosko's correlation matrix, incorrect recall of pattern),
Multiple training encoding strategy (Wang's generalized correlation matrix).

25
 4.1 Bidirectional Associative Memory (BAM) Operations

BAM is a two-layer nonlinear neural network.

Denote one layer as field A with elements Ai and the other layer
as field B with elements Bi.

The basic coding procedure of the discrete BAM is as follows.

Consider N training pairs { (A1 , B1) , (A2 , B2), . . , (Ai , Bi), . . (AN , BN) }
where Ai = (ai1 , ai2 , . . . , ain) and Bi = (bi1 , bi2 , . . . , bip) and
aij , bij are either in ON or OFF state.
in binary mode , ON = 1 and OFF = 0 and

in bipolar mode, ON = 1 and OFF = -1

N
the original correlation matrix of the BAM is M0 = T
[ Xi ] [ Yi ]
i=
1
where Xi = (xi1 , xi2 , . . . , xin) and Yi = (yi1 , yi2 , . . . , yip)

and xij(yij) is the bipolar form of aij(bij)

The energy function E for the pair ( , ) and correlation matrix M is

T
E= -M

With this background, the decoding processes, means the operations

to retrieve nearest pattern pairs, and the addition and deletion of
the pattern pairs are illustrated in the next few slides.

26
 Retrieve the Nearest of a Pattern Pair, given any pair
(ref : previous slide)

Example
Retrieve the nearest of (Ai , Bi) pattern pair, given any pair ( , ) .

The methods and the equations for retrieve are :

start with an initial condition which is any given pattern pair ( , ),

determine a finite sequence of pattern pairs (' , ' ) , (" , " ) . . . .

until an equilibrium point (f , f ) is reached, where

T
' = ( M ) and ' = ( ' M )
T
" = (' M ) and " = ( '' M )
(F) = G = g1 , g2 , . . . . , gr ,
F = ( f1 , f2 , . . . . , fr )
M is correlation matrix
1 if fi >0
0 (binary)

gi = , fi < 0

-1 (bipolar)
previous g i , fi = 0

Kosko has proved that this process will converge for any
correlation matrix M.

27
 Addition and Deletion of Pattern Pairs
Given a set of pattern pairs (Xi , Yi) , for i = 1 , 2, . . . , n and a set
of correlation matrix M :
a new pair (X' , Y') can be added or
an existing pair (Xj , Yj) can be deleted from the memory model.

Addition : add a new pair (X' , Y') , to existing correlation matrix M ,

them the new correlation matrix Mnew is given by

T T T T
Mnew = X1 Y1 + X1 Y1 + . . . . + Xn Yn + X' Y'

Deletion : subtract the matrix corresponding to an existing pair (Xj , Yj)

from the correlation matrix M , them the new correlation matrix Mnew
is given by
T
Mnew = M - ( Xj Yj )

Note : The addition and deletion of information is similar to the

functioning of the system as a human memory exhibiting learning
and forgetfulness.

28
 4.2 Energy Function for BAM

Note : A system that changes with time is a dynamic system. There are two
types of dynamics in a neural network. During training phase it iteratively update
weights and during production phase it asymptotically converges to the solution
patterns. State is a collection of qualitative and qualitative items that characterize
the system e.g., weights, data flows. The Energy function (or Lyapunov function)
is a bounded function of the system state that decreases with time and the
system solution is the minimum energy.
Let a pair (A , B) defines the state of a BAM.
to store a pattern, the value of the energy function for that pattern

has to occupy a minimum point in the energy landscape.

also adding a new patterns must not destroy the previously
stored patterns.
The stability of a BAM can be proved by identifying the energy function E

with each state (A , B) .

For auto-associative memory : the energy function is
T
E(A) = - AM A
For bidirecional hetero associative memory : the energy function is

E(A, B) = - T
AM B ; for a particular case A = B , it corresponds

to Hopfield auto-associative function.

We wish to retrieve the nearest of (Ai , Bi) pair, when any ( , ) pair
is presented as initial condition to BAM. The neurons change
their states until a bidirectional stable state (Af , Bf) is reached. Kosko
has shown that such stable state is reached for any matrix M when it
corresponds to local minimum of the energy function. Each cycle of
decoding lowers the energy E if the energy function for any point
T
( , ) is given E = M
by T
If the energy E = Ai M Bi evaluated using coordinates of the pair
(Ai , Bi) does not constitute a local minimum, then the point cannot

be recalled, even though one starts with = Ai. Thus Kosko's encoding
method does not ensure that the stored pairs are at a local minimum.

29
 Example : Kosko's BAM for Retrieval of Associated Pair
The working of Kosko's BAM for retrieval of associated pair.
Start with X3, and hope to retrieve the associated pair Y3 .

Consider N = 3 pattern pairs (A1 , B1 ) , (A2 , B2 ) , (A3 , B3 ) given by

A1 = ( 1 0 0 0 0 1 ) B1 = ( 1 1 0 0 0 )
A2 = ( 0 1 1 0 0 0 ) B2 = ( 1 0 1 0 0 )
A3 = ( 0 0 1 0 1 1 ) B3 = ( 0 1 1 1 0 )

Convert these three binary pattern to bipolar form replacing 0s by -1s.

X1 = ( 1 -1 -1 -1 -1 1 ) Y1 = ( 1 1 -1 -1 -1 )
X2 = ( -1 1 1 -1 -1 -1 ) Y2 = ( 1 -1 1 -1 -1 )
X3 = ( -1 -1 1 -1 1 1 ) Y3 = ( -1 1 1 1 -1 )

The correlation matrix M is calculated as 6x5 matrix

1 1 -3 -1 1
1 -3 1 -1 1

Y1 + T
T -1 -1 3 1 -1
+ X3 Y3 =
T X2 Y2 -1 -1 -1 1 3
M = X1
-3 1 1 3 1
-1 3 -1 1 -1

Suppose we start with = X3, and we hope to retrieve the associated pair

Y3 . The calculations for the retrieval of Y3 yield :

M = ( -1 -1 1 -1 1 1 ) ( M ) = ( -6 6 6 6 -6 )
( M) = ' = ( -1 1 1 1 -1 )
T
' M = ( -5 -5 5 -3 7 5 )
T
(' M ) = ( -1 -1 1 -1 1 1 ) = '
' M = ( -1 -1 1 -1 1 1 ) M = ( -6 6 6 6 -6 )
(' M) = " = ( -1 1 1 1 1 -1 )
= '

This retrieved patern ' is same as Y3 .

Hence, (f , f) = (X3 , Y3 ) is correctly recalled, a desired result .

30
 Example : Incorrect Recall by Kosko's BAM

The Working of incorrect recall by Kosko's BAM.

Start with X2, and hope to retrieve the associated pair Y2 .

Consider N = 3 pattern pairs (A1 , B1 ) , (A2 , B2 ) , (A3 , B3 ) given by

A1 = ( 1 0 0 1 1 1 0 0 0 ) B1 = ( 1 1 1 0 0 0 0 1 0 )
A2 = ( 0 1 1 1 0 0 1 1 1 ) B2 = ( 1 0 0 0 0 0 0 0 1 )
A3 = ( 1 0 1 0 1 1 0 1 1 ) B3 = ( 0 1 0 1 0 0 1 0 1 )

Convert these three binary pattern to bipolar form replacing 0s by -1s.

X1 = ( 1 -1 -1 1 1 1 -1 -1 -1 ) Y1 = ( 1 1 1 -1 -1 -1 -1 1 -1 )
X2 = ( -1 1 1 1 -1 -1 1 11 ) Y2 = ( 1 -1 -1 -1 -1 -1 -1 -1 1 )
X3 = ( 1 -1 1 -1 1 1 -1 1 1 ) Y3 = ( -1 1 -1 1 -1 -1 1 0 1 )

The correlation matrix M is calculated as 9 x 9 matrix

T T Y2 + T
M = X1 Y1 + X2 X3 Y3

-1 3 1 1 -1 -1 1 1 -1
1 -3 -1 -1 1 1 -1 -1 1
-1 -1 -3 1 -1 -1 1 -3 3
3 -1 1 -3 -1 -1 -3 1 -1
= -1 3 1 1 -1 -1 1 1 -1
-1 3 1 1 -1 -1 1 1 -1
1 -3 -1 -1 1 1 -1 -1 1
-1 -1 -3 1 -1 -1 1 -3 3
-1 -1 -3 1 -1 -1 1 -3 3

(Continued in next slide)

31
 SC - Bidirectional hetero AM
[Continued from previous
slide]
Let the pair (X2 , Y2 ) be recalled.
X2 = ( -1 1 1 1 -1 -1 1 11 ) Y2 = ( 1 -1 -1 -1 -1 -1 -1 -1 1 )

Start with = X2, and hope to retrieve the associated pair Y2 .

The calculations for the retrieval of Y2 yield :

M = ( 5 -19 -13 -5 1 1 -5 -13 13 )

( M) = ( 1 -1 -1 -1 1 1 -1 -1 1 ) = '
' MT = ( -11 11 5 5 -11 -11 11 5 5 )

(' = ( -1 1 1 1 -1 -1 1 1 1 ) = '
T
MM
' ) = ( 5 -19 -13 -5 1 1 -5 -13 13 )

(' M) = ( 1 -1 -1 -1 1 1 -1 -1 1 ) = "
= '

Here " = ' . Hence the cycle terminates with

F = ' = ( -1 1 1 1 -1 -1 1 1 1 ) = X2

F = ' = ( 1 -1 -1 -1 1 1 -1 -1 1 ) Y2

But ' is not Y2 . Thus the vector pair (X2 , Y2) is not recalled
correctly by Kosko's decoding process.

( Continued in next slide )

32
 Check with Energy Function : Compute the energy functions
T
= -71
for the coordinates of pair (X2 , Y2) , the energy E2 = - X2 M Y2

for the coordinates of pair (F , F) , the energy EF = - F M T

F = -75

However, the coordinates of pair (X2 , Y2) is not at its local

minimum can be shown by evaluating the energy E at a point which
is "one Hamming distance" way from Y2 . To do this consider a point

Y2 ' = ( 1 -1 -1 -1 1 -1 -1 -1 1 )

where the fifth component -1 of Y2 has been changed to 1. Now

T
E = - X2 M Y 2
= - 73
which is lower than E2 confirming the hypothesis that (X2 , Y2) is not
at its local minimum of E.

Note : The correlation matrix M used by Kosko does not guarantee

that the energy of a training pair is at its local minimum. Therefore , a
pair Pi can be recalled if and only if this pair is at a local minimum
of the energy surface.

33
 SC - Bidirectional hetero AM
4.3 Multiple Training Encoding Strategy

Note : (Ref. example in previous section). Kosko extended the unidirectional

auto-associative to bidirectional associative processes, using correlation matrix
T
M= Xi Yi computed from the pattern pairs. The system proceeds to

retrieve the nearest pair given any pair ( , ), with the help of recall
equations. However, Kosko's encoding method does not ensure that the stored
pairs are at local minimum and hence, results in incorrect recall.

Wang and other's, introduced multiple training encoding strategy which

ensures the correct recall of pattern pairs. This encoding strategy is an
enhancement / generalization of Kosko's encoding strategy. The Wang's
T
generalized correlation matrix is M = qi Xi Yi where qi is viewed
T
as pair weight for Xi Yi as positive real numbers. It denotes the
minimum number of times for using a pattern pair (Xi , Yi) for training to
guarantee recall of that pair.

To recover a pair (Ai , Bi ) using multiple training of order q, let us

augment or supplement matrix M with a matrix P defined as
T where (Xi , Yi) are the bipolar form of (Ai , Bi).
P = (q Xi Yi
1)

The augmentation implies adding (q - 1) more pairs located at (Ai , Bi) to

the existing correlation matrix. As a result the energy E' can reduced to
an arbitrarily low value by a suitable choice of q. This also ensures that
the energy at (Ai , Bi) does not exceed at points which are one Hamming
distance away from this location.

The new value of the energy function E evaluated at (Ai , Bi) then becomes
T XTi Yi Bi T
E' (Ai , Bi) = Ai M Bi (q 1) Ai

The next few slides explains the step-by-step implementation of

Multiple training encoding strategy for the recall of three pattern pairs
(X1 , Y1 ) , (X1 , Y1 ) , (X1 , Y1 ) using one and same augmentation matrix

M . Also an algorithm to summarize the complete process of multiple

training encoding is given.
34
 Example : Multiple Training Encoding Strategy

The working of multiple training encoding strategy which ensures

the correct recall of pattern pairs.

Consider N = 3 pattern pairs (A1 , B1 ) , (A2 , B2 ) , (A3 , B3 ) given by

A1 = ( 1 0 0 1 1 1 0 0 0 ) B1 = ( 1 1 1 0 0 0 0 1 0 )
A2 = ( 0 1 1 1 0 0 1 1 1 ) B2 = ( 1 0 0 0 0 0 0 0 1 )
A3 = ( 1 0 1 0 1 1 0 1 1 ) B3 = ( 0 1 0 1 0 0 1 0 1 )

Convert these three binary pattern to bipolar form replacing 0s by -1s.

X1 = ( 1 -1 -1 1 1 1 -1 -1 -1 ) Y1 = ( 1 1 1 -1 -1 -1 -1 1 -1 )
X2 = ( -1 1 1 1 -1 -1 1 11 ) Y2 = ( 1 -1 -1 -1 -1 -1 -1 -1 1 )
X3 = ( 1 -1 1 -1 1 1 -1 1 1 ) Y3 = ( -1 1 -1 1 -1 -1 1 0 1 )

Let the pair (X2 , Y2) be recalled.

X2 = ( -1 1 1 1 -1 -1 1 11 ) Y2 = ( 1 -1 -1 -1 -1 -1 -1 -1 1 )
T
the augmented correlation matrix M
T
Choose q=2, so that P = X2 Y2 ,
T T

becomes M = X1 Y1 + 2 X2 Y2 + X 3 Y3

4 2 2 0 0 2 2 -2
2 -4 -2 -2 0 0 -2 -2 2
0 -2 -4 0 -2 -2 0 -4 4
= 4 -2 0 -4 -2 -2 -4 0 0
-2 4 2 2 0 0 2 2 -2
-2 4 2 2 0 0 2 2 -2
2 -4 -2 -2 0 0 -2 -2 2
0 -2 -4 0 -2 -2 0 -4 4
0 -2 -4 0 -2 -2 0 -4 4

( Continued in next slide )

35
 SC - Bidirectional hetero AM
[Continued from previous slide]

Now give = X2, and see that the corresponding pattern pair = Y2

is correctly recalled as shown below.

M = ( 14 -28 -22 -14 -8 -8 -14 -22 22 )

( M) = ( 1 -1 -1 -1 -1 -1 -1 -1 1 ) = '
T
' M = ( -16 16 18 18 -16 -16 16 18 18 )

(' MT) = ( -1 1 1 1 -1 -1 1 1 1 ) = '

' M = ( 14 -28 -22 -14 -8 -8 -14 -22 23 )

(' M) = ( 1 -1 -1 -1 1 1 -1 -1 1 ) = "

Here " = ' . Hence the cycle terminates with

F = ' = ( -1 1 1 1 -1 -1 1 1 1 ) = X2

F = ' = ( 1 -1 -1 -1 1 1 -1 -1 1 ) = Y2

Thus, (X2 , Y2 ) is correctly recalled, using augmented correlation

matrix M . But, it is not possible to recall (X1 , Y1) using the same
matrix M as shown in the next slide.

( Continued in next slide )

36
 Note : The previous slide showed that the pattern pair (X2 , Y2 ) is correctly
recalled, using augmented correlation matrix
T
T T
M = X 1 Y1 X3 Y3
+ 2 X2 Y2 +

but then the same matrix M can not recall correctly the other
pattern pair (X1 , Y1 ) as shown below.

X1 = ( 1 -1 -1 1 1 1 -1 -1 -1 ) Y1 = ( 1 1 1 -1 -1 -1 -1 1 -1 )

Let = X1 and to retrieve the associated pair Y1 the calculation shows

M = ( -6 24 22 6 4 4 6 22 -22 )

( M) = ( -1 1 1 1 1 1 1 1 -1 ) = '
T
' M = ( 16 -16 -18 -18 16 16 -16 -18 -18 )
T
(' M ) = ( 1 -1 -1 -1 1 1 -1 -1 -1 ) = '
' M = ( -14 28 22 14 8 8 14 22 -22 )

(' M) = ( -1 1 1 1 1 1 1 1 -1 ) = "

Here " = ' . Hence the cycle terminates with

F = ' = ( 1 -1 -1 -1 1 1 -1 -1 -1 ) = X1

F = ' = ( -1 1 1 1 1 1 1 1 -1 ) Y1

Thus, the pattern pair (X1 , Y1 ) is not correctly recalled, using augmented

correlation matrix M.

To tackle this problem, the correlation matrix M needs to be further

augmented by multiple training of (X1 , Y1 ) as shown in the next slide.
( Continued in next slide )
37
 The previous slide shows that pattern pair (X1 , Y1) cannot be recalled
under the same augmentation matrix M that is able to recall (X2 , Y2).

However, this problem can be solved by multiple training of (X1 , Y1)

which yields a further change in M to values by defining
T Y2 + T
T Y1 + 2 X2 X3 Y3
M = 2 X1

-1 5 3 1 -1 -1 1 3 -3
1 -5 -3 -1 1 1 -1 -3 3
-1 -3 -5 1 -1 -1 1 -5 5
5 -1 1 -5 -3 -3 -5 1 -1
= -1 5 3 1 -1 -1 1 3 -3
-1 5 3 1 -1 -1 1 3 -3
1 -5 -3 -1 1 1 -1 -3 3
-1 -3 -5 1 -1 -1 1 -5 5
-1 -3 -5 1 -1 -1 1 -5 5

Now observe in the next slide that all three pairs can be correctly recalled.

( Continued in next slide )

38
 Recall of pattern pair (X1 , Y1 )
X1 = ( 1 -1 -1 1 1 1 -1 -1 -1 ) Y1 = ( 1 1 1 -1 -1 -1 -1 1 -1 )

Let = X1 and to retrieve the associated pair Y1 the calculation shows

-3 31 -31 M = ( 3 33 31 -3 -5 -5 )
-1 ( M) = ( 1 1 1 -1 -1 -1 1 -1 ) = '
(' M T ) = ( 13 -13 -19 23 13 13 -13 -19 -19 )
(' M T ) = ( 1 -1 -1 1 1 1 -1 -1 -1 ) = '
' M = ( 3 33 31 -3 -5 -5 -3 31 -31 )
(' M) = ( 1 1 1 -1 -1 -1 -1 1 -1 ) = "

Here " = ' . Hence the cycle terminates with

F = ' = ( 1 -1 -1 1 1 1 -1 -1 -1 ) = X1
F = ' = ( 1 1 1 -1 -1 -1 -1 1 -1 ) = Y1

Thus, the pattern pair (X1 , Y1 ) is correctly recalled

Recall of pattern pair (X2 , Y2 )

X2 = ( -1 1 1 1 -1 -1 1 1 1) Y2 = ( 1 -1 -1 -1 -1 -1 -1 -1 1 )

Let = X2 and to retrieve the associated pair Y2 the calculation shows

M = ( 7 -35 -29 -7 -1 -1 -7 -29 29 )

( M) = ( 1 -1 -1 -1 -1 -1 -1 -1 1 ) = '
(' M T ) = ( -15 15 17 19 -15 -15 15 17 17 )
(' M T ) = ( -1 1 1 1 -1 -1 1 1 1 ) = '
' M = ( 7 -35 -29 -7 -1 -1 -7 -29 29 )
(' M) = ( 1 -1 -1 -1 -1 -1 -1 -1 1 ) = "

Here " = ' . Hence the cycle terminates with

F = ' = ( -1 1 1 1 -1 -1 1 1 1 ) = X2
F = ' = ( 1 -1 -1
-1 -1 -1 -1 -1 1 ) = Y2

Thus, the pattern pair (X2 , Y2 ) is correctly recalled

Recall of pattern pair (X3 , Y3 )

X3 = ( 1 -1 1 -1 1 1 -1 1 1) Y3 = ( -1 1 -1 1 -1 -1 1 0 1)

Let = X3 and to retrieve the associated pair Y3 the calculation shows

M = ( -13 17 -1 13 -5 -5 13 -1 1 )
( M) = ( -1 1 -1 1 -1 -1 1 -1 1 ) = '
(' M T ) = ( 11 -11 27 -63 11 11 -11 27 27 )
(' M T ) = ( 1 -1 1 -1 1 1 -1 1 1 ) = '
' M = ( -13 17 -1 13 -5 -5 13 -1 1 )
(' M) = ( -1 1 -1 1 -1 -1 1 -1 1 ) = "

Here " = ' . Hence the cycle terminates with

F = ' = ( 1 -1 1 -1 1 1 -1 1 1 ) = X3
F = ' = ( -1 1 -1 1 -1 -1 1 0 1 ) = Y3

Thus, the pattern pair (X3 , Y3 ) is correctly recalled

( Continued in next slide )

39
 Thus, the multiple training encoding strategy ensures the correct recall
of a pair for a suitable augmentation of M . The generalization of the
correlation matrix, for the correct recall of all training pairs, is written as
N T
M = qi X i Yi where qi 's are +ve real numbers.
i=1

This modified correlation matrix is called generalized correlation matrix.

Using one and same augmentation matrix M, it is possible to recall all
the training pattern pairs .

40
 SC - Bidirectional hetero AM
Algorithm (for the Multiple training encoding strategy)

To summarize the complete process of multiple training encoding an

algorithm is given below.

Algorithm Mul_Tr_Encode ( N , Xi , , qi ) where

Yi

N : Number of stored patterns set

X,i Yi: the bipolar pattern pairs

( X1, X2, . . . . , where
X= XN) = (xi ,x i , . . .x i n )
Y Xi 1
2

Y2 YN (
= (xj 1 ,xj 2

..
.
,

wh
er
e
Y
 SC - Bidirectional hetero AM
,...xj n ) = , )
q : is the weight vector (q1 , q2 , . . . . , qN )

Step 1 Initialize correlation matrix M to null matrix M [0]

Step 2 Compute the correlation matrix M as

For i 1 to N
M M [ qi Transpose Xi ) ( Xi ) end
(

(symbols matrix addition, matrix multiplication, and

scalar multiplication)

Step 3 Read input bipolar pattern A

Step 4 Compute A_M where A_M A M

Step 5 Apply threshold function to A_M to get B'

ie B'
( A_M )
where is defined as (F) = G = g1 , g2, . . . . , gn

Step 6 Output B' is the associated pattern pair

end
41
 SC AM References
5. References : Textbooks

1. "Neural Network, Fuzzy Logic, and Genetic Algorithms - Synthesis and

Applications", by S. Rajasekaran and G.A. Vijayalaksmi Pai, (2005), Prentice Hall,
Chapter 4, page 87-116.

2. "Elements of Artificial Neural Networks", by Kishan Mehrotra, Chilukuri K.

Mohan and Sanjay Ranka, (1996), MIT Press, Chapter 6, page 217-263.

3. "Fundamentals of Neural Networks: Architecture, Algorithms and Applications",

by Laurene V. Fausett, (1993), Prentice Hall, Chapter 3, page 101-152.

4. "Neural Network Design", by Martin T. Hagan, Howard B. Demuth and Mark

Hudson Beale, ( 1996) , PWS Publ. Company, Chapter 13, page 13-1 to 13-37.

5. "An Introduction to Neural Networks", by James A. Anderson, (1997), MIT

Press, Chapter 6-7, page 143-208.

6. Related documents from open source, mainly internet. An exhaustive list is

being prepared for inclusion at a later date.

42