This document outlines topics on error backpropagation training algorithms, Kohonen self-organizing maps, and Hopfield neural networks. It then lists several applications of artificial neural networks, including statistical pattern recognition, control of robotics and industrial processes, automatic synthesis of digital systems, adaptive telecommunications, image compression, radar classification, optimization problems, sentence understanding, and applying expertise to conceptual domains.

1. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications of ANN
Sanjay Shitole
Department of Information Technology
Usha Mittal Institute of Technology for Women
SNDT Women’s University, Santacruz(w), Mumbai.
14 Oct 2011
Sanjay Shitole Applications of ANN

2. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Outline of Topics
1 Error Backpropagation Training Algorithm
2 Kohonen Self Organizing Map
Applications
Devanagari Character Recognition
3 Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Sanjay Shitole Applications of ANN

3. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Inputs
( Fixed
Input)
Layer Layer of neuronsk
z
v
v
v v
v
v
z
z1
i
zi
i−1
=−1
11
j1
1i ji
1i
vj−1,i
j1
jiv
y
1
j
j−1
j
=−1
y
y
y
wj
w
w
w
w
ww
w
w
11
1j
1J
K1
Kj
KJ
kJ
1J
j of neurons
−th column
−th column
−th column
of nodes of nodes
of nodes
i j
k
Dummy
neurons
(Fixed Input)
o
o
o
1
k
K
Sanjay Shitole Applications of ANN

4. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Detection of Lung Cancer
Introduction
Sanjay Shitole Applications of ANN

5. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Detection of Lung Cancer
Introduction
Current Medical Techniques
Sanjay Shitole Applications of ANN

6. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Figure: Block Diagram
Sanjay Shitole Applications of ANN

7. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Sanjay Shitole Applications of ANN

8. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Sanjay Shitole Applications of ANN

9. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Inputs
( Fixed
Input)
Layer Layer of neuronsk
z
v
v
v v
v
v
z
z1
i
zi
i−1
=−1
11
j1
1i ji
1i
vj−1,i
j1
jiv
y
1
j
j−1
j
=−1
y
y
y
wj
w
w
w
w
ww
w
w
11
1j
1J
K1
Kj
KJ
kJ
1J
j of neurons
−th column
−th column
−th column
of nodes of nodes
of nodes
i j
k
Dummy
neurons
(Fixed Input)
o
o
o
1
k
K
Sanjay Shitole Applications of ANN

10. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Given are P training pairs written as ((z1, d1), (z2, d2), . . . , (zp, dp))
, Where z1 and d1 are as explained in above section, zi is (I × 1),
di is (K × 1), and i = 1, 2, . . . , P. Note that the Ith component of
each zi is of value −1 since input vectors have been augmented.
Size J − 1 of the hidden layer having outputs y is selected. Note
that the Jth component of y is of value −1, since hidden layer
outputs have also been augmented; y is (J × 1) and o is (K × 1).
Sanjay Shitole Applications of ANN

11. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
η > 0, Emax chosen. This value of η is used in next equations
used for weight adjustments.
Weights W and V are initialized at small random values; W is
(K × J), V is (J × I).
q ← 1, p ← 1, E ← 1
Training step starts here. Input is presented and the layer’s
outputs computed:
z ← zp, d ← zp
Where vj , a column vector, is the jth row of V.
ok ← f (wt
k y), fork=1,2,...,K
Where wk, a column vector, is the kth row of W.
Sanjay Shitole Applications of ANN

12. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Error value is computed:
E ←
1
2
(dk − ok)2
+ E fork = 1, 2, . . . , K
Error signal vectors δo and δy of both layers are computed.
Vectors δo is (K × 1), is (J × 1).
The error signal terms of the output layer in this step are
δok =
1
2
(dk − ok)(1 − o2
k ), fork = 1, 2, . . . , K
The error signal terms of the hidden layer in this step are
δyj =
1
2
(1 − o2
yj
)
K
k=1
δokwkj , forj = 1, 2, . . . , J
The steps of this algorithm are
Sanjay Shitole Applications of ANN

13. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Output layer weights are adjusted:
wkj ← wkj + ηδokyj , fork = 1, 2, . . . , Kandj = 1, 2, . . . , J
Hidden layer weights are adjusted:
vji ← vji + ηδyj zi , forj = 1, 2, . . . , Jandi = 1, 2, . . . , I
If p < P then p ← p + 1, q ← q + 1 , and go to step 2;
Otherwise, go to step 8.
The training cycle is completed. For E < Emax terminate the
training session. Output weights W,V,q, and E.
If E > Emax, then E ← 0, p ← 1, and initiate the new
training cycle by going to Step 2.
Sanjay Shitole Applications of ANN

14. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Sanjay Shitole Applications of ANN

15. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Sanjay Shitole Applications of ANN

16. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Sanjay Shitole Applications of ANN

17. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Sanjay Shitole Applications of ANN

18. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Sanjay Shitole Applications of ANN

19. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Image compression.
Sanjay Shitole Applications of ANN

20. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Image compression.
Radar classification of sea-ice.
Sanjay Shitole Applications of ANN

21. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Image compression.
Radar classification of sea-ice.
Optimization problems.
Sanjay Shitole Applications of ANN

22. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Image compression.
Radar classification of sea-ice.
Optimization problems.
Sentence understanding.
Sanjay Shitole Applications of ANN

23. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Image compression.
Radar classification of sea-ice.
Optimization problems.
Sentence understanding.
Application of expertise in conceptual domain.
Sanjay Shitole Applications of ANN

24. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Applications
Statistical pattern recognition, especially recognition of
speech.
Control of robot arms, and other problems in robotics.
Control of industrial process, especially diffusion processes in
the production of semiconductor substrates.
Automatic synthesis of digital systems.
Adaptive devices for various telecommunications tasks.
Image compression.
Radar classification of sea-ice.
Optimization problems.
Sentence understanding.
Application of expertise in conceptual domain.
Classification of insect courtship songs.
Sanjay Shitole Applications of ANN

25. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Devanagari Character Recognition
Sanjay Shitole Applications of ANN

26. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Sanjay Shitole Applications of ANN

27. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Sanjay Shitole Applications of ANN

28. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Figure: SOM Grid
Sanjay Shitole Applications of ANN

29. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
SOM Algorithm
Intialize the weights Wij (1 < i ≤ 64, 1 < j < m) to small
random values,where m is the total number of nodes in the
map.Set the initial radius of the neighborhood around node j
as Nj (t).
Present inputs X1(t), X2(t), X3(t), . . . , X64(t).
Calculate the distance dj between the inputs and node j by
dj = 64
i=1 (Xi (t) − Wij (t))2
.
Determine j∗which minimizes dj .
Update weights for j∗ and its neighbors mNj (t), the new
weights for j in Nj∗(t) are
Wij (t + 1) = Wij + α(t) (Xi (t) − Wij (t))
Where α(t) and Nj∗(t) are controlled so as to decrease in t.
If process reaches the maximum number of iterations,stop
otherwise go to step 2.Sanjay Shitole Applications of ANN

30. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Devanagari Character Recognition
Output nodes Nc Cycles Training time Classification accuracy
125 60 500 10hrs 65
150 75 500 11hrs 67
175 80 750 13hrs 70
200 99 750 15hrs 75
225 120 900 18hrs 77
250 125 900 23hrs 88
275 130 1000 24hrs 90
300 145 1000 25hrs 91
Sanjay Shitole Applications of ANN

31. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Content addressable memory: which involves the recall of
stored pattern by presenting a partial or distorted version of it
to the memory.
Sanjay Shitole Applications of ANN

32. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Content addressable memory: which involves the recall of
stored pattern by presenting a partial or distorted version of it
to the memory.
Combinatorial optimization problems: The class of
optimization problems includes the Traveling salesman
problem
Sanjay Shitole Applications of ANN

33. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Figure: Hopfield NetworkSanjay Shitole Applications of ANN

34. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
The total input neti of the ith neuron as
neti =
n
j=1
j=i
wij vj + ii − Ti i = 1, 2, ..., n
The external input to the ith neuron has been denoted here as ii .
Introducing the vector notation for synaptic weights and neuron
output, the total input neti of the ith neuron can be written as
neti = wt
i v + ii − Ti , for i = 1, 2, · · · , n
Sanjay Shitole Applications of ANN

35. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
The complete matrix description of the linear portion of the system
shown in Figure is given by net = Wv + i − T where
net
∆
=





net1
net2
...
netn





i
∆
=





i1
i2
...
in





t
∆
=





T1
T2
...
Tn





Sanjay Shitole Applications of ANN

36. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Matrix W , sometimes called the connectivity matrix, is an n ∗ n
matrix containing network weights arranged in rows of vectors
equal to wj as defined and it is equal to
W =





wt
1
wt
2
...
wt
n





W =





0 w12 w13 · · · w1n
w21 0 w23 · · · w2n
...
...
...
...
...
wn1 wn2 · · · wn3 0





Sanjay Shitole Applications of ANN

37. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Responce or Update rule of the ithneuron excited in
net = Wv + i − T
vi → −1 if neti < 0
vi → +1 if neti > 0
The update algorithm for a discrete-time recurrent network and we
can obtain the following update rule:
vk+1
i = sgn(wt
i vk
+ii −Ti ), for i = 1, 2, ..., n and k = 0, 1, ...
Where k denotes the index of recursive update.
Sanjay Shitole Applications of ANN

38. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Energy
The scalar-valued energy function for the discussed system is a
quadratic form and has the matrix form
E
∆
= −
1
2
vt
Wv − it
v + tt
v
Sanjay Shitole Applications of ANN

39. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Energy
Let us study the changes of the energy function for the system
which is allowed to update. Assume that the output node i has
been updated at the kth instant so that vk+−1
i − vk
i = ∆v,. Since
only the single neuron computes, the scheme is one of
asynchronous updates. Let us determine the related energy
increment in this case. Computing the energy gradient vector,
E = −
1
2
(W t
+ W )vt
− it
+ tt
Sanjay Shitole Applications of ANN

40. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Energy
which reduces for symmetrical matrix W for which W t = W to
the form
E = −Wv − it
+ tt
The energy increment becomes equal:
∆E = ( E)t
∆v
Since only the ith output is updated.
Sanjay Shitole Applications of ANN

41. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Energy
∆v
∆
=








0
...
∆vi
...
0








and the energy increment reduces to the form
∆E = (−wt
i v + it
i + ti )∆vi
This can be rewritten as:
∆E =


n
j=1
wij vj + ii + ti

 ∆vi
for j = i Sanjay Shitole Applications of ANN

42. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
To train the Hopfield Network say for x = [0, 1, 0, 1]
Step1: Convert [0, 1, 0, 1] to bipolar. This results in
x1 = [-1,1,-1,1].
Step2:Calculate the Transpose of x1 say y1
Step3:Multiply x1 and y1
Step4:Replace diagonal elements by 0
The final weight matrix is
.




0 −1 1 −1
−1 0 −1 1
1 −1 0 −1
−1 1 −1 0




.
Sanjay Shitole Applications of ANN

43. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
To Train the Hopfield Network for more number of patterns, all the
matrices created for each pattern should be added to get the final
weight matrix.
Sanjay Shitole Applications of ANN

44. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
1 Storage capacity scales
linearly with size N of the
network
2 Storage capacity must be
maintained small for the
fundamental memories to
be recoverable.
Mmax =
N
2logeN
200 400 600 800 1000
0
0
20
40
60
80
100
Without Errors
With Errors
Sanjay Shitole Applications of ANN

45. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Sanjay Shitole Applications of ANN

46. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Thank you
Sanjay Shitole Applications of ANN

47. Outline
Error Backpropagation Training Algorithm
Kohonen Self Organizing Map
Hopfield Neural Network
Applications
Architecture
Mathematical Foundation and algorithm
Deriving weight matrix
Storage Capacity
Case Study
Doubts???
email: shitoless@rediffmail.com
Sanjay Shitole Applications of ANN