1- The Competitive network Design
The hamming network one of a competitive network.
First layer:
Function Type: purelinear feed forward network
Wight and bias represented as a matrix
Wight of the first layer are the transpose of the prototype vectors
p1p2p Q
weight matrix :
And bias matrix:
output of the first layer is given by:
Second Layer:
a recurrent layer or (the competitive layer) with positive linear transfer function
The second layer is initialized with the output of the first layer.
A 2(0) = a 1
a2 (t + 1) poslin W 2a 2 = ( (t))
The neuron with the largest initial condition will win the competiton
ai2 t + 1poslin ( ai2 taj2 t )
i j
Algorithm and Pseudocode
Initialize weights
Find the winning weight vector which is closest to presented Input vector
Modify the winner so as to move closer to the input vector
Modify weights so as to make them more similar to the values in the input vector.
Pseudocode:
Initialize weights
{wi}, i = 1, 2, ..., k
select a pattern X
Find the winner neuron:
b = argmink| x – wi |
Update the winner neuron
∆wb = γ(x − wb)
Goto step 2 until no significant change in weights.
3- using the competitive network to distinguish between the following two patterns
p1=[1 -1 1 -1] , p2 = [1 1 -1 -1]
W1= , b =
a11 = purelin (w1p1+b1) a11= purelin ( + )
a11 = purelin ( )
a12 = purelin (w1p2+b1) a12 = purelin ( + )
a12 = purelin ( )
w2 =
a2 = poslin (w2 a1) a21 = poslin ( )
a21 = poslin ( )
a22 = poslin ( )
a22 = poslin ( )
| 
