%
% ocrtrain.m
%
% train a 2 level feedforward net to recognize the
% members of the DNA alphabet. The letters are bit patterns
% associated with a 5x5 LED array.
%
% usage [V,W] = dnaocr(V0,W0,maxiter,rate)
%
% where		V0 is the 25x25 matrix of weights from input to hidden
%               W0 is the 2x25 matrix of weights from hidden to output
%	        maxiter is the number of training iterations
%		rate is the learning rate
% 
% e.g., V0 = randn(25);  W0 = randn(2,25);
%       [V,W] = dnaocr(V0,W0,5000,0.1)
%

function [V,W] = ocrtrain(V0,W0,maxiter,rate)

V = V0;		% initialize V, 
W = W0;		% initialize W,

state = zeros(4,25);
state(1,[1:5 6 13 18 23]) = 1; 			% T
state(2,[1:5 6 10 11:15 16 20 21 25]) = 1;	% A
state(3,[1:5 6 11 16 21:25]) = 1;		% C 
state(4,[1:5 6 11 14 15 16 20:25]) = 1;		% G 

target = [0 0;	% T	(state T should produce this output)
          0 1;  % A
          1 0;  % C
          1 1]; % G


for iter=1:maxiter,

      j = ceil(rand*4);		% index of input
      i = ceil(rand*2);		% index of output

      p = state(j,:)';	% 25-by-1

      h = s(V*p);		% 25-by-1

      o = s(W(i,:)*h);		% 1-by-1

      tmp = (o-target(j,i))*o*(1-o);    % the common part of the update

      W(i,:) = W(i,:) - rate*tmp*h';	% update row i of W
  
      V = V - rate*tmp*(W(i,:)'.*h.*(1-h))*p';  % update V

end

% the soft theshold

function val = s(x)
val = 1./(1+exp(-x+.5));