compgrad.m

MIT的一个工具箱

% Compute gradient relative to weights

for subnet=1:ninputs
   for column=1:ninputs
      grad12{subnet}(:,column) = sum(jback2i{subnet,column},2);
      grad23{subnet}(:,1:nhidden) = grad23{subnet}(:,1:nhidden) + jback3{subnet,column} * jacob2o{subnet,column}';
      grad13{subnet}(:,column) = sum(jback3{subnet,column},2);
      grad34{subnet}(:,1) = grad34{subnet}(:,1) + jback4i{subnet,column} * jacob3{subnet,column}';
      grad45{subnet}(:,1:nextra) = grad45{subnet}(:,1:nextra) + jback5{column}(subnet,:) * jacob4o{subnet,column}';
   end
   grad12{subnet} = grad12{subnet} + back2{subnet} * output1';
   grad12{subnet}(:,1:ninputs) = grad12{subnet}(:,1:ninputs) + wdecayf12 * weight12{subnet}(:,1:ninputs);
   grad23{subnet} = grad23{subnet} + back3{subnet} * output2{subnet}';
   grad23{subnet}(:,1:nhidden) = grad23{subnet}(:,1:nhidden) + wdecayf23 * weight23{subnet}(:,1:nhidden);
   grad13{subnet} = grad13{subnet} + back3{subnet} * output1';
   grad34{subnet} = grad34{subnet} + back4{subnet} * output3{subnet}';
   grad34{subnet}(:,1) = grad34{subnet}(:,1) + wdecayf34 * weight34{subnet}(:,1);
end