encog-java-core icon indicating copy to clipboard operation
encog-java-core copied to clipboard

Published 20 hours ago • verified publisher icon jeffheaton

Backpropagation comments

Open JohnVorwald opened this issue 10 years ago • 0 comments

Consider reformulating the formuation, and coding, including threshold functions, using matrix expressions, instead of all the low level loops. Then the output error function can be more general, depending on more than just the single neuron output. Also, the coding is quite a bit simpler.

For example, feed forward is given by for(iLayer:layers) iLayer.linearOutput = iLayer.weights * iLayer.inputs + iLayer.bias; iLayer.squashedOutput = iLayer.threshold(iLayer.linearOutput); end

JohnVorwald avatar Mar 03 '14 14:03 JohnVorwald