Hi, David.
In the file trainStochasticSquaredErrorTwoLayerPerceptron.m
, the weight matrices are updated after propagating and back-propagating on each item, this is the standard SGD optimize, while I try to use mini-batch optimize. I modify inputVector = inputValues(:, n(k));
to inputVector = inputValues(:, n(k): n(k) + minBatch );
and change targetVector
to targetVector = targetValues(:, n(k): n(k) + minBatch);
. I hope the weight matrices are updated epochs
times, and this looks like the mini-batch way, However, the experimental results are totally wrong, correct percent is only 10%.
Would you explain why?
Thanks