Neural Networks

<p>I am having some trouble with coding a back-propagation neural network.
When I get to the back propagation part, i take the Mean Square Error, and hold that as a constant, against which I multiply the gradient. For the gradient, what should I do? Right now, I approach each weight one at a time, hold everything else constant, increase the weight a little, calculate new MSE, decrease it a little, calc MSE, and then multiply the learning rate by the direction that causes least MSE, and change weight(well, I store it, and change it after all weights in the newtork have been operated on). is this right? I understand optimization, but this is a tricky situation.</p>

<p>Is this part of the project?
What are you trying to do?</p>

<p>no, it is just the neural network implementation. I will explain at school.</p>

<p>um…what’s the point of posting this?</p>

<p>Well, I know there are a few people here who know about this stuff, and I need some help with it.</p>

<p>To compute the deltas for the output units, multiply the error for each node by the gradient vector. To compute the deltas for the hidden units, multiply the gradient for the hiddens by the (output weights x the delta for each output unit). To compute the changes in the weights, multiply the delta vectors by the learning rate and then by either the activation vector of the hidden units (output layer) or the inputs (hidden layer). Then simply add the changes in the weights to the orignal weights.</p>

<p>I hope this helped. :]</p>