About counterpropagation neural networks. I only need some advice, not code.

[Kanj]

I'm working on a project which is mostly image classification using counter propagation neural network. I've got the program running but I don't have a clue about values like learning rate, decay, etc. I've tried many values, but the program just doesn't seem to "learn" anything. It's just some kind of a random number generator.

If you have any experience working with counterprop networks please reply. I'll post details if anyone shows interest.

This is my first project and honestly, I haven't read a lot on the topic. I'd also really appreciate it if you could provide a link to some relevant material.

Thanks.

[Peter87]

I don't know about counterpropagation networks but I have used feed-forward networks before. I have a vague memory that 1/(number of weights) could be a good learning rate try. I'm not at all sure about this and maybe the learning rate should be totally different for a counterpropagation network.

As for the decay I don't know. Maybe you can leave it out until you see that it learns something.

[Kanj]

Thanks Peter87, I've got the code working now. I made two major changes in my program before I found it working:-

1> Train using multiple input patterns corresponding to the same output pattern. (I hadn't realised earlier that it was necessary.)

2> I figured that if I'm training on garbled input pattern then maybe I should garble the input in the recall function too. This was the thing that completely turned around the performance.