A simple classifier...does it work?

Hi, I've just trained a Pattern Recognition Neural Network, that has to classify if there's or not a subject.

I gave it a target constituted by binary values organized in two columns that I had subsequently transposed to make it compatible with the network.

I also gave the network a dataset, constituted by n raws (signals) and m columns (samples), that I had to transposed for the same reason.

I choose a Hidden Size of 100, a validation checks of 50 and I obtained the results shown in attached figure, as gradient, performance, confusion, ROC and so on...

Can someone tell me if it works well or not?

Thank you, regards

Answer 1

[Himanshu tripathi]

Yes, looking at the results it seems fine !

[wolss]

How can I justify the convergence due to validation checks? I mean.. we have usually the net convergence thanks to the reachment of the minimum gradient. In this case the gradient is not so low, as the performance, but we have anyway very good results, and so.. how can I justify the well working of the net even if I don't have a so good gradient?

Hope I've made myself clear...

Thank you anyway!

[wolss]

I've another question for this answer... This pictures I've loaded on represent the net functioning (patternnet) with 10240 signals (and 512 samples). I've tried to do the same thing with 22528 signals (always with 512 samples for each one) but the net does not work so well.... Unfortunately it reaches at most 80%. How can I explain this?