The way to calculate the generalization ability of CNN

I’ve suggested some training method can increase the accuracy of classification with shallow networks. That method includes the training method with added loss function.

Actually, I really wanted to write some SCI paper but I don’t know why the network’s accuracy. I just experimented my suggesting method with my insight and that worked :slight_smile:

I’m in a postgraduate master’s degree course(3rd semester). Because of that, I think I’m a little poor about mathematical things and various things about deep learning.

I really want to find out the causalities about the result of my method. I already tried loss landscape visualization (https://github.com/tomgoldstein/loss-landscape) but it only shows little differences between original classification training method and proposing training method, not the significant ones.

So my question is
Is there any way or paper to calculate the generalization ability of deep neural network?? Can anybody help me or just let me know which paper is related?

Help me friends :slight_smile: