felixgruen / featurevis Goto Github PK

Hi Felix -

Thanks for providing the toolbox. It's great to have this functionality in MatConvnet. However, I'm having one serious problem. I'm not able to generate good reconstructions. I've tried starting from the zero image as well as the noise image, as suggested in the function description, and have gone up to 10,000 runs with little success at moving the image towards its target. I believe Yosinski generated his activation-maximization results with less than 1000 iterations, so I don't think the number of runs is the issue. I haven't tried messing with the regularization parameters, as I assume you set the defaults to the ones that worked best in practice, and which were established by prior work.

Can you supply an example in which you get a good reconstruction?

Thanks.

-Nick

Hi, thanks for providing nice toolbox.
I have one question about "layer-wise relevance propagation (LRP)" method in the code.
It seems like it was implemented based on the equation [7] in the paper (Samek, W., Binder, A., Montavon, G., Bach, S., & Müller, K. R. (2015). Evaluating the visualization of what a deep neural network has learned. arXiv preprint arXiv:1509.06321.).

In the code, following that dzdy = res(i+1).dzdx ./ (res(i+1).x + eps), dzdy is put into convolution backpropagation, then run res(i).dzdx = dzdx .* res(i).x. But I am wondering how z_ij can be same with res(i).x here. Also, I'm wondering how the conservation principle is maintained.
Actually, I am trying to implement alpha-beta version of LRP.
Could you let me know the intuition of the code about this part in detail?
Thanks.

Hi Felix,
This is a very useful library.
I do have one query which is what does the colors in the heat map denote? What are the colors for maximum, minimum and centre values? It'd be good to have the color index.
Thanks.