I wanted to use the visualization of the activation maximization of the filters that is described in the following keras tutorial/blog:


I'd like to know what is the intention behind the decision that filters that produce a loss <= 0 are skipped. I know for 0 that would be reasonable since their would be no gradient flowing then (I think) but what about negative values? And is it also reasonable to use the mean of the outputs of the filters as a loss? What if there are weights of a filter that have high negative and positive numbers. Would that be a problem?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.