Is there a classification scenario with cross-entropy loss such that the loss as a function of the predictor/neural net's parameters is a function s.t it satisfies the properties of (a) having a global minima (b) being not convex and (c) its gradient being upperbounded?

The first and the third conditions I wrote above are the basic requirements for a SGD to converge to criticality for non-convex objective.


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.