# Candidate sampling for softmax neural networks in mathematica

Though I've been using mathematica for a few years, I've never really delved into its ML capabilities (except for the occasional clustering). I'm curious if anyone who has some experience using mathematica's Neural Networks has done anything with candidate sampling with a softmax output, and if so, how one might go about doing it.

• I checked all the methods for Predict[] including Method -> "NeuralNetwork" on tasks from the site kaggle.com . The result is always very average. It is possible that we can improve to good by simply using Wolfram's language. But I did not try. – Alex Trounev Nov 28 '18 at 21:38
• @ericmarkmartin It would be useful if you could add some sample data and output to your question, to make it easier to test an answer. Have you tried anything specifically? SoftmaxLayer would be a good place to start :) – Carl Lange Nov 28 '18 at 22:36
• @AlexTrounev I'm not certain that your comment is relevant to the question. The Predict functionality is quite apart from the much lower-level Neural Networks framework. – Carl Lange Nov 28 '18 at 22:37
• @CarlLange Then indicate which level you want to check and give an example. – Alex Trounev Nov 28 '18 at 22:42
• @AlexTrounev I didn't write the question. Sorry for the misunderstanding. – Carl Lange Nov 28 '18 at 22:49