Divergence Errors for Arbitrary Feature Expectations

[nquaz]

Hello,

I'm not sure if this is the best place for such a discussion, but I am having issues with divergence and wonder if it stems from how I have formulated my feature set. My feature set consists of random perceptrons ie. sigmoidal functions with weights drawn from a standard normal. This design comes from a paper I am studying on how to compress distributions. I have tried different sizes (n = 5 to n = 500) of this feature set to reconstruct a simple uniform discrete distribution, but I am getting divergence errors in all cases. Do you have any intuition for why my features would be ill-defined? If this is best discussed over email, you can reach me at nnq1@rice.edu.

Thank you