amazaspshumik / mixture-models Goto Github PK

I would like to be able to sample from a fitted VBGMMARD. For this I would need the weights, means and covariances of the fitted model. The first 2 are available but I can't find the covar. I think it can be derived from the scale, but I don't know exactly how.

Returns probabilities instead of responses

Since optimization is non convex we can end up in local maximum. So to find best fit we may need to rerun algorithm several times therefore we need a method to compare goodness of fit of two models.

diagflat operation is has awful performance when applied to large arrays, substitute it in MDA

In deep trees there are cases where there is complete separability of data , in that case softmax regression (version which is not overparameterised) suffers from the same problem as logistic regression.
Should use overparameterised version (slower convergence since non-unique solution for short trees).

Should help with numerical stability and prevent weights going to infinity in deep hierarchical mixture of experts

Add new functionality to weighted gaussian discriminant analysis so that it can be used as expert for classification problems

Do not copy arrays. Save into existing arrays. Use * instead of dot() where possible.

Use likelihood on test set and stop iterations when it starts dropping. This will prevent overfitting in case of deep HMEs