HMM doesn't fit Variational distribution of VAE

[Baileyswu]

HMM doesn't fit Variational distribution of VAE

[Baileyswu]

After sampling data from posterior, we can get newdata which fit the HMM likelihood.

so it's reasonable that hmm have such distributions.

[lucasondel]

The thread is closed but I'd like to add a couple of precisions. As you have realized, in the VAE-HMM model, the HMM does not model the posteriors' mean but the density formed from all the posteriors. Hence, that's why you see the HMM's emissions fit the sampled data and not just the means.

Regarding your other comment, I agree that BEER has evolved to the state where it is difficult to use as there is no real documentation. Furthermore, technical documentation is not enough as the whole design of the software requires some theoretical knowledge about very specific machine learning algorithms. This theoretical background is also not easily assimilated/distributed since it is spread across many publications but, to my knowledge, has not been gather in a single textbook or similar support.

Anyway, I'm glad BEER gathers the implementation of several unsupervised/low-resource speech learning algorithms (thanks again for the VAE part btw) andI hope it will serve as a starting point for other people who work in this field.