Scott Linderman

Make HMM learning work with variable length time series

6

I don't think the current `hmm_fit_sgd` function is using the length of the time series as we hoped. At least with the default [loss function](https://github.com/probml/ssm-jax/blob/main/ssm_jax/hmm/learning.py#L36), the length is just scaling...

Add spectral methods for HMM learning

2

In addition to SGD and EM, we should support spectral learning methods. See for example, > Anandkumar, Animashree, et al. "Tensor decompositions for learning latent variable models." Journal of machine...

The [`_sample_minibatches`](https://github.com/probml/ssm-jax/blob/main/ssm_jax/optimize.py#L12) function produces an generator that, under the hood, mutates an internal state. This doesn't play nicely within `lax.scan`. Instead, it silently fails and returns the first value of...

Now that we've implemented Bernoulli, NB and binomial are very simple extensions. We just need to implement the proper `a_func`, `b_func`, and `c_func` methods to extend the base class. See...

The old versions of pyglm relied on [graphistician](https://github.com/slinderman/graphistician) to support different network models. To reduce dependencies, I've copied the most important network models into pyglm. I still need to copy...

As @tansey pointed out, we could combine these interfaces and make them more familiar to numpy users by making `pgs` (i.e. the vector to be populated with samples upon output)...

This could be a good diagnostic tool.

We should make an example of how to warm start the SLDS with PCA/factor analysis followed by a few Gibbs steps on the discrete states and their dynamics.

The weak limit transitions class does not support expectations necessary for mean field. We should either support those expectations (probably doesn't even make sense for "weak limit" models) or we...

Now that more folks are using pylds, it would be good to go through and clean up the code a bit. - [ ] Fix or make issues for TODO's...