Scott Linderman

icon indicating a website link http://slinderman.web.stanford.edu

icon company Stanford University icon location Stanford, CA icon location I'm an assistant professor at Stanford University in the Statistics Department and the Wu Tsai Neurosciences Institute.

Results 42 comments of Scott Linderman

add support for HMM inference with sparse transition matrices

+1. Lots of models have constrained transition matrices (e.g. left-to-right HMMs and change-point models).

add support for HMM inference with sparse transition matrices

Ah, there's a slight confusion then. One issue is a bug that the current message passing code is returning NaNs in some cases, as @emdupre showed above. Another is a...

Parallel hmm posterior sample

Thanks @calebweinreb! To clarify, I would say that the associative operator takes in two sets of samples, $$z_s \sim p(z_s \mid x_{1:s}, z_{s+1}) $$ and $$z_t \sim p(z_t \mid x_{1:t},...

Add spectral methods for HMM learning

I didn't realize TensorLy worked on JAX — cool! I am most familiar with the tensor decomposition approach described in the paper above, but to be honest I need to...

implement basic particle filtering examples (possibly using blackjax)

Let's loop in my lab on this one — Dieterich Lawson and Andy Warrington have written a lot of particle filtering code recently. I think we'll have lots of stuff...

Add support for parallel hmm sampler

@AdrienCorenflos, I think @calebweinreb has actually come up with a clever way to implement the HMM (and LGSSM) sampling with `associative_scan`! Curious what you think of his implementation in this...

Add support for parallel hmm sampler

I think the other necessary ingredient to the proof is that the function composition $E_{su} = E_{st} \circ E_{tu}$ defined by $E_{su}(i) = E_{st}(E_{tu}(i))$ ensures that $E_{su}(i) \mapsto z_s \sim...

Make HMM learning work with variable length time series

Just commenting here to note that this request (or variants of it) has come up multiple times in the past few weeks. A simple change would be to make the...

Make HMM learning work with variable length time series

Sure, that would be great Alex!

Code fails to raise error when user specifies a non-positive real number for the first parameter `b`

Good catch! I think you can add an assert in [`pgdrawv`](https://github.com/slinderman/pypolyagamma/blob/master/pypolyagamma/pypolyagamma.pyx#L28) before the C code gets called.