Fork me on GitHub
#uncomplicate
<
2017-06-07
>
whilo08:06:00

argh, the backlog is gone 😞

whilo08:06:02

@blueberry I am exploring anglican for a distributed inference model where models talk to each other in p2p fashion.

whilo08:06:34

Anglican supports high-dimensional data, but you are right, the inference methods do not necessarily scale.

whilo08:06:52

I like though that it has a distinction between model specification and inference method.

whilo08:06:07

It is also pretty cool that it is a state of the art probabilistic programming language in clojure.

whilo08:06:32

You can even describe stochastic processes and build fairly complicated models.

whilo08:06:34

They have different inference methods, not just monte carlo ones. In general black box variational inference is more scalable than MCMC and now they also tinker with neural nets in the background to improve the SMC sampler by making the neural net create better proposals.

whilo08:06:08

I am not familiar with the details yet, but I don't think the two approaches of bayadera and anglican exclude each other so far.

whilo08:06:48

Exploiting fast GPU facilities is very much desired for anglican, while some models might not be run on GPUs.

whilo08:06:03

But I might oversimplify some things, I am still not familiar with all the details.