Why Edward for Gaussian processes?


#1

Hi @dustin,
I really enjoyed your GP summer school talk; the is one of my favorites :slight_smile:. You motivate probabilistic programming for GPs, explaining the how but not really the why. That is, in what scenarios would a prob prog GP be advantageous over a standard implementation – why use Edward over e.g. GPFlow? One advantage I can see is the decoupling of modeling and inference in Edward is a little nicer to work with. Are there differences in computational efficiency? Are there GP models Edward enables that wouldn’t be possible in a deterministic paradigm? Maybe uncertainty propagation easier? Thanks in advance!

Cheers,
Alex


#2

Thanks @BoltzmannBrain for asking. I think the “how” is also the “why”. The slide you links to explains how (why) probabilistic programming in terms of modeling. Each of the points are difficult to do in GPy/GPflow.

There’s another slide on how (why) probabilistic programming in terms of inference. They’re also difficult to do in GPy/GPflow.


#3

Hi, @dustin, is there any case of industry usage of GP or PPL? just new to PPL and it is very powerful, but it seems bit hard to be adapted in production stage, any advices?

PS: I just saw the tensorflow repo got a sub-repo called probability https://github.com/tensorflow/probability, is it the Edward 2.0? or another group doing the same staff?