What companies are using probabilistic programming

2 points by maxwellsdeamons 1 year ago | 3 comments

I know there is already a question with the exact same title. But it's from 2018. I want to follow up on that 6 years later. A lot of people mentioned Stan there as SOTA. Is this still true? What companies are successfully using probabilistic programming? Are there still new startups that focus on PP? What market verticals are they focusing on? I have the feeling a lot of startups are focusing on deep learning, while PP is completely neglected. And tbh, I haven't seen a lot of deep learning frameworks being able to compete with PP in production. That might be a bit of a hot take, happy to be proven wrong.

This is the 2018 post I am talking about: https://news.ycombinator.com/item?id=17220861

extr 1 year ago |

What is PP to you? Just any application of bayesian techniques? You should reframe your question in terms of what kinds of problems bayesian techniques are SOTA for and then explore what companies operate in those spaces. No startup is "focusing on" PP, it's not a field where there is significant R&D overhang relative to industrial applications like deep learning has seen the last few years. Nobody is getting VC dollars to research new posterior sampling methods.

There are some areas where it's USEFUL to apply existing techniques, usually contexts where it's useful to have a confidence measurement attached to the prediction. Think big decisions based on small data (not always though). Investing, pharma, etc. STAN is SOTA in many ways but if you're interested in higher throughput inference Variational Inference is the preferred technique and that's best supported by something like PyMC.

maxwellsdeamons 1 year ago | |

I think you actually answered my question. I was not 100% sure if there is no VC money for a company that solves some verticals with STAN. I assume there are still some quant companies out there using some flavor of HMMs. But yes, the question may be phrased a bit poorly. Are people still using probabilistic graphical models without deep learning as their go-to model in startups in some verticals? What verticals would that be? And if there is no VC money for this? What are they actually doing? Some combination of PGMs with deep neural network nodes?

maxwellsdeamons 1 year ago |

But yes, the question may be phrased a bit poorly. Are people still using probabilistic graphical models without deep learning as their go-to model in startups in some verticals? What verticals would that be? And if there is no VC money for this? What are they actually doing? Some combination of PGMs with deep neural network nodes?