Motivating Probabilistic Programming
summary
Built proml, a purely functional Scala library that makes model parameters distributions instead of points, using Metropolis–Hastings MCMC to sample posterior weight spaces. Backends swap via typeclasses (Breeze or tensorflow-scala), with typesafe matrix ops and notebooks for inspection. Example: generate 20,000 samples and burn 15,000 to recover slope ≈2 and intercept ≈−2 for linear regression; extends to Bayesian deep nets; currently MH only, plans to add ADVI.
AI-written summary · about
chapters (9)
- 0:00Introduction and background
- 1:15Thought experiment on NZT drug
- 4:55Problem with current models
- 8:53Probabilistic models explained
- 13:15Introducing Primal library
- 17:54Linear regression demo
- 22:34MCMC and Bayesian inference
- 25:28Bayesian deep learning
- 28:02Closing remarks and Q&A
related talks
31mProbabilistic Programming: The What, Why and How
29mProbabilistic Programming: What It Is and How It Works
38mProbabilistic programming and meta-programming in Clojure
01h 03mAn Overview of Probabilistic Programming
47mThe State of the Art for Probabilistic Programming
30mProbabilistic modelling with Scala
01h 03mA Personal Viewpoint on Probabilistic Programming
42mBayes and Bool Walk into a Bar