A Personal Viewpoint on Probabilistic Programming
summary
Argues that probabilistic programs represent distributions by sample-generating code and that Bayesian conditioning can be phrased as conditioning as program transformation; gives concrete demos (binomial coin inference, synthetic block counting, 3D pose via generative vision) and cites Prakash Mansinghka’s group beating baselines on pose. Describes implemented inference as trace MCMC on execution traces (Metropolis–Hastings over random choices), notes rejection sampling is exact but impractical, continuous observations need added noise, and practical success depends on model structure and conditional independence.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:34Probabilistie programming
- 3:32Example: simple probabilistic Python program
- 6:41The stochastic inference problem
- 8:39Example: inferring bias of a coin
- 21:18Fantasy example: extracting 3D structure from images
- 37:24Perturb and traces
- 44:06MIT-Church GMR+08
- 58:32Conditional independence and traces
- 59:36Sampling is easier than computing probabilities
related talks
01h 03mAn Overview of Probabilistic Programming
31mProbabilistic Programming: The What, Why and How
43mThe Power of Probabilistic Programming
01h 07mProbabilistic Programming and Readable Models
24mExact Bayesian Inference for Loopy Probabilistic Programs Using Generating Functions
20mProbabilistic Programming with Stochastic Probabilities
29mProbabilistic Programming: What It Is and How It Works
59mA Welcome to Causal Probabilistic Programming