Probabilistic Programming and Readable Models
summary
Argues that probabilistic programming produces readable, interpretable models and demonstrates it with real tools and results rather than platitudes. Uses Pyro on PyTorch and variational inference with GPU acceleration to scale classic Bayesian models, moving from tens of thousands to millions of sequences. Builds a multilevel logistic growth model with half a million latent variables that trains in about 10 minutes on GPU, yields a genome‑wide association Manhattan plot highlighting an unexpected nucleocapsid spike signal, and produces reproducible forecasts and insights.
AI-written summary · about
chapters (15)
- 0:00Probabilistic programming intro
- 6:53Early modeling challenges
- 10:45Graphical models to programs
- 15:57What makes a PPL
- 22:52Code organization in PP
- 28:05Deep learning meets PP
- 36:30Applications of probabilistic programming in genetics
- 40:42Tracking SARS-CoV-2 variants
- 52:08Future directions in PP
- 53:09Q&A
- 56:20Q&A: Quantum computing
- 57:24Q&A: Predictive modeling
- 59:37Q&A: Probabilistic programming vs neural networks
- 1:01:24Q&A: Adoption barriers
- 1:04:56Q&A: Variational vs MCMC
related talks
01h 03mAn Overview of Probabilistic Programming
31mProbabilistic Programming: The What, Why and How
43mThe Power of Probabilistic Programming
01h 03mA Personal Viewpoint on Probabilistic Programming
40mProbabilistic Programming for Augmented Intelligence
47mThe State of the Art for Probabilistic Programming
21mProbabilistic Programming with Programmable Variational Inference
43mProbabilistic Programming in the Real World