Probabilistic Programming: The What, Why and How
summary
Probabilistic programming models become programs and the goal is to infer posterior beliefs about parameters from observed data; widely used Stan relies on Hamiltonian Monte Carlo, which forbids discrete parameters because gradients are needed. The claim: translate programs with discrete choices into gradient-friendly form by marginalize discrete variables (sum them out) and then reconstruct them afterward. Implementation uses information flow analysis and type inference to find conditional independences and slice code for efficiency; the translator lets Stan-style HMC run on these models and shows practical speedups.
AI-written summary · about
chapters (20)
- 0:00Intro
- 3:53Monty Hall
- 6:55Scenarios
- 8:07Pseudocode
- 11:38Progressive Programming vs Deep Learning
- 13:20Why Progressive Programming
- 14:30Picture
- 15:15Particle Physics
- 16:20Astronomy
- 17:15Success Stories
- 17:50The Catch
- 19:52The Dream
- 20:23My Research
- 20:57Star
- 21:45Slickstand
- 22:04Change Point Model
- 24:08Gradient
- 24:41Discrete Parameters
- 29:02Key Idea
- 30:23Conclusion
related talks
01h 03mA Personal Viewpoint on Probabilistic Programming
43mThe Power of Probabilistic Programming
20mProbabilistic Programming with Stochastic Probabilities
01h 03mAn Overview of Probabilistic Programming
29mMotivating Probabilistic Programming
01h 07mProbabilistic Programming and Readable Models
32mProbabilistic and Differentiable Programming in Scientific Simulators
29mProbabilistic Programming: What It Is and How It Works