Random Projections for Probabilistic Inference
summary
Combining random parity constraints with variational inference yields a practical project-then-optimize pipeline that converts counting the partition function into a small set of MAP-style optimizations and promises a constant-factor approximation with high probability; it treats parity hashing as a projection, then performs an I-projection into a tractable family (tested with mean-field), is applied to inpainting/denoising and generative-model learning, and empirically reports huge gains (author cites up to 500 orders of magnitude over plain mean‑field) but requires solving nonconvex variational optimizations to near global optimality.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:21Recent progress in Artificial Intelligence
- 1:18Supervised Learning
- 14:27Learning Latent Variable Models
- 20:10Variational Methods and Random Projections
- 27:16How Might One Count?
- 27:56Brute Force Counting
- 30:10Let's Try Something Different ...
- 31:36Making the Intuitive Idea Concrete
- 32:07Random parity constraints
- 36:15Hashing as a random projection
- 38:05Variational Inference with Random Projections
- 45:21Conclusions
related talks
21mProbabilistic Programming with Programmable Variational Inference
31mProbabilistic Programming: The What, Why and How
26mUsing Random Projections to Make Sense of High-Dimensional ...
21mGuaranteed inference for probabilistic programs: a parallelisable, small-step o...
43mThe Power of Probabilistic Programming
37mVariational Inference in Python
20mNonparametric Involutive Markov Chain Monte Carlo
01h 07mProbabilistic Programming and Readable Models