Probabilistic Programming with Programmable Variational Inference
summary
Proposes a compositional, provably correct route to variational inference that turns user-written model, guide, and objective programs into unbiased gradient estimators via three program transformations; implements the design as genjax.vi on JAX and supplies denotational semantics plus a soundness proof. Evaluations on AIR/multi-MNIST show a combinatorial space of previously unsupported objectives and estimators, with best algorithms 5–10× faster across options and per-epoch implementations about 10x faster than Pyro, reducing AIR training to ~40s on a consumer GPU.
AI-written summary · about
chapters (7)
- 0:00Introduction
- 1:49Variational inference challenges
- 3:56Modular variational inference
- 6:03Variational inference tutorial
- 11:59Evaluating on deep learning
- 16:32Formal contributions summary
- 19:02Q&A
related talks
38mBlack-Box Variational Inference for Probabilistic Programs
37mVariational Inference in Python
51mRandom Projections for Probabilistic Inference
01h 07mProbabilistic Programming and Readable Models
31mProbabilistic Programming: The What, Why and How
23mProbabilistic Programming with Vectorized Programmable Inference
40mProbabilistic Programming for Augmented Intelligence
01h 29mA Beginner's Guide to Variational Inference