Models and Programs: Better Together
summary
Argues that combining models and programs yields concrete systems: use DNNs plus program synthesis (PROS) to auto-recover parsers for changing HTML and email formats, deployed as a heterogeneous data extraction framework that saves about 50% of data annotation bandwidth; use ML to propose labels and synthesis to produce interpretable runtime programs. Presents a programming-by-rewards (PBR) engine called self-tune that learns decision functions from execution rewards, and PL techniques—weakest preconditions, program slicing, and a rejection-free Metropolis–Hastings sampler—to make probabilistic-program inference practical.
AI-written summary · about
chapters (10)
- 0:00Start
- 3:43Two ways to construct computer systems
- 10:41Programs and Models: Serving Each Other
- 29:26Heterogeneous Data Extraction Framework
- 37:24Programming By Rewards (PBR)
- 39:00Learning with black-box rewards
- 48:39Tutorial on probabilistic programs (1)
- 50:43Tutorial on probabilistic programs (3)
- 54:31Kidney Disease Estimation
- 55:48Probabilistic Inference
related talks
01h 00mBetter Learning through Programming Languages
01h 03mAn Overview of Probabilistic Programming
19mPrimo: Practical Learning-Augmented Systems with Interpretable Models
01h 10mNeurosymbolic Program Synthesis: Bridging Perception and Reasoning in Real-World(…)
01h 07mProbabilistic Programming and Readable Models
43mThe Power of Probabilistic Programming
11mCollaborative Natural Language Inference
32mStructured Deep Learning with Probabilistic Neural Prog