Probabilistic and Differentiable Programming in Scientific Simulators
summary
A framework connects existing stochastic simulators to probabilistic engines using a lightweight flatbuffers-based PPX protocol and a Python library (piperop) that intercepts RNG calls, records stack-trace addresses, and exports execution traces. Large C++ particle-physics simulators become probabilistic programs; training runs count in the millions to fit recurrent proposal networks for amortized inference, reducing end-to-end posterior estimation from days to under an hour on HPC. The system supports cross-language, distributed inference, learned surrogate replacements, and an updated DiffSharp backend for differentiable workflows.
AI-written summary · about
chapters (7)
- 0:00Start
- 3:57Inverting a simulator
- 6:20Probabilistic execution
- 20:55Inference results
- 22:25Interpretability
- 26:00Design optimization
- 27:02MIAPP program
related talks
43mThe Power of Probabilistic Programming
31mProbabilistic Programming: The What, Why and How
20mProbabilistic Programming with Stochastic Probabilities
47mThe State of the Art for Probabilistic Programming
01h 07mProbabilistic Programming and Readable Models
01h 03mAn Overview of Probabilistic Programming
40mProbabilistic Programming for Augmented Intelligence
14mMulti-Language Probabilistic Programming