Fast Inference for Probabilistic Graphical Models
summary
Fast-PGM is a C++ system for importance-sampling PGM inference that claims 3–20× speedup over prior implementations by fusing CPT and proposal tables, using a compact probability table layout with index math and rightmost-variable reordering, simplifying computation of conditional vectors, and using sample-level parallelism to avoid load imbalance. Experiments show per-optimization gains (memory ≈25%, computation ≈14%, parallelism ≈61%), supports plug-in algorithms, and the source is on GitHub.
AI-written summary · about
chapters (11)
- 0:00Probabilistic graphical models intro
- 2:56Importance sampling challenges
- 4:19Unified importance sampling
- 7:49Fast PGM architecture
- 8:31Memory optimizations in Fast PGM
- 10:38Irregular memory access
- 12:06Computation simplification
- 14:13Parallelism strategies
- 15:40Experimental results
- 16:42Accuracy validation
- 17:05Conclusion
related talks
01h 23mProbabilistic Graphical Models, HMMs using PGMPY
30mHybrid Recommendation Systems using Probabilistic Graphical Models
32mLearning Ill-Conditioned Gaussian Graphical Models
35mScalable structure learning and inference for domain-specific probabilistic prog
23mDynamic inference in probabilistic graphical models
11mExact Inference for Discrete Probabilistic Programs via Generating Functions
01h 44mpomegranate: fast and flexible probabilistic modeling in python
20mGenSQL: A Probabilistic Programming System for Querying Generative Models of Database(…)