State Space Model Programming in Turing.jl
summary
Two Julia packages, SSMProblems.jl and GeneralisedFilters.jl, provide a composable SSM API inside the Turing.jl ecosystem that implements Kalman, particle and Rao‑Blackwellised particle filters and hooks into particle‑marginal MH and particle Gibbs. Benchmarks show a ~5× speedup by swapping to StaticArrays and many orders‑of‑magnitude gains from custom CUDA kernels versus CPU; forward‑mode AD via ForwardDiff.jl is supported, multi‑target tracking and modular resampling are implemented and the system is production‑ready.
AI-written summary · about
chapters (16)
- 0:00State Space Models in Turing.jl
- 1:03What are State Space Models
- 2:28Why standard inference fails
- 3:30Rao-Blackwellized SSM explained
- 4:54General-purpose RBPF
- 7:03Optimizing linear algebra with static arrays and GPUs
- 8:50Performance benchmarks
- 9:55Modular resampling and custom algorithms
- 10:37Automatic differentiation for particle filters
- 11:42Multi-target tracking extensions
- 12:44Turing.jl and PMMH integration
- 13:05Q&A
- 14:35Q&A
- 15:59Q&A
- 16:19Q&A
- 17:44Closing remarks
related talks
46mState-Space Models and Deep Learning: Is there a new revolution on our doorstep?
36mProbabilistic Programs Which Make (Common) Sense
24mSoss - Lightweight Probabilistic Programming with No Performance Ceiling
32mhow compiler smarts can help improve the performance of numerical methods
27mOpen Source Component-Based Modeling with ModelingToolkit
28mFunctional Sparse Tensor Compilation
43mJulia: A Fresh Approach to Machine Learning
47mThe State of the Art for Probabilistic Programming