Welcome to Statistics seminar on Wednesday 14th October at 12.15. Miika Kailas (JYU) is presenting his work on Bayesian computation with MIMCS. The lecture will take place at Agora Ag B336.2 Eta.
Anyone interested is warmly welcome!
Abstract
Modular Intelligent Markov Chain Samplers (MIMCS) is a new Python library for general-purpose Bayesian computation with MCMC. It provides several state-of-the-art sampler components such as the No U-Turn Sampler (NUTS), Riemannian Manifold Hamiltonian Monte Carlo (RMHMC) and Within-orbit Adaptive Leafrog integration for HMC, and allows them to be combined in a modular fashion. The user need not be an expert in MCMC samplers, though, as MIMCS also provides heuristics for choosing the appropriate combination of sampler components given the model and available evidence from pilot runs. The computational backend is provided by the JAX library which supports easy access to parallel-computing devices (GPUs) and is compiled to efficient low-level code by the XLA compiler. Models may be specified via a small Stan-like language that compiles into JAX functions, or written directly as JAX functions. Further MIMCS features include parallel tempering and (rudimentary) support for discrete variables, and target-aware convergence diagnostics via Stein discrepancies. MIMCS is currently available via GitHub; an outgrowth of my research code, the library is in early development but is ready for use.
