Scientific Computation

Bayesian model comparison requires computing intractable high-dimensional integrals. SunBURST exploits GPU parallelism to achieve O(d^0.67) scaling through 1024+ dimensions — a paradigm shift enabling model selection problems previously considered computationally impossible.

See the dedicated SunBURST project page for more details.

ChiSao (Convergence–Halt–Invert–Stick–And–Oscillate, named for the Wing Chun "sticky hands" drill) is a GPU-native population optimizer that runs an entire sample batch at once and rides a deliberate convergence–anticonvergence oscillation to escape local traps while freezing confirmed modes. Samples that reach true peaks are frozen ("stuck"); the rest keep exploring via momentum-based anti-convergence and stochastically smoothed gradients, with adaptive reseeding (Repulse Monkey, Golden Rooster) preserving diversity. Across all 42 Simon Fraser University benchmark functions through d=64 it achieves 100% mode recovery — where CPU baselines (basin-hopping, CMA-ES, multistart) collapse at d≥8 — at up to 34× speedup over basin-hopping and 39× on unimodal functions, and stays fully reliable under likelihood noise up to σ=1.0. A sibling to SunBURST, aimed at optimization rather than Bayesian evidence. Open-source via chisao on PyPI.