Repositories

Many statistical methods see limited use because their implementations are hard to run, slow, or, worse, nonexistent. I am committed to developing reliable, efficient software, both for domain-specific models that let scientists obtain reproducible results from their own data and for general-purpose sampling methods that can be widely adopted. All of my methods are released as open-source software, with efficient implementations and documentation aimed at practitioners.

AdaptEllipticalSliceSampler.jl implements the adaptive generalized elliptical slice sampler (AGESS), a gradient-free, black-box sampler that performs well on a wide variety of target distributions, including those that are non-differentiable, multimodal, or high-dimensional. Because the package is integrated with Turing.jl through AbstractMCMC.jl, users can write models in general probabilistic programming syntax and use AGESS as a drop-in sampler, without deriving anything specific to their model. The package also supports running multiple chains in parallel, so convergence can be checked with standard multi-chain diagnostics. In addition, I have developed two R packages with computationally intensive routines written in C++: BayesFMMM, for fitting functional mixed membership models, and NeuralComp, for testing whether neurons multiplex when encoding multiple stimuli.