Pusℍ: Concurrent Probabilistic Programming with Function Spaces
share
We introduce a prototype probabilistic programming language (PPL) called Pusℍ for performing Bayesian inference on function spaces with a focus on Bayesian deep learning (BDL). We describe the core abstraction of Pusℍ based on particles that links models, specified as neural networks (NNs), with inference, specified as procedures on particles using a programming model inspired by message passing. Finally, we test Pusℍ on a variety of models and datasets used in scientific machine learning (SciML), a domain with natural function space inference problems, and we evaluate scaling of Pusℍ on single-node multi-GPU devices. Thus we explore the combination of probabilistic programming, NNs, and concurrency in the context of Bayesian inference on function spaces. The code can be found at https://github.com/lbai-lab/PusH.
READ FULL TEXT
