NSF Award
2311983
The goal of this project is to make powerful Bayesian models and inference algorithms more usable, accessible, and reliable in challenging data science problems. Bayesian inference provides a principled approach to learning probabilistic models by combining prior modeling assumptions with observed data. It enables state-of-the-art results in problems from diverse areas including biostatistics, robotics, computational physics, quantitative finance, cognitive science, and machine learning. Advantages of Bayesian inference include the ability to incorporate prior domain-specific knowledge, to quantify uncertainty about parameters and predictions, and to generalize well to novel data. A key challenge, however, is correctly implementing and diagnosing Bayesian inference algorithms, especially those that target sophisticated probabilistic models. The project's novelty is to address this challenge by developing rigorous programming-language techniques that make sound and effective Bayesian inference more easily applicable. The project's impact is to boost the development and exploration of more flexible Bayesian methods among researchers and help domain experts more reliably leverage these technologies for real-world problems.<br/><br/>The research plan of the project builds on probabilistic programming languages (PPLs) such as Stan, Gen, and Pyro, which provide interfaces that cleanly separate model development from the specification of the corresponding inference algorithm. To make Bayesian learning feasible for more flexible models and larger data sets, several PPLs have enabled users to write custom probabilistic inference algorithms through "programmable inference" interfaces that automate many complex computations needed to develop effective inference algorithms. However, it is easily possible for users to accidentally write incorrect inference programs in such a way that breaks convergence and leads to unsound results. Even worse, such mistakes often go unnoticed. The research in this project aims to alleviate the fundamental tension between soundness and flexibility of programmable inference by (1) applying new programming-language techniques such as static analysis and type systems to verify whether a user-written inference program satisfies theoretical conditions for soundness; and (2) developing new dynamic statistical program analyses to empirically assess the quality of approximate posterior samples produced from the sound inference program. In this way, the system ensures that approximate inference algorithms are not only soundly implemented but are also effective for a given problem in practice. The practicality of the developed techniques is validated through evaluations on challenging data science problems. Moreover, the research results are integrated in the graduate and undergraduate education at Carnegie Mellon University.<br/><br/>This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
