Grant: Discovery Grants Program - Individual

Probabilistic programming languages (PPL) are on the cusp of becoming practically useful for expressing and solving a wide-range of model-based statistical reasoning problems. The high-level hypothesis I propose to test is that continuing PPL research and development will make it possible for the artificial intelligence (AI) community to rapidly develop key new probabilistic models for perception, reasoning, and action selection that go far beyond what current deep learning systems are thought to be capable of now. The foundation of deep learning consists of supervised learning, big data, fast computers, neural net architectures, and differentiation automation software tools. I draw an analogy between what I propose here and the development of such programming language tools for automating differentiation. Such tools arguably led to the deep learning revolution by making it easier for academics and practitioners both to quickly and easily experiment with novel neural net architectures. PPLs aim to play the same role for unsupervised learning and inference. PPLs subsume the denotation of and automate inference in Bayes nets, graphical models, factor graphs, Bayesian nonparametric models, and Bayesian deep learning models. Until very recently this has unfortunately required PPLs to use very general purpose inference algorithms that are capable of “solving” “all” inference problems. The main part of this proposal is to advance so-called “inference compilation,” a very recently introduced technique that bridges between probabilistic programming and deep learning, leveraging deep learning techniques to dramatically speed amortized inference in richly structured PPL models. The general aim is to develop theory and software that allows a) derivation of the structure of so-called inference network (IN) architectures from PPL models (and vice versa) and b) efficient training of such INs so that they perform rapid inference. A second part of this proposal addresses the problem that even writing interpretable, detailed, and accurate generative models is itself very hard so I will also investigate generative model learning within the PPL/inference compilation framework. Languages that automate semi- and un-supervised, generative-model learning with compiled/amortized inference are, in my opinion, the key constituents of the next toolchain for advanced AI. From theory to implementation this proposed research and development aims for the realization of practical and scalable implementations of such languages and demonstrations of how they can be used for next-generation AI applications such as fully-autonomous driving, unrestricted question-answering systems, etc.