Zoom: https://mit.zoom.us/j/99363394196
Email abrower@mit.edu for password

ABSTRACT
Many AI architectures based on probabilistic inference in generative models are not supported by existing AI programming languages. Existing probabilistic programming languages can express rich generative models, but do not support many of the inference techniques needed to solve problems like perceiving 3D objects and scenes, inferring others’ probable beliefs and goals, and learning programs from sparse data. This thesis presents Gen, the first general-purpose, high-level probabilistic programming platform. Gen enables users to specify an open universe of models and inference algorithms at a high level of abstraction using a new data type for probabilistic inference that automates implementation details like storing latent states and computing acceptance probabilities, importance weights, and gradients. Gen users can train deep neural networks and neuro-symbolic inference programs that approximate conditional distributions in generative models, and apply them to accelerate model-based Monte Carlo. Users can also employ custom trans-dimensional MCMC and sequential Monte Carlo schemes that bridge between different latent representations of the same domain.  Gen’s black box interfaces enable human users to migrate code to progressively more specialized implementations and extend Gen via new domain-specific languages. Gen’s open-source prototype is being used by an active and growing community for engineering artificial intelligence, reverse-engineering natural intelligence, and teaching probabilistic modeling and inference. The talk will illustrate Gen applied to the problem of inferring the probable goals and future motion of a person. It will also present inference engineering benchmarks showing Gen outperforms less flexible probabilistic programming systems.