Invited talk at the Association for the Advancement of Artificial Intelligence’s Thirty-Second AAAI Conference on Artificial Intelligence (AAAI-18), New Orleans, LA (February 2018).

Probabilistic programming is able to build rich models of systems that combine prior knowledge with the ability to learn from data. One of the reasons for the success of deep learning is the ability to discover hidden features of the domain through complex multi-layered, nonlinear functions; another reason is the ability to learn and reason effectively about these functions in a scalable way. Our goal is to develop generative probabilistic programs that have the same properties.

Recent trends in cognitive science view perception and action in a unified framework based on downward prediction using a generative probabilistic model and upward propagation of errors. Scruff is intended to be a probabilistic programming cognitive architecture based on this idea. Scruff provides many different mechanisms for accomplishing intelligent behavior, all within a neat generative probabilistic framework. Scruff uses Haskell’s rich type system to create a library of models, where each kind of model is able to support certain kinds of inference efficiently. The type system ensures that only compatible models can be linked together. Current mechanisms include learning parameters via gradient ascent backpropagation (as in deep neural networks), reinforcement learning to perform inference, conditioning on various kinds of evidence, and different ways of computing probabilities. Using Scruff, we are exploring a range of new kinds of deep models, such as deep noisy-or networks, deep probabilistic context free grammars, and deep conditional linear Guassian networks.