Our minds rapidly explore vast hypothesis spaces, inferring probable explanations for sparse data in just fractions of a second. So far, the data efficiency, flexibility, speed, robustness, and energy efficiency of human intelligence have yet to be matched by AI systems. How can we narrow these gaps between artificial and natural intelligence? This talk will review progress towards solving these problems drawing on probabilistic computation, an emerging paradigm that integrates generative models, probabilistic inference, and Monte Carlo into the building blocks of software and hardware. Rather than represent complex probability distributions explicitly, via factorization of density functions, probabilistic computing systems represent them implicitly, via generative programs that model the causal processes unfolding in the world and in agents’ minds. Rather than calculate exact probabilities for inference, probabilistic computing systems sample informed guesses, via massively parallel, online inference programs that combine sequential Monte Carlo, stochastic gradient descent, and variational inference with deep learning and symbolic reasoning. I will illustrate these principles using Gen, a new, multi-paradigm AI programming platform developed by my group, that integrates probabilistic, symbolic, and differentiable approaches. Examples will be drawn from models of core aspects of common-sense scene understanding --- inferring peoples’ probable goals from noisy observations of their motion, and perceiving 3D scene structure --- as well as the problem of learning the structure of generative programs from limited data. I will also briefly review behavioral data that supports probabilistic computing models of human intelligence, and progress towards biologically plausible and energy efficient implementations.

Please use the following Zoom link to attend this seminar:

https://mit.zoom.us/j/91438373475