Vikash Mansinghka Principal Investigator, MIT Probabilistic Computing Project
What kinds of computations give rise to human intelligence? And how can they be scaled in silicon? A great deal of enthusiasm has been focused on answering these questions by building increasingly large deep learning systems. This talk shows how an alternate scaling route, based on probabilistic programs and spiking probabilistic hardware, integrating the best of deep learning, is being used to make new kinds of generative software models and agents that are engineered — not learned from data! — and guaranteed by design to be safe, assistive, and rational. It will also show evidence that this kind of AI better explains the computations unfolding in our own minds and brains than today’s artificial neural networks, and how we can apply this understanding to deliver computational intelligence that is much easier for us to talk to, teach, and justifiably trust.