Cerebellar Modeling and Learning

Over the past three centuries, humankind has learned a remarkable amount about the chemistry, physics, and biology of the universe. For example, most of the complex workings of the human body are known even down to the molecular level. The main missing piece from this impressive body of knowledge is the brain, whose inner workings remain one of the most notable scientific frontiers. In particular, understanding the nature of intelligence, and especially the nuts and bolts of how it can be realized physically, is one of the main grand challenges of the the 21st century.

For the past few decades, the fields of Neuroscience and Artificial Intelligence (AI) have been making steady progress towards meeting this challenge, but from completely opposite directions. Neuroscience begins from the best physical manifestation of intelligence we have, the human brain, and tries to understand how the nuts and bolts within it interact to produce intelligence. AI, on the other hand, starts with a problem to solve (one that requires an intelligent agent) and attempts to create a structure of interacting nuts and bolts that solves this problem.

While the drive of these fields is in opposite directions, their underlying questions are similar. Nonetheless, due in large part to the large gaps in biological knowledge, their research and communities have been almost entirely independent.

This project is predicated on the claim that for at least one major part of the brain, the cerebellum, our knowledge of the nuts and bolts has recently advanced sufficiently to allow us to bridge the gap. We therefore propose to combine complementary strengths from the fields of Neuroscience and AI: 1) the power of biologically constrained computer simulations of the cerebellum for modeling control tasks; and 2) the vast array ofcontrol tasks tasks that have been successfully addressed by AI algorithms.

With this motivation in mind, the two key questions that this project seeks to address are:

  1. How can our experience with AI solutions to complex tasks help us understand the nuts and bolts of the human cerebellum?
  2. How can our knowledge of the nuts and bolts of the cerebellum help us tackle novel AI tasks where other algorithms have failed?