Principles for Control When Computation Is Costly
Event Details
- Type
- Center for Studies in Physics and Biology Seminars
- Speaker(s)
-
Xaq Pitkow, Ph.D.(opens in new window), associate professor, Carnegie Mellon University
- Speaker bio(s)
-
Thinking is hard. Sometimes it seems better just to hack a solution than to plan it carefully. Here we develop this idea quantitatively, defining a version of stochastic control that accounts for computational costs of inference. We apply this to Linear Quadratic Gaussian (LQG) control with an added internal cost on information. This creates a trade-off: an agent can obtain more utility overall by sacrificing some task performance, if doing so saves enough mental effort during inference. We discover that the rational strategy that solves the joint inference and control problem goes through phase transitions depending on the task demands, switching from a costly but optimal inference to a family of suboptimal inferences, each interpretable as misestimating the structure of the world. In all cases, the agent moves more to think less. This work provides a foundation for a new type of rational computations that could be used by both brains and machines under strong energy constraints.
- Open to
- Public
- Phone
- (212) 327-8636(opens in new window)
- Sponsor
-
Melanie Lee
(212) 327-8636(opens in new window)
leem@rockefeller.edu(opens in new window)