ARNI Emerging Researchers Talk Series #2: Itzel Olivos-Castillo

Bio: Itzel is a Ph.D. student at Rice University working with Prof. Xaq Pitkow. She studies perception and control mechanisms that give biological organisms an advantage over machines. She believes understanding how the brain works using mathematical principles is essential to build the next generation of AI systems which are more robust, more general-purpose, less artificial, and more intelligent. She holds a bachelor’s degree (Telematics Engineering) and master’s degree (Computer Science) from Instituto Politécnico Nacional (IPN-Mexico).

Title: Resource-Efficient Control in Brains and Machines

Abstract:
The brain can turn noisy stimuli into rational behaviors that address a wide variety of tasks using limited
experience, relying on limited processing capacity, and consuming less energy than a lightbulb. What makes the
brain such an efficient control system? Cognitive studies have identified meta-reasoning, the ability to reason
about one’s own reasoning process, as a crucial factor behind this remarkable performance. However, it remains
unclear how meta-level rational agents—whether biological or artificial—successfully balance internal
computation costs against task performance in uncertain environments. To help bridge this gap, we develop a
novel approach to stochastic control where the internal computation cost of inference (a resource-intensive
mechanism that aids in mitigating uncertainty) is optimized alongside task performance. We apply our framework
to quantitatively examine how meta-level rational agents solve Linear Quadratic Gaussian problems. Our findings
reveal that when the estimation error is a meta-control variable the agent can regulate, the dynamics of inference
and control become tightly coupled. This coupling leads to intriguing phase transitions in what is worth
representing, switching from a costly but maximally informative strategy to a family of solutions that differ in
how the agent integrates new evidence, corrects estimation errors, and models the world to lessen the burden of
optimal inference. The fundamental principles we found generalize efficient coding ideas, extend the principle of
minimal intervention in control, and provide a foundation for a new type of rational behavior that both brains and
machines could use for effective but computationally constrained control.

Zoom Link: https://columbiauniversity.zoom.us/j/91436346202?pwd=Fa0ohRBhckitrJqVF5gWrUPo5774U2.1