Kseniia Shilova, QBioS Thesis Proposal

Quantitative Biosciences Thesis Proposal

Kseniia Shilova

School of Mathematics

Uncertainty-Driven Active Sensing in Mice: Linking Perception, Action, and Cortex

Date: Friday, October 3

Time: 11:00 AM - 1:00PM ET

Location: (in person) Skiles Classroom Building, Room 114, or virtually in Zoom (https://gatech.zoom.us/j/94340349536 )

Open to the community

Advisor:

Dr. Hannah Choi (School of Mathematics)

Committee Members:

Dr. Audrey Sederberg (School of Psychology, School of Physics)

Dr. Chris Rodgers (Emory University School of Medicine, Department of Neurosurgery)

Dr. Simon Sponberg (School of Biological Sciences, School of Physics)

Abstract:

Natural behavior is active: animals move to reduce sensory uncertainty, yet it is unclear whether uncertainty minimization alone explains action choice or whether behavior reflects mixtures of intrinsic and extrinsic objectives, and how these uncertainty computations map onto specific cortical populations. This proposal will (1) build a probabilistic perception model to compute per-step actual and expected information gain and test their relationship to behavior and cortical signals, (2) train short-horizon, uncertainty-minimizing agents and evaluate whether their behavioral trajectories reproduce identifiable segments of mouse behavior, and (3) place the agent and the animal in a shared discrete-choice space to infer time-varying mixtures of objectives. Empirically, I use two mouse datasets: a freely moving vision dataset with egocentric video, eye/head tracking, and V1 electrophysiology, and a head-fixed active-touch dataset with high-speed whisker videography and SSp/MOp recordings. Model-derived variables (priors, posteriors, and prediction-error-like signals) will test population-specific encoding across the cortex. Together, these aims identify when uncertainty reduction suffices, when additional drives are required, and how cortical circuits represent these computations during natural behavior.