Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Discovering efficient maps for navigating a reconfigurable 3D environment

Tzuhsuan Ma¹, Alison E. Comrie¹, Jakob Voigts¹, Ann Hermundstad¹; ¹HHMI Janelia Research Campus

Presenter: Tzuhsuan Ma

How do biological and artificial agents build internal maps to efficiently and robustly navigate ambiguous environments? Prior work in animals has focused on characterizing spatially-tuned neural activity in simple arenas, while computational approaches often rely on resource-intensive methods that are specialized for specific settings. As a result, little is known about general cognitive algorithms that enable efficient learning and navigation in complex terrains that lack well-defined routes. We study this question in mice that freely explore a 3D maze in complete darkness, without any explicit trials or reinforcements. We enumerate assumptions about cognitive processes that underlie their map formation, and we use this to derive algorithms that can be validated in further experiments. Specifically, we combine a noisy path integrator with sensory localization by tactile perception; by sweeping the parameters of these model components, we infer likely loop closure events when mice correctly recognize previously-visited locations during early exploration of the maze. To validate these assumptions, we combine theoretical and experimental approaches. Theoretically, we simulate a navigation task to test whether loop closure locations are good for constructing compressed maps that support accurate self-localization. We find that these compressed, loop-closure-based maps outperform occupancy-based maps, suggesting that mice could use early loop closure locations to self-localize and guide future navigation. Experimentally, we develop a reconfigurable maze that allows us to rapidly enumerate 3D environments and conduct trial-based experiments with targeted perturbations. Together, this work aims to derive algorithmic principles for efficient online map-building in ambiguous environments. Such principles shed light on cognitive strategies for navigation and may also support the design of efficient algorithms for resource-constrained autonomous robots.

Topic Area: Methods, Tools, Theory & Neural Coding