Spatial subgoal learning in the mouse: behavioral and computational mechanisms
Here we aim to better understand how animals navigate structured environments. The prevailing wisdom is that they can select among two distinct approaches: querying a mental map of the environment or repeating previously successful trajectories to a goal. However, this dichotomy has been built around data from rodents trained to solve mazes, and it is unclear how it applies to more naturalistic scenarios such as self-motivated navigation in open environments with obstacles. In this project, we leveraged instinctive escape behavior in mice to investigate how rodents use a period of exploration to learn about goals and obstacles in an unfamiliar environment. In our most basic assay, mice explore an environment with a shelter and an obstacle for 5-20 minutes and then we present threat stimuli to trigger escapes to shelter. After 5-10 minutes of exploration, mice took inefficient paths to the shelter, often nearly running into the obstacle and then relying on visual and tactile cues to avoid it. Within twenty minutes, however, they spontaneously developed an efficient subgoal strategy, escaping directly to the obstacle edge before heading to the shelter. Mice escaped in this manner even if the obstacle was removed, suggesting that they had memorized a mental map of subgoals. Unlike typical models of map-based planning, however, we found that investigating the obstacle was not important for updating the map. Instead, learning resembled trajectory repetition: mice had to execute `practice runs’ toward an obstacle edge in order to memorize subgoals. To test this hypothesis directly, we developed a closed-loop neural manipulation, interrupting spontaneous practice runs by stimulating premotor cortex. This manipulation successfully prevented subgoal learning, whereas several control manipulations did not. We modelled these results using a panel of reinforcement learning approaches and found that mice behavior is best matched by systems that explore in a non-uniform manner and possess a high-level spatial representation of regions in the arena. We conclude that mice use practice runs to learn useful subgoals and integrate them into a hierarchical cognitive map of their surroundings. These results broaden our understanding of the cognitive toolkit that mammals use to acquire spatial knowledge.
https://discovery.ucl.ac.uk/id/eprint/10169447/7/thesis_shamash.pdf