Flatiron Institute community members are cordially invited to a CCN Seminar with Adrien Peyrache, Associate Professor, Montreal Neurological Institute and Hospital at McGill University.
To schedule a meeting with Adrien during his visit please be in touch with Jessica Hauser at firstname.lastname@example.org
Talk title: Cognitive maps: origin and function
The mammalian hippocampus contains a “cognitive map”, which is a set of representations of the environment whose structure reflects the topology of space. It can be used to support navigation and generate offline simulations for the purposes of recall, imagination, planning, and the consolidation of memories. Yet, what exactly constitutes cognitive maps and how they form is still debated. Here, we report that while spatially-tuned cells reliably emerge in recurrent neural networks trained to predict sensory inputs, the presence of spatially tuned cells is not sufficient to guarantee the presence of a cognitive map. Instead, the emergence of a cognitive map requires the use of recurrent connections to predict multi-step sequence predictions. Once learned, this map can autonomously simulate plausible trajectories offline, which can be queried by internal signals and biased towards recently experienced locations, effectively “replaying” those experiences. We develop a predictive RNN architecture, inspired by hippocampal theta sequences, which can rapidly learn cognitive maps. Together, these results provide a unifying model for hippocampal functions and a platform for hippocampal-inspired AI.