P006 Neuromimetic models of mammalian spatial navigation circuits learn to navigate in complex simulated environments
1Haroon Anwar,3Christopher Earl,4Hananel Hazan,1,2Samuel Neymotin
1Center for Biomedical Imaging and Neuromodulation, Nathan Kline Institute for Psychiatric Research, Orangeburg, NY, USA.
2Department of Psychiatry, NYU Grossman School of Medicine, New York, NY, USA.
3Department of Computer Science, University of Massachusetts, Boston, MA, USA.
4Allen Discovery Center, Tufts University, Boston, MA, USA.
Email: haroon.anwar@gmail.com
Introduction
Hippocampal place cells and entorhinal grid cells play a central role in navigation. Grid cells support vector-based navigation relying primarily on internally generated motion related cues like speed and head directions, whereas place cells, mainly driven by external sensory cues, capture relationships among temporal and spatial cognitive variables. Most theoretical models [1-3] capture physiological properties of the grid and place cells but lack learning and spatial navigation functions. In this work, we extend theoretical models to incorporate learning and function. Our aim is to increase understanding of the neural basis of navigation, and use it to improve fully autonomous or hybrid artificial systems with humans in the loop.
Methods
We use integrate-and-fire neuron models to represent head-direction (North, South, East, West), motion-direction (Forward, Backward, Left, Right), landmark, conjunctive, place, and motor neurons. The number of conjunctive cells scales with the number of landmark cells and is adjusted to ensure unique landmark encoding relative to the agent’s orientation. Initially, all conjunctive cells form weak connections to place cells. As the agent navigates, only synapses from activated conjunctive cells to place cells are strengthened, forming place fields. Consequently, synapses from place cells to motor neurons representing rewarding actions are potentiated via reward-based spike-timing dependent plasticity [4], guiding the agent toward its target.
Results
Our modeling results highlight the strengths of place cell-based navigation models in learning complex pathways. While grid cell-based models alone struggle with complex and multi-linear navigation, place cell-based models - integrating inputs from grid circuits - demonstrate superior learning capabilities. The capacity of our place cell-based model to encode diverse places and environments scales with the number of landmark and conjunctive cells included. Additionally, our findings suggest that non-Hebbian synaptic plasticity mechanisms may play a crucial role in the development of place fields, further enhancing navigational learning.
Discussion
Although our place cell-based navigation model successfully learns how to navigate in complex environments, its capacity is limited by the categories of neurons utilized. Such limitations are inherent to our modeling approach, which requires predefining the number of neurons, neuron types, and synaptic plasticity mechanisms. We encountered scaling challenges due to all-to-all weak connections from conjunctive cells to place cells. Once a place field is established, all remaining weak connections to that place cell must be removed to prevent spurious activation outside its designated field. To address these constraints, we plan to incorporate structural plasticity rules in future models to remove excessively weak synaptic connections.
Acknowledgements
Research supported by ARL Cooperative Agreement W911NF-22-2-0139 and ARL/ORAU Fellowships
References
[1] Burak Y, Fiete IR (2009) Accurate path integration in continuous attractor network models of grid cells.PLoS Comp Biol5(2): e1000291.
[2] Giocono LM, Moser M-B, Moser EI (2011) Computational models of grid cells.Neuron71, 589-603.
[3] Bush D, Barry C, Manson D, Burgess N (2015) Using grid cells for navigation.Neuron87, 507-520.
[4] Hasegan D, Deible M, Earl C, D’Onofrio D, Hazan H, Anwar H, Neymotin S (2022) Training spiking neuronal networks to perform motor control using reinforcement and evolutionary learning. Front. Comput Neurosci 2022 16:1017284
