Navigating by decoding grid cells: Investigating the role of entorhinal grid cells through computational modeling at the intersection of neuroscience and artificial intelligence
Abstract
Ever since grid cells were discovered in the mammalian entorhinal cortex over a decade ago, the striking representation of space generated by these neurons has offered a peek at the inner workings of navigational processes and possibly other high-level cognitive tasks in the brain. The hexagonal, grid-like patterns these neurons produce as the animal travels across space appear intriguingly algorithmic in their nature, raising the question of whether the same general principles could successfully be applied in artificial neural networks, potentially leading to new algorithms in artificial intelligence. Conversely, while grid cells are believed to play an important role in spatial computation and navigation, their specific role in the brain is not fully understood; building artificial navigational agents using the principles of grid cells could help illuminate their role in biology. In this thesis, we investigate the possible role of grid cells in “vector navigation”, where agents navigate by calculating goal vectors using an internal coordinate system. We first develop a neural network able to perform vector navigation by reading out spatial coordinate information from grid cell populations. We then show that the proposed decoding mechanism can work over long distances, and that it can be integrated with other kinds of spatial information known from the hippocampal formation in order to enable the agent to traverse obstacles in complex environments. This demonstrates that grid cells can play the role of a coordinate system for vector navigation within larger navigational architectures, either in artificial (simulated or robotic) settings, or in the brain. We finally show that the proposed decoding mechanism remains functioning in the face of noisy and distorted grid cell signals, which is important for the model to be biologically plausible. Our results could inspire further neuroscientific investigation into grid cells’ potential role in vector navigation. This shows that close interaction between neuroscience and artificial intelligence, e.g. on the topic of neural representations, might lead to valuable insights for both fields.
Has parts
Article A: Edvardsen, Vegard. A Passive Mechanism for Goal-Directed Navigation using Grid Cells. I: Proceedings of the European Conference on Artificial Life 2015. s. 191-198 © 2015 Massachusetts Institute of Technology Published under a Creative Commons Attribution 4.0 International (CC BY 4.0) license https://doi.org/10.7551/978-0-262-33027-5-ch039Article B: Edvardsen, Vegard. Goal-directed navigation based on path integration and decoding of grid cells in an artificial neural network. Natural Computing 2016 https://doi.org/10.1007/s11047-016-9575-0
Article C: Edvardsen, Vegard. Long-range navigation by path integration and decoding of grid cells in a neural network. I: 2017 International Joint Conference on Neural Networks (IJCNN). IEEE, s. 4348-4355 https://doi.org/10.1109/IJCNN.2017.7966406
Article D: Edvardsen, Vegard; Bicanski, Andrej; Burgess, Neil. Navigating with grid and place cells in cluttered environments. Hippocampus 2019 - The final published version is available at https://doi.org/10.1002/hipo.23147 Attribution 4.0 International (CC BY 4.0)
Article E: Edvardsen, Vegard. Navigating with distorted grid cells. I: The 2018 Conference on Artificial Life. MIT Press. © 2018 Massachusetts Institute of Technology Published under a Creative Commons Attribution 4.0 International (CC BY 4.0) license https://doi.org/10.1162/isal_a_00053