An Exploration of Grid Cells in Machine Learning

The study of grid cells has been one of the most exciting areas of neuroscience over the past decade. In this Brains@Bay meetup, we are focusing on grid cells and how they act as an inspiration for machine learning architectures. We are thrilled to invite a great lineup of speakers that will present their exciting work on grid cells.

Speaker Lineup:
➤ Marcus Lewis, Senior Researcher at Numenta
Title: Grid cell intro + Quickly forming structured memories

➤ James Whittington, Postdoctoral Research Associate at University of Oxford
Title: Generalisation in the hippocampal formation

➤ Kimberly Stachenfeld, Research Scientist at DeepMind
Title: Representation Learning with Grid Cells

The talks will be followed by a discussion panel and Q&A where Prof. Tim Behrens (University of Oxford) will join. More details to follow soon.

We look forward to seeing you there!

---
➤ Marcus Lewis, Grid cell intro + Quickly forming structured memories

Abstract: The first half of this talk aims to build a basic intuition of grid cells, and it will introduce two different ways that scientists think about grid cells: a blank canvas for memory vs. something that encodes information directly. The second half will dive into using grid cells and other cells to quickly form spatial memories. I'll focus on how grid cells might fit into two fundamental operations: learning spatial maps (which is relatively slow), and learning composite maps as arrangements of spatial maps (which is fast). Viewing the system through this lens, I suggest that the hippocampal formation uses memory associations to create graph- or tree-like composite maps, quickly representing environments as arrangements of parts, and I suggest that each grid cell "module" may support an independent mapping system. This suggests it will likely be fruitful to create AI systems that have multiple independent graph- or tree-based memory systems that associate arbitrary information with nodes and arbitrary "displacement" information with edges.

➤ James Whittington, Generalisation in the hippocampal formation
Abstract: I’ll be talking about unifying space and relational memory in the hippocampal formation. In particular I will introduce, the Tolman-Eichenbaum Machine, a model that learns and generalises relational knowledge, while also exhibiting many neural representations and phenomena similar to the hippocampus and entorhinal cortex.

➤ Kimberly Stachenfeld, Representation Learning with Grid Cells
Abstract: I'll present a relational view of grid cells in which grid cells represent geometry over graphs. I'll go over how these cells could support sophisticated reasoning, and the predictions this model makes for grid cells in relational environments.