Please note: This PhD seminar will take place in E7 6443 and online.
Nicole Sandra-Yaffa Dumont, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Chris Eliasmith, Jeff Orchard
In this seminar, we propose a novel state embedding method for deep reinforcement learning (RL) that enhances sample efficiency, particularly in navigation tasks. Inspired by neuroscience and cognitive science, our approach leverages grid cell-like codes and compositional vector representations. Grid cells, neurons located in the medial entorhinal cortex, are thought to provide a spatial framework for navigation in the brain.
We utilize grid cell-inspired Fourier features in both navigation and control tasks, alongside a Vector Symbolic Algebra (VSA) to integrate continuous and discrete variables in compositional embeddings. Testing on multi-step navigation tasks, our results demonstrate that these embeddings significantly improve sample efficiency and transfer learning, highlighting the potential of neuroscience-inspired representations in enhancing RL models.
To attend this PhD seminar in person, please go to E7 6443. You can also attend virtually using Zoom.