We're interested in modular intelligence inspired by brain regions with complementary functions and timescales: left and right hemispheres, fast and slow learning in hippocampus and neocortex.
Hippocampal replay improves world knowledge separately from policy learning. Adding a distribution-matching replay buffer to a latent world model prevents catastrophic forgetting in continual RL, at far lower compute.
A fast, episodic short-term memory learns to steer a slow, generalising long-term memory, so new inputs are described in terms it already knows. It resists forgetting better than a ResNet baseline in continual learning, and generalises better in few-shot learning than other continual learners.
We supervise Masters and PhD projects with Monash and UNSW, and publish open Requests for Research for anyone to take on.
