RL-dreams

With my PhD advisors Cecilia Diniz Behn and Samy Wu Fung, we explored how replay in REM sleep can help us build more aligned RL systems. Specifically, we built a model-based RL system with a mixture-of-experts transformer model as the world model, and disrupted the gating function of the WM in order to generate synthetic data. I had the chance to present on this work at NAISys 2022.