Cohere talk: predictive representations and synaptic consolidation for continual RL
Cohere · youtube · 2026-09-12
A Cohere Labs Open Science Community talk by Raymond Chua (recently PhD at McGill/Mila under Doina Precup and Blake Richards, soon postdoc at Columbia) on continual reinforcement learning.
- Simple Successor Features: predictive state representations inspired by the Successor Representation and hippocampal cognitive maps, improving continual RL while preserving SR's theoretical properties.
- Synaptic consolidation: biologically inspired memory consolidation combined with predictive representations; results show consolidation works best when it stabilizes predictive representations rather than action-value mappings alone, suggesting structured internal models are a more robust substrate for lifelong learning.
The work also offers new hypotheses on how biological systems acquire, adapt, and preserve knowledge.
More from Research
- Amazon FAR studies image tokenizers as visual languages in unified multimodal models — Amazon-FAR · 2026-09-12
- Fly connectome work framed as the inflection point of AGI and techbio — shakoistsLog · 2026-09-12
- Ethan Mollick: METR long-horizon benchmark effectively saturated as agents hit 18-week tasks — emollick · 2026-09-12
- ResNet by hand: a 10-step walkthrough showing how one identity matrix enabled thousand-layer networks — ProfTomYeh · 2026-09-12
- Deep learning lore: BatchNorm (2015) influenced early TPU v2 design; architecture research boom expected — gordic_aleksa · 2026-09-12
- Fly Spa: full fruit fly connectome (166,700 neurons) simulated live in your browser — juanbenet · 2026-09-12