NeuroAI
Learning from the neuroscience of human consciousness to build safer, more capable AI.
We study conscious access, predictive coding, memory consolidation, and neuromodulation — and translate these principles into engineering strategies for continuous learning, compositional representations, long‑term credit assignment, and energy efficiency.
Replay + consolidation to learn new tasks without forgetting (sleep‑inspired, hippocampal‑cortical loops).
Workspace‑style broadcasting and language‑like codes enable flexible reuse and rapid recombination.
Eligibility traces + neuromodulators propagate credit over long temporal spans.
Sparse routing, attention gating, and local learning reduce compute and power.
Research programs
Global Workspace for Agents
Inspired by Global Workspace Theory: specialized modules (vision, memory, language, action) compete for access to a broadcast “workspace”. We implement differentiable routers that dynamically gate modules and synchronize latent states — improving sample efficiency and zero‑shot generalization.
Replay & Long‑Horizon Credit Assignment
Combining hippocampal‑style replay with eligibility traces and modulatory signals allows credit assignment across long delays. We evaluate improvements to continual RL and sequence modeling with realistic task gaps and delayed outcomes.
Current prototypes
A controller routes tokens into expert modules (tools, memory, vision) and broadcasts summaries back into context.
Task‑aware replay schedules that interleave practice with off‑policy consolidation to prevent forgetting.
Attention‑based binding and key‑value routing to build language‑like compositional structure in latent codes.
Meta‑gradients shape local plasticity rules, approximating dopamine/acetylcholine signals for rapid adaptation.
Dynamic sparsity and activity regularization to reduce FLOPs while maintaining accuracy under tight budgets.
Self‑supervised predictive coding objectives to build compact, reusable situation models across tasks.
Collaborate with us
We’re building agents that learn continuously, compose knowledge, and reason over long horizons — guided by insights from human neuroscience. If you’re excited about the science and the engineering, let’s talk.
Contact OLAB