NeuroSynth's brain-inspired RL cuts catastrophic forgetting by 54x
Dual-pathway design preserves early-task knowledge where PPO collapses to near zero.
Yash Kini's new paper (arXiv:2607.28663) presents NeuroSynth, a biologically inspired continual reinforcement learning architecture designed to combat catastrophic forgetting—the tendency for AI to overwrite old skills when learning new ones. Drawing from neuroscience, NeuroSynth uses a dual-pathway consolidation mechanism that separates rapid task acquisition ('plan') from long-term retention ('habit'), augmented by replay buffers and knowledge distillation. This mirrors how biological brains combine fast hippocampal encoding with slower cortical consolidation.
Tested on the NeuroMaze-CL benchmark with three sequential navigation tasks, NeuroSynth dramatically outperformed standard PPO. Across six seeds, it preserved 18.00% Task A success versus PPO's 0.33% (p=0.0149, d=1.49) and 35.33% on Task B versus PPO's 0.00% (p=0.0024, d=2.31). It also beat Elastic Weight Consolidation on final Task C performance (9.00% vs 2.00%), though this advantage wasn't statistically significant. The results suggest that biologically plausible consolidation mechanisms can significantly improve the stability-plasticity balance, pointing toward more practical lifelong learning systems for robotics and real-world AI applications.
- NeuroSynth achieves 18% Task A success vs 0.33% for PPO in continual learning (54x improvement)
- Dual-pathway 'plan' and 'habit' architecture mimics hippocampal-cortical consolidation with replay
- Outperforms EWC on final task retention (9% vs 2%) though not statistically significant
Why It Matters
Biological consolidation could unlock lifelong learning AI agents that don't forget, enabling robust real-world robotics and adaptive systems.