SynGAP's brain-inspired method delivers 4x accuracy on continual learning

A team of researchers—Isabelle Aguilar, Zayn Andre Zainal, and Omid Kavehei—has published a new paper on arXiv introducing SynGAP (Synaptic Geometric Adaptive Preconditioning), a continual learning framework designed to solve the stability-plasticity dilemma without relying on task labels. The method is inspired by synaptic metaplasticity, the brain's history-dependent neuromodulation of synapses, which underpins Complementary Learning Systems (CLS) theory. Instead of storing episodic memory or requiring task boundaries, SynGAP maintains an exponential moving average of the Fisher Information Matrix across a continuous data stream. This dynamic metaplastic state is converted into a bounded multiplicative mask that preconditions raw gradients, selectively reducing updates to parameters critical to previously learned knowledge. This results in a task-free, memory-efficient approach that aligns with biological learning.

Empirical results show SynGAP substantially outperforms established baselines. On the Split CIFAR-100 benchmark, SynGAP delivers a 4x increase in accuracy compared to EWC++ and beats Experience Replay (ER) by almost 10%, while reducing the forgetting measure by over 10% against both methods. On the CORe50 benchmark, SynGAP achieves about 68% accuracy, a 10% improvement over standard optimizer baselines. By formalizing continuous biological metaplasticity as stable gradient-based regularization, SynGAP offers a robust, low-memory solution for adaptive intelligence—making it particularly suitable for edge devices and real-time learning applications where storing large replay buffers is impractical.