Reproduce “Order parameters and phase transitions of continual learning in deep neural networks” on OpenResearch

About this paper

Haozhe Shan, Qianyi Li, Haim Sompolinsky

A statistical-mechanics theory for continual learning in deep, wide neural networks introduces scalar order parameters to quantify task relations, revealing distinct performance regimes and phase transitions, including a novel "catastrophic anterograde interference" where new learning impedes generalization despite perfect memory of old tasks.

Key insights

• In single-head continual learning, short-term forgetting is primarily driven by "conflict" between tasks, defined by the difference between relevant-feature and rule similarities, rather than simple input feature overlap.
• Network depth has a non-trivial effect: deeper networks reduce short-term forgetting but can increase the time constant for long-term forgetting accumulation, suggesting an optimal depth for minimizing maximal long-term forgetting.
• Multi-head continual learning exhibits distinct phase transitions in performance (Fixed Representations, Overfitting, Generalization regimes) governed by network load and a newly defined task similarity order parameter, where "catastrophic anterograde interference" can occur.

Summary by alphaXiv.