Reproduce “Learning Fast, Learning Slow: A General Continual Learning Method based on Complementary Learning System” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Elahe Arani, Fahad Sarfraz, Bahram Zonooz

CLS-ER introduces a dual-memory experience replay method for continual learning, drawing inspiration from the brain's complementary learning system to mitigate catastrophic forgetting. It leverages distinct plastic and stable semantic memories to consolidate knowledge, achieving state-of-the-art performance in class-incremental, domain-incremental, and general incremental learning benchmarks, while also producing better calibrated and more robust models.

Key insights

  • Integrating a dual-memory system inspired by the brain's hippocampus (episodic memory) and neocortex (plastic/stable semantic memories) effectively balances plasticity and stability in continual learning.
  • Dynamically guiding the working model with consolidated knowledge from both short-term (plastic) and long-term (stable) semantic memories prevents forgetting and promotes robust knowledge integration.
  • The CLS-ER approach naturally leads to models that converge to flatter loss minima, exhibit less bias towards recent tasks, and provide better calibrated uncertainty estimates.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.