Reproduce “A Unified Continual Learning Framework with General Parameter-Efficient Tuning” on OpenResearch

About this paper

Qiankun Gao, Chen Zhao, Yifan Sun, Teng Xi, Gang Zhang, Bernard Ghanem, Jian Zhang

Peking University Shenzhen Graduate School and KAUST researchers developed LAE, a unified Continual Learning framework that leverages diverse Parameter-Efficient Tuning (PET) methods across various pre-trained model architectures. This memory-free framework consistently surpasses prior state-of-the-art methods, achieving over 3.5% A10 improvement on ImageNet-R by effectively mitigating catastrophic forgetting.

Key insights

• Calibrating the adaptation speed of diverse Parameter-Efficient Tuning (PET) modules is crucial for achieving consistent and robust continual learning performance across different architectures.
• A dual-expert system, with an "online" module for rapid new task adaptation and an "offline" module for stable multi-task knowledge accumulation via EMA, effectively balances plasticity and stability.
• An energy-based ensemble of these two expert models (online and offline) allows for robust inference by leveraging their complementary strengths in recent and historical tasks without explicit task ID.

Summary by alphaXiv.