Achieving a Better Stability-Plasticity Trade-off via Auxiliary Networks in Continual Learning

Sanghwan KimLorenzo NociAntonio OrvietoThomas Hofmann

Mar 2023

摘要

In contrast to the natural capabilities of humans to learn new tasks in asequential fashion, neural networks are known to suffer from catastrophicforgetting, where the model's performances on old tasks drop dramatically afterbeing optimized for a new task. Since then, the continual learning (CL)community has proposed several solutions aiming to equip the neural networkwith the ability to learn the current task (plasticity) while still achievinghigh accuracy on the previous tasks (stability). Despite remarkableimprovements, the plasticity-stability trade-off is still far from being solvedand its underlying mechanism is poorly understood. In this work, we proposeAuxiliary Network Continual Learning (ANCL), a novel method that applies anadditional auxiliary network which promotes plasticity to the continuallylearned model which mainly focuses on stability. More concretely, the proposedframework materializes in a regularizer that naturally interpolates betweenplasticity and stability, surpassing strong baselines on task incremental andclass incremental scenarios. Through extensive analyses on ANCL solutions, weidentify some essential principles beneath the stability-plasticity trade-off.

展开全部

图表提取

论文十问

参考文献

被引用

提问与回答

暂无人提供速读十问回答

被引用

笔记

问答

领域