Revisiting Class-Incremental Learning with Pre-Trained Models: Generalizability and Adaptivity are All You Need

Paper (opens new window) | Code (opens new window) |

0. Abstract

• 基于预训练模型，考虑了类增量学习的工作，pre-trained models (PTMs) + CIL
• 预训练模型核心因素：adaptivity for model updating and generalizability for knowledge transferring
• SimpleCIL：不在下游任务训练，基于预训练模型获得的类中心，直接进行分类任务的分类。
• ADapt And Merge

1. Introduction

2. Related Work

3. Methodology

4. Experiments

5. Conclusion