Long-Tailed Class Incremental Learning

TL;DR

This paper introduces long-tailed class incremental learning scenarios, evaluates existing methods within them, and proposes a two-stage baseline with a learnable weight scaling layer that improves performance on imbalanced datasets.

Contribution

The paper defines new long-tailed CIL scenarios, systematically evaluates existing methods, and proposes a novel two-stage learning baseline with a learnable weight scaling layer.

Findings

Existing methods perform differently in long-tailed scenarios compared to balanced ones.

The proposed baseline improves incremental accuracy by up to 6.44 points.

The approach benefits both long-tailed and conventional CIL settings.

Abstract

In class incremental learning (CIL) a model must learn new classes in a sequential manner without forgetting old ones. However, conventional CIL methods consider a balanced distribution for each new task, which ignores the prevalence of long-tailed distributions in the real world. In this work we propose two long-tailed CIL scenarios, which we term ordered and shuffled LT-CIL. Ordered LT-CIL considers the scenario where we learn from head classes collected with more samples than tail classes which have few. Shuffled LT-CIL, on the other hand, assumes a completely random long-tailed distribution for each task. We systematically evaluate existing methods in both LT-CIL scenarios and demonstrate very different behaviors compared to conventional CIL scenarios. Additionally, we propose a two-stage learning baseline with a learnable weight scaling layer for reducing the bias caused by long-tailed distribution in LT-CIL and which in turn also improves the performance of conventional CIL due to the limited exemplars. Our results demonstrate the superior performance (up to 6.44 points in average incremental accuracy) of our approach on CIFAR-100 and ImageNet-Subset. The code is available at https://github.com/xialeiliu/Long-Tailed-CIL