Publication Type

Journal Article

Version

acceptedVersion

Publication Date

12-2025

Abstract

In this paper we review the concept of “phase” defined in Class-Incremental Learning (CIL), i.e., learning new classes while not forgetting old ones. Due to this design, classic CIL algorithms are mostly offline or can handle only intensive data distribution shifts across the phases. However, real-world data streams are often online, usually with uncertain or untraceable changes in their data distributions. To this end, we design the per-step distribution shifts by modeling the class sampling weights using bell-shaped curves. Such a design respects the rise-and-fall nature and presents realistic but underexplored challenges for CIL: 1) The data non-stationarity across steps requires the models to identify the recent dynamics and adopt an appropriate learning strategy for knowledge memorization and adaptation. 2) Over all steps, the proposed streams exhibit various class-imbalance patterns, with different majority classes and time-varying imbalance ratios. To address the challenges, we propose a novel Rate-Dependent Coreset Selector (RDCS), which essentially presents an adaptive and robust sample selection criterion when constructing memory for replay. We conduct extensive experiments by generating the proposed data streams on multiple image benchmarks and implementing RDCS in an efficient approximation, showing its superior performance.

Keywords

Class-incremental learning, Non-stationary data streams, Class imbalance, Coreset selection

Discipline

Software Engineering | Theory and Algorithms

Research Areas

Software and Cyber-Physical Systems

Publication

Neurocomputing

Volume

656

First Page

1

Last Page

9

ISSN

0925-2312

Identifier

10.1016/j.neucom.2025.131519

Publisher

Elsevier

Copyright Owner and License

Authors

Additional URL

https://doi.org/10.1016/j.neucom.2025.131519

Download

Share

COinS