Retrospective Class Incremental Learning

Qingyi Tao, Chen Change Loy, Jianfei Cai, Zongyuan Ge, Simon See

Research output: Chapter in Book/Report/Conference proceedingConference PaperResearchpeer-review

Abstract

Existing works study the Class Incremental learning (CIL) problem with the assumption that the data for previous classes are absent, or only a small subset of samples (known as exemplars) are accessible. Differently, we propose a new and practical setting called retrospective CIL, where all the previous data are accessible, but with bounded training budgets for old data replay. Since only a small subset of old samples can be replayed, it brings a new research problem, i.e., dynamically sampling old data along the incremental training process. As incremental learning particularly suffers from catastrophic forgetting, we propose to use the forgettability of the old samples as the sampling priorities to favour the forgotten samples during the dynamic sampling process. To achieve this, we introduce a forgetting rate metric with graph-based propagation to estimate the sample forgettability. The proposed method brings improvements on two benchmark datasets.

Original languageEnglish
Title of host publication2021 IEEE International Conference on Multimedia and Expo, ICME 2021
EditorsGiuseppe Valenzise, Wengang Zhou
Place of PublicationPiscataway NJ USA
PublisherIEEE, Institute of Electrical and Electronics Engineers
Number of pages6
ISBN (Electronic)9781665438643
ISBN (Print)9781665411523
DOIs
Publication statusPublished - 2021
EventIEEE International Conference on Multimedia and Expo 2021 - Shenzhen, China
Duration: 5 Jul 20219 Jul 2021
https://ieeexplore.ieee.org/xpl/conhome/9428049/proceeding (Proceedings)

Publication series

NameProceedings - IEEE International Conference on Multimedia and Expo
PublisherIEEE, Institute of Electrical and Electronics Engineers
ISSN (Print)1945-7871
ISSN (Electronic)1945-788X

Conference

ConferenceIEEE International Conference on Multimedia and Expo 2021
Abbreviated titleICME 2021
Country/TerritoryChina
CityShenzhen
Period5/07/219/07/21
Internet address

Keywords

  • catastrophic forgetting
  • continual learning
  • lifelong learning

Cite this