ICAF：具有插值一致性和自适应筛选的无监督持续学习

打开文本图片集

中图分类号：TP18；TP391.41 文献标识码：A 文章编号：2096-4706（2025）11-0049-06

1，²，1，2（1.，；2.，)

ICAF: Unsupervised Continual Learning with Interpolation Consistency and Adaptive Filtering

YE Gencheng1², WANG Xiang1,2 (1.Beijing KeyLabofTraffc Data Analysisand Mining (Beijing Jiaotong University),Beijingl0o044,China; 2.School ofComputer Science and Technology,Beijing Jiaotong University,Beijing 10oo44,China)

Abstract: Unsupervised Continual Learning addresss the challenge of sequentiall learning tasks without supervision information.However,thissetingoftenleadstocatastropicforgeting.Toovercomethisisue,memorybanksareusedtostore samplesfrompastasks.However,storingandreplayingappropriatesamplesisacomplextask,andthesimpleandcommonly usedrandomselectionis noteffective,especiallyforlong-sequence taskswithlarge-scale samples.Tosolve this isue,an Unsupervised ContinualLearning method with Interpolation Consistencyand AdaptiveMemoryFiltering (ICAF)is proposed.It canefectivelyselect high-qualitysamples byusing thefeaturesleamedfromrandomlyaugmentedsamples,adjust thethreshold during training,andachieveinterpolationconsistency throughaspeciallydesigned linear interpolationanddatacombination strategy,thereby preventing insuffcientlearning causedbyrelying solelyoninterpolatedsamplesfrompasttasks.This method has achieved the best results on multiple experimental datasets.

Keywords: Continual Learning; Unsupervised Learning; Representation Learning; Contrastive Learning; Selfsupervised Learning

0 引言

近年来，持续学习在人工智能领域受到了广泛关注。（剩余12294字）