Utilizing Auxiliary Data as Neutral References to Enhance Decision Boundary Robustness in Continual Learning
- 주제(키워드) Continual learning , Neural networks , Deep learning , Computer vision
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 Jongbin Ryu
- 발행년도 2026
- 학위수여년월 2026. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000035800
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
Maintaining stable decision boundaries remains challenging in continual learning, as the introduction of new classes frequently shifts old boundaries and leads to catastrophic forget- ting. To address this challenge, we revisit the function of auxiliary samples, which include neutral and general information. Typically, when a dataset is constructed through the random collection of enough samples, it contains neutralized and generic information, and it does not belong to any specific class. Concretely, we introduce a framework that treats auxiliary samples without labels: (i) it draws auxiliary samples from an external out-of-distribution dataset and maximizes predictive entropy with respect to the class distribution of current task, discouraging reliance on spurious and shared features; and (ii) it utilizes a self-supervision between current-task data and auxiliary samples to expand the learning over a wider input distribution, hence enhancing discriminative and class-specific features. In our experiments, we demonstrate that the incorporation of auxiliary samples strengthens decision boundaries and reduces feature overlap with new classes, yielding improved retention and reduced forget- ting. Our method achieves strong performance across the S-CIFAR-10/100, S-ImageNet100, S-TinyImageNet, and DomainNet datasets. Our analysis also shows that conventional training approaches struggle to differentiate between classes as continual learning evolves; however, our approach mitigates this effect by encouraging robust and task-specific decision boundaries.
more목차
1 Introduction 1
2 Related Work 3
2.1 Continual Learning 3
2.2 Auxiliary Samples 3
2.3 Continual Self-supervised Learning 4
3 Method 4
3.1 Entropy Maximization on Auxiliary Sample 4
3.2 Self-supervised Learning with Bi-Level Fusion 5
3.3 Empirical Evidence 7
4 Experiments 8
4.1 Experimental Settings 9
4.2 Experimental Results 10
5 Analysis 12
6 Conclusion 14
A Appendix 19
A.1 GPU Overhead in Training Phase 19
A.2 Analysis on Auxiliary Dataset 19
A.3 Analysis on Task Recency Bias 20
A.4 Auxiliary Dataset 20
A.5 Hyperparameter Settings 21
