Learnability and Algorithm for Continual Learning

Gyuhak Kim, Changnan Xiao, Tatsuya Konishi, Bing Liu
Proceedings of the 40th International Conference on Machine Learning, PMLR 202:16877-16896, 2023.

Abstract

This paper studies the challenging continual learning (CL) setting of Class Incremental Learning (CIL). CIL learns a sequence of tasks consisting of disjoint sets of concepts or classes. At any time, a single model is built that can be applied to predict/classify test instances of any classes learned thus far without providing any task related information for each test instance. Although many techniques have been proposed for CIL, they are mostly empirical. It has been shown recently that a strong CIL system needs a strong within-task prediction (WP) and a strong out-of-distribution (OOD) detection for each task. However, it is still not known whether CIL is actually learnable. This paper shows that CIL is learnable. Based on the theory, a new CIL algorithm is also proposed. Experimental results demonstrate its effectiveness.

Cite this Paper

BibTeX
@InProceedings{pmlr-v202-kim23x, title = {Learnability and Algorithm for Continual Learning}, author = {Kim, Gyuhak and Xiao, Changnan and Konishi, Tatsuya and Liu, Bing}, booktitle = {Proceedings of the 40th International Conference on Machine Learning}, pages = {16877--16896}, year = {2023}, editor = {Krause, Andreas and Brunskill, Emma and Cho, Kyunghyun and Engelhardt, Barbara and Sabato, Sivan and Scarlett, Jonathan}, volume = {202}, series = {Proceedings of Machine Learning Research}, month = {23--29 Jul}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v202/kim23x/kim23x.pdf}, url = {https://proceedings.mlr.press/v202/kim23x.html}, abstract = {This paper studies the challenging continual learning (CL) setting of Class Incremental Learning (CIL). CIL learns a sequence of tasks consisting of disjoint sets of concepts or classes. At any time, a single model is built that can be applied to predict/classify test instances of any classes learned thus far without providing any task related information for each test instance. Although many techniques have been proposed for CIL, they are mostly empirical. It has been shown recently that a strong CIL system needs a strong within-task prediction (WP) and a strong out-of-distribution (OOD) detection for each task. However, it is still not known whether CIL is actually learnable. This paper shows that CIL is learnable. Based on the theory, a new CIL algorithm is also proposed. Experimental results demonstrate its effectiveness.} }
Endnote
%0 Conference Paper %T Learnability and Algorithm for Continual Learning %A Gyuhak Kim %A Changnan Xiao %A Tatsuya Konishi %A Bing Liu %B Proceedings of the 40th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2023 %E Andreas Krause %E Emma Brunskill %E Kyunghyun Cho %E Barbara Engelhardt %E Sivan Sabato %E Jonathan Scarlett %F pmlr-v202-kim23x %I PMLR %P 16877--16896 %U https://proceedings.mlr.press/v202/kim23x.html %V 202 %X This paper studies the challenging continual learning (CL) setting of Class Incremental Learning (CIL). CIL learns a sequence of tasks consisting of disjoint sets of concepts or classes. At any time, a single model is built that can be applied to predict/classify test instances of any classes learned thus far without providing any task related information for each test instance. Although many techniques have been proposed for CIL, they are mostly empirical. It has been shown recently that a strong CIL system needs a strong within-task prediction (WP) and a strong out-of-distribution (OOD) detection for each task. However, it is still not known whether CIL is actually learnable. This paper shows that CIL is learnable. Based on the theory, a new CIL algorithm is also proposed. Experimental results demonstrate its effectiveness.
APA
Kim, G., Xiao, C., Konishi, T. & Liu, B.. (2023). Learnability and Algorithm for Continual Learning. Proceedings of the 40th International Conference on Machine Learning, in Proceedings of Machine Learning Research 202:16877-16896 Available from https://proceedings.mlr.press/v202/kim23x.html.

Related Material