Z. Shi, J. Jing, Y. Sun, J. -H. Lim and M. Zhang, "Unveiling the Tapestry: The Interplay of Generalization and Forgetting in Continual Learning," in IEEE Transactions on Neural Networks and Learning Systems, doi: 10.1109/TNNLS.2025.3546269.
Abstract:
In AI, generalization refers to a model’s ability toperform well on out-of-distribution data related to the giventask, beyond the data it was trained on. For an AI agent toexcel, it must also possess the continual learning capability,whereby an agent incrementally learns to perform a sequenceof tasks without forgetting the previously acquired knowledgeto solve the old tasks. Intuitively, generalization within a taskallows the model to learn underlying features that can readily beapplied to novel tasks, facilitating quicker learning and enhancedperformance in subsequent tasks within a continual learningframework. Conversely, continual learning methods often includemechanisms to mitigate catastrophic forgetting, ensuring thatknowledge from earlier tasks is retained. This preservation ofknowledge over tasks plays a role in enhancing generalizationfor the ongoing task at hand. Despite the intuitive appeal ofthe interplay of both abilities, existing literature on continuallearning and generalization has proceeded separately. In thepreliminary effort to promote studies that bridge both fields,we first present empirical evidence showing that each of thesefields has a mutually positive effect on the other. Next, buildingupon this finding, we introduce a simple and effective techniqueknown as Shape-Texture Consistency Regularization (STCR),which caters to continual learning. STCR learns both shape andtexture representations for each task, consequently enhancinggeneralization and thereby mitigating forgetting. Remarkably,extensive experiments validate that our STCR, can be seamlesslyintegrated with existing continual learning methods, includingreplay-free approaches. Its performance surpasses these continuallearning methods in isolation or when combined with establishedgeneralization techniques by a large margin. Our data and sourcecode are available here.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the National Research Foundation, Singapore - AI Singapore Programme
Grant Reference no. : AISG2-RP-2021-025
This research / project is supported by the National Research Foundation, Singapore - National Research Foundation Fellowship Award
Grant Reference no. : NRF-NRFF15-2023-0001
This research / project is supported by the Agency for Science, Technology, and Research (A*STAR) - Start Up Grant
Grant Reference no. :
This research / project is supported by the Nanyang Technological University - Start Up Grant
Grant Reference no. :