SATCH: Specialized Assistant Teacher Distillation to Reduce Catastrophic Forgetting
Keywords: Continual Learning, Catastrophic Forgetting, Knowledge Distillation, Class Incremental Learning
Abstract: Continual learning enables models to learn new tasks sequentially without forgetting previously learned knowledge. Knowledge distillation reduces forgetting by using a single teacher model to transfer previous knowledge to the student model. However, existing methods face challenges, specifically loss of task-specific knowledge, limited diversity in the transferred knowledge, and delays in teacher availability. These issues stem from self-distillation, where the teacher is a mere snapshot of the student after learning a new task, inheriting the student’s biases and becoming available only after learning a task. We propose Specialized Assistant TeaCHer distillation (SATCH), a novel method that uses a smaller assistant teacher trained exclusively on the current task. By incorporating the assistant teacher early in the learning process, SATCH provides task-specific guidance, improves the diversity of transferred knowledge, and preserves critical task-specific insights. Our method integrates seamlessly with existing knowledge distillation techniques, and experiments on three standard continual learning benchmarks show that SATCH improves accuracy by up to 12% when combined with four state-of-the-art methods. Code is available in supplementary materials.
Supplementary Material: zip
Primary Area: transfer learning, meta learning, and lifelong learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 9863
Loading