Poor Teaching: Explore and Question Knowledge Distillation under Distribution Shift

Songming Zhang, Ziyu Lyu, Xiaofeng Chen

Keywords: Knowledge Distillation, Distribution Shift

TL;DR: We propose a unified framework for evaluating the performance of knowledge distillation under distribution shift.

Abstract:

Knowledge distillation techniques transfer knowledge from a complex or large learning model into a small model, and have made remarkable achievements in recent decades. However, few studies has investigated and explored the mechanism of the knowledge distillation against distribution shifts in real scenarios. In this paper, we reconsider the knowledge distillation paradigm under the shift situations, by reformulating the objectives of distillation with multiple domains. Under the novel paradigm, we propose a unified and systematic evaluation framework to benchmark knowledge distillation against two general distributional shifts including diversity and correlation shift.
The evaluation benchmark covers more than 20 methods from algorithmic, data-driven, and optimization perspectives for five benchmark datasets. Extensive experiments are constructed and some constructive findings are summarized to explain when and how the existing knowledge distillation methods work against distribution shifts.

Primary Area: datasets and benchmarks

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/2024/AuthorGuide.

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: 3183