Can Information-Theoretic Generalization Bound Explain the Generalization of Pre-trained Language Model?
Keywords: Information-Theoretic Generalization Bound, Pre-trained Language Model
TL;DR: We use information-theoretic generalization bound to explain the generalization of pre-trained language model, and demonstrate the limitations of current information-theoretic generalization bound through experiments and theory.
Abstract: Although language models exhibit exceptional generalization capabilities in downstream tasks after extensive text pre-training, the underlying causes behind this generalization remain unclear. Existing studies on information-theoretic generalization bounds suggest that the compression of information stored in the weights (IIW) is a crucial factor influencing a model's ability to generalize, with some experiments indicating a correlation between lower IIW and improved generalization. However, it remains uncertain whether IIW is applicable to pre-trained language models. In this work, we find that using IIW can explain why the pre-trained language models have better generalization compared to non-pre-trained language models. Unfortunately, we also discover that IIW does not consistently reflect the degree of generalization when applying IIW to study the fine-tuning process of pre-trained language models. We revisit existing IIW estimation methods, highlighting their limitations in accurately estimating IIW based on theoretical and empirical evidence. Our findings suggest that current information-theoretic generalization bounds, constrained by the limitations of IIW estimation methodologies, fail to accurately capture the generalisation performance of pre-trained language models.
Supplementary Material: zip
Primary Area: learning theory
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: 747
Loading