Data-Evolution Learning

Peng Sun, Rui Yin, Tianyu Du, Tao Lin

Keywords: Data-centric Learning

Abstract:

Recent advancements in machine learning have been driven by models trained on large-scale, high-quality datasets. However, the practical application of these models faces two significant challenges: the infeasibility of acquiring precise labels in real-world settings and the substantial computational burden imposed by training large models. While existing approaches—such as self-supervised learning, weak supervision, noisy label learning, and dataset distillation—address these challenges from a model-centric perspective, they often overlook the potential benefits of optimizing the data itself.
This paper introduces a novel data-centric learning paradigm where both the dataset and the model co-evolve during the learning process.
We formalize this paradigm and propose a Data-evolution Learning Algorithm (DeLA), which offers three key advantages: optimized dataset generation, versatile dataset compatibility, and effective utilization of prior knowledge.
Extensive experiments demonstrate that DeLA enables the creation of optimized datasets for reuse in subsequent training, effectively addressing diverse datasets with varying target types. Moreover, DeLA accelerates learning by utilizing architecture-agnostic, open-source prior models for efficient data creation.
Notably, DeLA frequently outperforms traditional SOTA model-centric methods in self-supervised and noisy label learning.
Furthermore, its simplicity enables implementation in only two lines of PyTorch code, offering significant potential for advancements in representation learning.
Our code will be made publicly available.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation 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: 10334