Equally Critical: Samples, Targets, and Their Mappings in Datasets
Keywords: Data-efficient Learning
Abstract: Neural scaling laws highlight the trade-off between test error reduction and increased resources in machine learning, revealing diminishing returns as data volume, model size, and computational power increase.
This inefficiency poses sustainability challenges, as marginal performance gains necessitate exponential resource consumption.
Recent works have investigated these laws from a data-efficient standpoint, primarily concentrating on sample optimization, while largely neglecting the influence of target.
In this study, we first demonstrate that, given an equivalent training budget, employing soft targets on a 10% subset can outperform the use of one-hot targets on the full dataset. Building on this observation, we review existing paradigms in the sample-target relationship, categorizing them into distinct sample-to-target mapping strategies.
Subsequently, we propose a unified loss framework to assess their impact on training efficiency. Finally, we conduct a comprehensive analysis of how variations in target and sample types, quantities, and qualities influence training efficiency across three training strategies, providing six key insights to enhance training efficacy.
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.
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: 10518
Loading