Choose Before You Label: Efficient Node and Data Selection in Distributed Learning
Keywords: distributed learning, node selection
TL;DR: We propose and evaluate a strategy to select the nodes to include in distributed learning that does not require labelled data
Abstract: We consider one of the most relevant problems of distributed learning, i.e., the selection of the learning nodes to include in the training process as well as the selection of the samples from each of the learning nodes' local datasets, so as to make learning sustainable. Traditional approaches rely on pursuing a balanced label distribution, which requires label statistics from all datasets, including those not selected for learning. This may be costly and may raise privacy concerns. To cope with this issue, we aim at selecting few and small datasets. To this end, we propose a new metric, called loneliness, which is defined on unlabelled training samples. First, through both a theoretical and an experimental analysis, we show that loneliness is strongly linked with learning performance (i.e., test accuracy). Then, we propose a new node- and data-selection procedure, called Goldilocks, that uses loneliness to make its decisions. Our performance evaluation, including three state-of-the-art datasets and both centralized and federated learning, demonstrates that Goldilocks outperforms approaches based upon a balanced label distribution by providing over 70% accuracy improvement, in spite of using information that is both less sensitive privacy-wise and less onerous to obtain.
Primary Area: other topics in machine learning (i.e., none of the above)
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: 3086
Loading