Memory Savings by Sharing One Source: Insights into Subsetsum Approximation
Keywords: Strong lottery tickets, Subsetsum approximation, Mixture of Experts, Ensembles, Memory savings, Discrete optimization
Abstract: Large deep neural networks, often fine-tuned from foundation models, dominate modern machine learning, but their high memory requirements limit deployment on resource-constrained devices. Strong lottery tickets (SLTs) offer a promising solution by significantly reducing memory usage, as they are fully characterized by a seed for generating a random source network and a binary mask. Notably, multiple models can share the same source network without increasing its width requirement. As we show, this source sharing can lead to memory savings when experts share specific sparsity patterns. Based on novel insights into optimized subset sum approximations, we also show how the masks can be adjusted to further reduce memory overhead. To validate these theoretical findings, we provide explicit SLT constructions in experiments.
Supplementary Material: zip
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: 10323
Loading