Generalized Resource-Aware Distributed Minimax Optimization
Keywords: Distributed minimax optimization, Generalization
Abstract: Traditional distributed minimax optimization algorithms cannot be applied in resource-limited clients dealing with large-scale models. In this work, we present *SubDisMO*, a generalized resource-aware distributed minimax optimization algorithm. *SubDisMO* prunes the global large-scale model into adaptive-sized submodels to accommodate varying resources during each communication round. However, the randomly pruned submodels are susceptible to *arbitrary submodel sharpness*, which can hinder generalization and lead to slow convergence. To address this issue, *SubDisMO* trains the arbitrarily pruned submodels with perturbations by optimizing the minimax objectives, enhancing the *generalization* performance of the aggregated full model. We theoretically analyze our proposed resource-aware *SubDisMO* algorithm, demonstrating that it achieves an asymptotically optimal convergence rate of $O(1/\sqrt{QT\mathcal{C}^*})$, which is dominated by the minimum covering number $\mathcal{C}^*$. We also show the generalization bound of *SubDisMO* corresponding to the remaining rate in each layer. Extensive experiments on *CIFAR-10* and *CIFAR-100* datasets demonstrate that *SubDisMO* achieves superior generalization and effectiveness compared to state-of-the-art baselines.
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.
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: 9485
Loading