Exploring Modality Collaboration with Modality-Agnostic Transformers in Multi-Modal Federated Learning
Primary Area: societal considerations including fairness, safety, privacy
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.
Keywords: federated learning, computer vision, vision transformer, multi-modal learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: Our paper introduces Modality-Collaborated Federated Learning as a novel setting in federated learning, which focuses on enabling collaboration between uni-modal clients, and a baseline framework, FedCola, based on modality-agnostic tranformers.
Abstract: In Federated Learning (FL), the focus has predominantly been on uni-modal scenarios, limiting the system's ability to leverage multi-modal data. This paper introduces a novel setting, Modality-Collaborated Federated Learning (MCFL), designed to facilitate collaboration among uni-modal clients with different data modalities. Unlike existing frameworks that emphasize multi-modal clients and tasks, MCFL aims to be more practical by focusing on uni-modal clients and ensuring performance gains across individual modalities. To address the challenges of model heterogeneity and modality gaps in MCFL, we propose Federated Modality Collaboration (FedCola), a framework based on a modality-agnostic transformer. FedCola explores optimal strategies in cross-modal parameter-sharing, model aggregation, and temporal modality arrangement. Our comprehensive evaluations demonstrate that FedCola significantly outperforms existing solutions, serving as a robust baseline for MCFL and marking a substantial advancement in federated learning.
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: 1455
Loading