Dynamic Routing Mixture of Experts for Enhanced Multi-Label Image Classification
Keywords: Multi-label image classification, Dynamic Routing Mixture of Experts, Computer vision, Dynamic gating networks, Label heterogeneity
Abstract: Multi-label image classification (MLC) is a fundamental task in computer vision, requiring the identification of multiple objects or attributes within a single image. Traditional approaches often rely on shared backbones and static gating mecha-nisms, which can struggle to effectively capture complex label correlations and handle label heterogeneity, leading to issues such as negative transfer. In this pa-per, we introduce the Dynamic Routing Mixture of Experts (DR-MoE) model, a novel architecture that integrates input-dependent dynamic gating networks into the mixture-of-experts (MoE) framework for MLC. Unlike static gating in exist-ing models like the Hybrid Sharing Query (HSQ) Yin et al. (2024), our dynamic gating mechanism adaptively selects and weights both shared and task-specific experts based on the input image features. This allows DR-MoE to better capture varying label dependencies and mitigate negative transfer, resulting in improved overall and per-label classification performance. We conduct extensive experi-ments on benchmark datasets MS-COCO Lin et al. (2014) and PASCAL VOC 2007 Everingham et al. (2015), demonstrating that DR-MoE achieves state-of-the-art results, outperforming existing methods including HSQ, Q2L Liu et al.(2021), and ML-GCN Chen et al. (2019). Additionally, ablation studies confirm the effectiveness of dynamic gating in enhancing model adaptability and perfor-mance, particularly for labels with high heterogeneity. Our findings suggest that incorporating dynamic routing mechanisms into MoE architectures is a promising direction for advancing multi-label image classification.
Primary Area: applications to computer vision, audio, language, and other modalities
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: 13868
Loading