ConfSMoE: Confidence-Guided Gating for Sparse MoE with Missing Modalities

Researchers have introduced ConfSMoE, a novel approach to address the challenge of missing modalities in Sparse Mixture-of-Experts (SMoE) architectures within multimodal learning. Real-world data often suffers from incompleteness due to sensor failures or collection errors, causing existing SMoE models to experience performance degradation and expert collapse. To mitigate this, the team proposes a two-stage imputation module that leverages expert opinions to handle missing inputs effectively. Theoretically grounded, the method introduces a confidence-guided gating mechanism that detaches softmax routing scores from task confidence scores relative to ground truth signals. This innovation naturally alleviates expert collapse without requiring additional load balance loss functions, aligning with insights from Gaussian and Laplacian gating mechanisms. The proposed model was rigorously evaluated across four real-world datasets under three distinct experimental settings. Results demonstrate ConfSMoE's superior resistance to missing modality issues and highlight the positive impact of the new gating mechanism on model generalization and stability. This advancement offers a robust solution for deploying SMoE architectures in practical, imperfect data environments.