Gradient-Guided Modality Decoupling for Missing-Modality Robustness

TL;DR

This work tackles missing-modality robustness in multimodal learning by identifying modality dominance as a key bottleneck. It introduces Gradient-guided Modality Decoupling (GMD) to cancel conflicting gradient components across modalities, coupled with Dynamic Sharing (DS) to adaptively switch off parameters for absent modalities, avoiding misleading imputations. The method demonstrates strong robustness on BraTS 2018 segmentation and CMU-MOSI/MOSEI sentiment tasks, outperforming existing baselines and addressing both unimodal and multimodal cases. The approach offers theoretical and empirical insights into gradient interactions across modalities and provides a practical, scalable solution with released code for community use.

Abstract

Multimodal learning with incomplete input data (missing modality) is practical and challenging. In this work, we conduct an in-depth analysis of this challenge and find that modality dominance has a significant negative impact on the model training, greatly degrading the missing modality performance. Motivated by Grad-CAM, we introduce a novel indicator, gradients, to monitor and reduce modality dominance which widely exists in the missing-modality scenario. In aid of this indicator, we present a novel Gradient-guided Modality Decoupling (GMD) method to decouple the dependency on dominating modalities. Specifically, GMD removes the conflicted gradient components from different modalities to achieve this decoupling, significantly improving the performance. In addition, to flexibly handle modal-incomplete data, we design a parameter-efficient Dynamic Sharing (DS) framework which can adaptively switch on/off the network parameters based on whether one modality is available. We conduct extensive experiments on three popular multimodal benchmarks, including BraTS 2018 for medical segmentation, CMU-MOSI, and CMU-MOSEI for sentiment analysis. The results show that our method can significantly outperform the competitors, showing the effectiveness of the proposed solutions. Our code is released here: https://github.com/HaoWang420/Gradient-guided-Modality-Decoupling.

Figures (4)

• Figure 1: Overall architecture including Dynamic Sharing (DS) and Gradient-guided Modality Decoupling (GMD). Multimodal data is first mapped to a common feature space through modality encoders $\{f_1, f_2,...,f_m\}$. Then the outputs of these encoders are fed to a shared architecture: $T$. Then the outputs $\{h_1, h_2,...,h_m\}$ go through a flexible fusion operator (e.g., average pooling) to obtain modal-incomplete representations $\{H_1, H_2,..., H_z\}$. GMD decouples the conflicted gradients, and the calibrated gradients are back-propagated to reduce modality dominance.
• Figure 2: (a) demonstrates one gradient dominating another. The combined gradient $\mathcal{G}$ only has a small projected component at the direction of $\mathcal{G}_k$, highlighted in green; (b) shows the conflicting components to be removed, marked as ✗; (c) visualizes the calibrated gradients. The corrected gradients are now of a similar scale and direction. Both modal-incomplete cases can be effectively updated without conflict during optimization.
• Figure 3: Gradients from modalities vary significantly in norm (std: $4\times10^4$). Previous works ignore gradient conflicts, leading to bias towards dominant gradients (red). Our GMD removes conflicting gradients and reduces angle differences, achieving more balanced optimization (blue).
• Figure 4: Visualization of weights given by GMD on different modal-incomplete cases, smoothed with a window size of 100 steps. For cases with T1CE, the dominant modality, presented, its gradient (purple line) is relatively suppressed to allow for optimization of cases with only weak modalities (red and blue line).