Improving Multimodal Learning with Multi-Loss Gradient Modulation

TL;DR

This paper introduces a dynamic multi-loss gradient modulation technique for multimodal learning, improving the integration of audio and video data by balancing their contributions during training, leading to significant performance gains.

Contribution

It proposes a novel multi-loss objective with dynamic adjustment of modality learning rates, surpassing previous methods in multimodal audio-video classification tasks.

Findings

ResNet-based models improved by up to 12.4% on CREMA-D.

Conformer models achieved up to 14.1% enhancement on CREMA-D.

Performance gains of 2.7% to 7.7% on AVE, and up to 6.1% on UCF101.

Abstract

Learning from multiple modalities, such as audio and video, offers opportunities for leveraging complementary information, enhancing robustness, and improving contextual understanding and performance. However, combining such modalities presents challenges, especially when modalities differ in data structure, predictive contribution, and the complexity of their learning processes. It has been observed that one modality can potentially dominate the learning process, hindering the effective utilization of information from other modalities and leading to sub-optimal model performance. To address this issue the vast majority of previous works suggest to assess the unimodal contributions and dynamically adjust the training to equalize them. We improve upon previous work by introducing a multi-loss objective and further refining the balancing process, allowing it to dynamically adjust the learning pace of each modality in both directions, acceleration and deceleration, with the ability to phase out balancing effects upon convergence. We achieve superior results across three audio-video datasets: on CREMA-D, models with ResNet backbone encoders surpass the previous best by 1.9% to 12.4%, and Conformer backbone models deliver improvements ranging from 2.8% to 14.1% across different fusion methods. On AVE, improvements range from 2.7% to 7.7%, while on UCF101, gains reach up to 6.1%.