Abstract
Deep multimodal learning has achieved great progress in recent years. However, current fusion approaches are static in nature, i.e., they process and fuse multimodal inputs with identical computation, without accounting for diverse computational demands of different multimodal data. In this work, we propose dynamic multimodal fusion (DynMM), a new approach that adaptively fuses multimodal data and generates data-dependent forward paths during inference. DynMM can reduce redundant computations for "easy" multimodal inputs (that can be predicted correctly using only one modality or simple fusion techniques) and retain representation power for "hard" samples by adopting all modalities and complex fusion operations for prediction. Results on various multimodal tasks demonstrate the efficiency and wide applicability of our approach. For instance, DynMM can reduce the computation cost by 46.5% with a negligible accuracy loss on CMU-MOSEI sentiment analysis. For RGB-D semantic segmentation on NYU Depth data, DynMM achieves a +0.7% mIoU improvement with over 21% reductions for the depth encoder when compared with strong baselines. We believe this opens a novel direction towards dynamic multimodal network design, with applications to a wide range of multimodal tasks.
Abstract (translated)
URL
https://arxiv.org/abs/2204.00102
