Paper ID: 2403.20280

Sparse multimodal fusion with modal channel attention

Josiah Bjorgaard

The ability of masked multimodal transformer architectures to learn a robust embedding space when modality samples are sparsely aligned is studied by measuring the quality of generated embedding spaces as a function of modal sparsity. An extension to the masked multimodal transformer model is proposed which incorporates modal-incomplete channels in the multihead attention mechanism called modal channel attention (MCA). Two datasets with 4 modalities are used, CMU-MOSEI for multimodal sentiment recognition and TCGA for multiomics. Models are shown to learn uniform and aligned embedding spaces with only two out of four modalities in most samples. It was found that, even with no modal sparsity, the proposed MCA mechanism improves the quality of generated embedding spaces, recall metrics, and subsequent performance on downstream tasks.

Submitted: Mar 29, 2024