Abstract: Highlights•The network further explores the information integration between modality- specific and modality-invariant spaces.•A novel module named Gradient-based Representation Enhancement (GRE) is proposed which is effective for extracting significant variation of representation matrices.
Loading