Community Channel-Net: Efficient channel-wise interactions via community graph topology.

•We unify the channel interaction methodologies through channel graph perspectives, which provide a general and flexible framework for interpreting, modeling, and designing channel interaction approaches in DNNs;•We identify the optimal interaction pattern for channel interaction, where the channel graphs are connected with community-structure topologies. The topology can effectively increase channel diversity and bring dynamic and sparse properties into the learning process. Based on this, a novel channel interaction method called CC-Net is proposed, learning the channel interdependence and keeping a high degree of channel diversity concurrently;•Experimental results indicate that our method outperforms other channel interaction methods in terms of the accuracy (i.e., classification error) and efficiency (i.e., model size and computational complexity). Highlights (for review)