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

Abstract

The layer-wise structure of deep neural networks (DNNs) isolates the channel interactions in the same layer, which significantly impedes the efficient learning of DNNs. Several existing methods enable channel-wise information exchange via learning channel interdependence in a heuristic and empirical manner. Nevertheless, only informative channels are emphasized while other channels are suppressed in these approaches. This results in a low channel diversity, which impeds the generalization of DNNs. Our work aims to learn channel-wise interdependence and keep the channel diversity concurrently via designing optimal channel interaction patterns. We model the channel interaction pattern from a graph perspective, where the interactions can be regarded as information exchange on the channel graph. Based on this framework, we propose the Community Channel-Net (CC-Net), using a community-based graph topology for channel interaction. Each community contains channels with semantic commonalities, and the inter-community connections are activated among critical channels. With this structured and dynamic topology, the channels from the same community can learn channel interdependence, and those critical channels from distinct communities can gain more diverse features. CC-Net outperforms baselines on image classification tasks over various backbones with fewer computational costs.