Dynamic deep neural network inference via adaptive channel skipping

Deep neural networks have recently made remarkable achievements in computer vision applications. However, the high computational requirements needed to achieve accurate inference results can be a significant barrier to deploying DNNs on resource-constrained computing devices, such as those found in the Internet-of-things. In this work, we propose a fresh approach called adaptive channel skipping (ACS) that prioritizes the identification of the most suitable channels for skipping and implements an efficient skipping mechanism during inference. We begin with the development of a new gating network model, ACS-GN, which employs fine-grained channel-wise skipping to enable input-dependent inference and achieve a desirable balance between accuracy and resource consumption. To further enhance the efficiency of channel skipping, we propose a dynamic grouping convolutional computing approach, ACS-DG, which helps to reduce the computational cost of ACS-GN. The results of our experiment indicate that ACS-GN and ACS-DG exhibit superior performance compared to existing gating network designs and convolutional computing mechanisms, respectively. When they are combined, the ACS framework results in a significant reduction of computational expenses and a remarkable improvement in the accuracy of inferences