Learning Dynamic Computing Resource Allocation in Convolutional Neural Networks for Wireless Interference Identification

Wireless interference identification (WII) is a critical technology for the non-cooperative communication systems, and it is widely applied into military and civilian scenarios. With enormous success in many computer vision and language processing applications, deep learning (DL) has also achieved the remarkable performance for WII. However, existing DL based WII methods can not dynamically allocate proper computing resources conditioned on the inputs, causing serious waste of resources and low computational efficiency, which is unacceptable for edge computing devices, such as unmanned aerial vehicle (UAV)-aided systems. In this paper, the problem of dynamic computation resource allocation is modeled as adaptively learning forward propagation depth of networks on a per-input basis, and this novel method is termed as dynamic computing resource allocation in convolutional neural networks (DCRACNN). In particular, the proposed DCRACNN is composed of two subnetworks: a main network and a policy network. The main network is equipped with multiple classifiers at different depths, allowing test examples to stop early during inference so as to significantly improve computational efficiency. The policy network is optimized with reinforcement learning to address the bottleneck of non-differentiable decisions of optimal forward propagation depth for the main network while preserving recognition accuracy. To address the challenges of high error rates of early classifiers in the main network, novel methods are introduced, which can be seamlessly integrated into DRCACNN. In addition, a novel network pruning process is proposed for the main network to further reduce model sizes and computational complexity. Experiments demonstrate that DCRACNN can significantly reduce the calculation cost with boosting accuracy for WII.