Structured Bayesian Compression for Deep Neural Networks Based on the Turbo-VBI Approach

With the growth of neural network size, model compression has attracted increasing interest in recent research. As one most common techniques, pruning been studied for a long time. By exploiting structured sparsity network, existing methods can prune neurons instead individual weights. However, methods, surviving are randomly connected without any structure, and non-zero weights within each neuron also distributed. Such irregular sparse structure cause very high control overhead memory access hardware even increase computational complexity. In this paper, we propose three-layer hierarchical prior to promote more regular during pruning. The proposed achieve per-neuron weight-level neuron-level sparsity. We derive an efficient Turbo-variational Bayesian inferencing (Turbo-VBI) algorithm solve resulting problem with prior. Turbo-VBI low complexity support general priors than algorithms. Simulation results show that our pruned networks while achieving better performance terms rate accuracy compared baselines.