Eager pruning

Today's big and fast data and the changing circumstance require fast training of Deep Neural Networks (DNN) in various applications. However, training a DNN with tons of parameters involves intensive computation. Enlightened by the fact that redundancy exists in DNNs and the observation that the ranking of the significance of the weights changes slightly during training, we propose Eager Pruning, which speeds up DNN training by moving pruning to an early stage. Eager Pruning is supported by an algorithm and architecture co-design. The proposed algorithm dictates the architecture to identify and prune insignificant weights during training without accuracy loss. A novel architecture is designed to transform the reduced training computation into performance improvement. Our proposed Eager Pruning system gains an average of 1.91x speedup over state-of-the-art hardware accelerator and 6.31x energy-efficiency over Nvidia GPUs.