DropBack : continuous pruning during deep neural network training

In recent years, neural networks have regained popularity in a variety of fields such as image recognition and speech transcription. As deep neural networks grow more popular for solving everyday tasks, deployment on small embedded devices — such as phones — is becoming increasingly popular. Moreover, many applications — such as face recognition or health applications — require personalization, which means that networks must be retrained after they have been deployed. Because today’s state-of-the-art networks are too large to fit on mobile devices and exceed mobile device power envelopes, techniques such as pruning and quantization have been developed to allow pre-trained networks to be shrunk by about an order of magnitude. However, they all assume that the network is first fully trained off-line on datacenter-class GPUs, then pruned in a post-processing step, and only then deployed to the mobile device. In this thesis, we introduce DropBack, a technique that significantly reduces the storage and computation required during both inference and training. In contrast to existing pruning schemes, which retain the weights with the largest values and set the rest to zero, DropBack identifies the weights that have changed the most, and recomputes the original initialization values for all other weights. This means that only the most important weights must be stored in off-chip memory both during inference and training, reducing off-chip memory accesses (responsible for a majority of the power usage) by up to 72×. Crucially, networks pruned using DropBack maintain high accuracy even for challenging network architectures: indeed, on modern, compact network architectures such as Densenet and WRN-28-10, DropBack outperforms the current state-of-the- art pruning techniques in both accuracy and off-chip memory storage required for weights. On the CIFAR-10 dataset, we observe 5× reduction in weights on an already 9×-reduced VGG-16 network, which we call VGG-S, and 4.5× on Densenet and WRN-28-10 — all with zero or negligible accuracy loss — or 19×, 27×, and 36×, respectively, with a minor impact on accuracy. When the recomputed initial weights are decayed to zero, the weight memory footprint of WRN-28-10 can be reduced up to 72×.Applied Science, Faculty ofElectrical and Computer Engineering, Department ofGraduat