DeepThin: A Self-Compressing Library for Deep Neural Networks

Abstract: As the industry deploys increasingly large and complex neural networks to mobile devices, more pressure is put on the memory and compute resources of those devices. Deep compression, or compression of deep neural network weight matrices, is a technique to stretch resources for such scenarios. Existing compression methods cannot effectively compress models smaller than 1-2% of their original size. We develop a new compression technique, DeepThin, building on existing research in the area of low rank factorization. We identify and break artificial constraints imposed by low rank approximations by combining rank factorization with a reshaping process that adds nonlinearity to the approximation function. We deploy DeepThin as a plug-gable library integrated with TensorFlow that enables users to seamlessly compress models at different granularities. We evaluate DeepThin on two state-of-the-art acoustic models, TFKaldi and DeepSpeech, comparing it to previous compression work (Pruning, HashNet, and Rank Factorization), empirical limit study approaches, and hand-tuned models. For TFKaldi, our DeepThin networks show better word error rates (WER) than competing methods at practically all tested compression rates, achieving an average of 60% relative improvement over rank factorization, 57% over pruning, 23% over hand-tuned same-size networks, and 6% over the computationally expensive HashedNets. For DeepSpeech, DeepThin-compressed networks achieve better test loss than all other compression methods, reaching a 28% better result than rank factorization, 27% better than pruning, 20% better than hand-tuned same-size networks, and 12% better than HashedNets. DeepThin also provide inference performance benefits ranging from 2X to 14X speedups, depending on the compression ratio and platform cache sizes.