Exploring energy and accuracy tradeoff in structure simplification of trained deep neural networks

摘要

This paper presents a structure simplification procedure that allows efficient energy and accuracy tradeoffs in implementation of trained deep neural networks (DNNs). This structure simplification procedure identifies and eliminates redundant neurons in any layer of the DNN based on the trained weights connected from these neurons. This procedure may be applied to all layers of a DNN. For each layer different configurations with Pareto-optimal accuracy and energy consumption are realized. Our work is the first to use energy-accuracy trade-offs to guide optimal structure realization of trained DNNs. After redundant neurons are discarded, the weights of remaining neurons will be updated using matrix multiplication without retraining. Yet, retraining may still be applied if desired to further fine tune the performance. In our experiments, we show energy-accuracy tradeoff provides clear guidance to achieve efficient realization of trained DNNs. We also observe significant implementation cost reductions with up to 33X in energy and 12X in memory while the performance (accuracy) loss is negligible.