Deep artificial Neural Networks (DNNs) are currently one of the most intensively and widely used predictive models in the field of machine learning. However, the computational workload involved in DNNs is typically out of reach for lowpower embedded devices. The approximate computing paradigm can be exploited to reduce the DNN complexity. It improves performance and energy-efficiency by relaxing the need for fully accurate operations. There are a large number of implementation options leveraging many approximation techniques (e.g., pruning, quantization, weight-sharing, low-rank factorization, knowledge distillation, etc.). However, to the best of our knowledge, a few or no automated approach exists to explore, select and generate the best approximate version of a given DNN according to design objectives. The goal of this paper is to demonstrate that the design space exploration phase can enable significant network compression without noticeable accuracy loss. We demonstrate this via an example based on weight sharing and show that our direct conversion method can obtain a 4.85x compression rate with 0.14% accuracy loss in ResNet18 and 4.91x compression rate with 0.44% accuracy loss in SqueezeNet without involving retraining steps.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.