ApproxDNNFlow: An Evaluation and Exploration Framework for DNNs with Approximate Multipliers

Abstract

Widely used deep neural networks (DNNs) are proved error-tolerant, therefore accurate multipliers in DNNs can be replaced by approximate multipliers to reduce the power consumption. We set up a framework for training and evaluating DNNs based on approximate multipliers. Noisy training is proposed to adjust the parameters to tolerate the error caused by the approximate multipliers. Moreover, the framework can evaluate DNN accuracies with approximate multipliers. In the experiment, four approximate multipliers are evaluated. Based on the DNN inference results on MNIST and CIFAR10 by LeNet, the selected approximate multiplier can reach 99.17% and 65.76% accuracies respectively (the original accuracies are 99.27% and 74.88%) with significant reduction of the power consumption and area. In addition, the inference accuracies can be improved up to 99.21% and 69.5% by the proposed noise training methods. The proposed framework can contribute to the design of effective approximate computing for DNNs in the future.