A Novel Architecture for Early Detection of Negative Output Features in Deep Neural Network Accelerators
Mohammadreza Asadikouhanjani, Seok‐Bum Ko · IEEE Transactions on Circuits & Systems II Express Briefs · 2020
Deep Neural Networks (DNNs) perform billions of computations in applications such as image classification, object detection, and image segmentation. In most well-known DNNs, an activation function follows a convolutional or a fully connected layer. Several popular activation functions involve setting all negative inputs to zero. In this brief, we propose a novel architecture in which the activation function is merged with the prior computational layer. Our proposed dataflow can reduce the computations needed to generate a specific output feature in convolutional and fully connected layers by reordering the computation to achieve early detection of the sign of the output feature. In addition, our computation engine supports zero computation skipping, which further accelerates the layer computation. In this brief, we build on a state-of-the-art DNN accelerator and modify it to perform early detection of negative output features. When compared to the original design, our method achieves a speedup of ×2.19 and reduces energy consumption by ×1.94. The average reduction in the number of multiply-accumulate (MAC) operations is 10.64% and the average reduction in the number of the load operations is 3.86%. These improvements are achieved while maintaining classification accuracy in two popular benchmark networks.