A Multi-Mode Accelerator for Pruned Deep Neural Networks

Convolutional Neural Networks (CNNs) are constituted of complex, slow convolutional layers and memory-demanding fully- connected layers. Current pruning techniques can reduce memory accesses and power consumption, but cannot speed up the convolutional layers. In this paper, we introduce a pruning technique able to reduce the number of kernels in convolutional layers of up to 90% with negligible accuracy degradation. We propose an architecture to accelerate fully- connected and convolutional computations within a single computational core, with power$/$energy consumption below mobile devices budget. The proposed pruning technique speeds up convolutional computations by up to $ 6.9\times $, reducing memory accesses by the same factor.