Ternary Neural Networks for Resource-Efficient AI Applications. (arXiv:1609.00222v1 [cs.LG])

The computation and storage requirements for Deep Neural Networks (DNNs) are usually high. This issue limit their deployability on ubiquitous computing devices such as smart phones or wearables. In this paper, we propose ternary neural networks (TNNs) in order to make deep learning more resource-efficient. We train these TNNs using a teacher-student approach. Using only ternary weights and ternary neurons, with a step activation function of two-thresholds, the student ternary network learns to mimic the behaviour of its teacher network. We propose a novel, layer-wise greedy methodology for training TNNs. During training, a ternary neural network inherently prunes the smaller weights by setting them to zero. This makes them even more compact thus more resource-friendly. We devise a purpose-built hardware design for TNNs and implement it on FPGA. The benchmark results with our purpose-built hardware running TNNs reveal that, with only 1.24 microjoules per image, we can achieve 97.76% accuracy with 5.37 microsecond latency and with a rate of 255K images per second on MNIST.

from cs.AI updates on arXiv.org http://ift.tt/2bXxcY8
via IFTTT