2017/05/11 by Hokchhay Tann, Soheil Hashemi, Tann, Hokchhay +6
Computer Science · Engineering · #Advanced Neural Network Applications #Adversarial Robustness in Machine Learning #CCD and CMOS Imaging Sensors #FOS: Computer and information sciences #Neural and Evolutionary Computing (cs.NE)
paper · pdf · doi:10.48550/arxiv.1705.04288
openalex publication_date 2017/05/11 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28
While Deep Neural Networks (DNNs) push the state-of-the-art in many machine\nlearning applications, they often require millions of expensive floating-point\noperations for each input classification. This computation overhead limits the\napplicability of DNNs to low-power, embedded platforms and incurs high cost in\ndata centers. This motivates recent interests in designing low-power,\nlow-latency DNNs based on fixed-point, ternary, or even binary data precision.\nWhile recent works in this area offer promising results, they often lead to\nlarge accuracy drops when compared to the floating-point networks. We propose a\nnovel approach to map floating-point based DNNs to 8-bit dynamic fixed-point\nnetworks with integer power-of-two weights with no change in network\narchitecture. Our dynamic fixed-point DNNs allow different radix points between\nlayers. During inference, power-of-two weights allow multiplications to be\nreplaced with arithmetic shifts, while the 8-bit fixed-point representation\nsimplifies both the buffer and adder design. In addition, we propose a hardware\naccelerator design to achieve low-power, low-latency inference with\ninsignificant degradation in accuracy. Using our custom accelerator design with\nthe CIFAR-10 and ImageNet datasets, we show that our method achieves\nsignificant power and energy savings while increasing the classification\naccuracy.\n