期刊
IEEE ACCESS
卷 10, 期 -, 页码 84946-84959出版社
IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC
DOI: 10.1109/ACCESS.2022.3196688
关键词
Domain wall; synapse; quantized weight; deep neural network; energy efficient; neuromorphic; in-memory computing
资金
- National Science Foundation (NSF) [ECCS 1954589, CCF 1815033]
- Virginia Commonwealth Cyber Initiative (CCI) CCI Cybersecurity Research Collaboration Grant
We demonstrate that extremely low resolution quantized synapses with large stochastic variations in synaptic weights can achieve high testing accuracies comparable to Deep Neural Networks (DNNs) with floating-point precision synaptic weights. We propose in-situ and ex-situ training algorithms based on modified algorithms and train 5-layer DNNs on MNIST dataset using 2-, 3- and 5-state DW devices as synapses. The highest inference accuracies obtained after in-situ and ex-situ training are close to 96.67% and 96.63% respectively, which is similar to the baseline accuracy of 97.1% obtained from DNN with floating-point precision weights. Our proposed approach demonstrates potential energy savings of at least two orders of magnitude compared to the floating-point approach implemented in CMOS. It is particularly attractive for low power intelligent edge devices.
We demonstrate extremely low resolution quantized (nominally 5-state) synapses with large stochastic variations in synaptic weights can be energy efficient and achieve reasonably high testing accuracies compared to Deep Neural Networks (DNNs) of similar sizes using floating-point precision synaptic weights. Specifically, voltage-controlled domain wall (DW) devices demonstrate stochastic behavior and can only encode limited states; however, they are extremely energy efficient during both training and inference. In this study, we propose both in-situ and ex-situ training algorithms, based on modification of the algorithm proposed by Hubara et al., 2017 which works well with quantization of synaptic weights, and train several 5-layer DNNs on MNIST dataset using 2-, 3- and 5-state DW devices as a synapse. For insitu training, a separate high precision memory unit preserves and accumulates the weight gradients which prevents accuracy loss due to weight quantization. For ex-situ training, a precursor DNN is first trained based on weight quantization and DW device model. Moreover, a noise tolerance margin is included in both of the training methods to account for the intrinsic device noise. The highest inference accuracies we obtain after in-situ and ex-situ training are similar to 96.67% and similar to 96.63%, respectively, which is very close to the baseline accuracy of similar to 97.1% obtained from a similar topology DNN having floating-point precision weights with no stochasticity. Large inter-state intervals due to quantized weights and noise tolerance margin enables in-situ training with significantly lower number of programming attempts. Our proposed approach demonstrates a possibility of at least two orders of magnitude energy savings compared to the floating-point approach implemented in CMOS. This approach is specifically attractive for low power intelligent edge devices where the ex-situ learning can be utilized for energy efficient non-adaptive tasks and the in-situ learning can provide the opportunity to adapt and learn in a dynamically evolving environment.
作者
我是这篇论文的作者
点击您的名字以认领此论文并将其添加到您的个人资料中。
推荐
暂无数据