Lee Sung-Tae, Lee Jong-Ho
Department of Electrical and Computer Engineering, ISRC, Seoul National University, Seoul, South Korea.
Front Neurosci. 2020 Sep 18;14:571292. doi: 10.3389/fnins.2020.571292. eCollection 2020.
A novel operation scheme is proposed for high-density and highly robust neuromorphic computing based on NAND flash memory architecture. Analog input is represented with time-encoded input pulse by pulse width modulation (PWM) circuit, and 4-bit synaptic weight is represented with adjustable conductance of NAND cells. Pulse width modulation scheme for analog input value and proposed operation scheme is suitably applicable to the conventional NAND flash architecture to implement a neuromorphic system without additional change of memory architecture. Saturated current-voltage characteristic of NAND cells eliminates the effect of serial resistance of adjacent cells where a pass bias is applied in a synaptic string and IR drop of metal wire resistance. Multiply-accumulate (MAC) operation of 4-bit weight and width-modulated input can be performed in a single input step without additional logic operation. Furthermore, the effect of quantization training (QT) on the classification accuracy is investigated compared with post-training quantization (PTQ) with 4-bit weight. Lastly, a sufficiently low current variance of NAND cells obtained by the read-verify-write (RVW) scheme achieves satisfying accuracies of 98.14 and 89.6% for the MNIST and CIFAR10 images, respectively.
提出了一种基于与非闪存架构的用于高密度和高鲁棒性神经形态计算的新型操作方案。模拟输入由脉宽调制（PWM）电路通过时间编码输入脉冲来表示，4位突触权重由与非单元的可调电导来表示。模拟输入值的脉宽调制方案和所提出的操作方案适用于传统与非闪存架构，以实现神经形态系统，而无需对存储器架构进行额外更改。与非单元的饱和电流 - 电压特性消除了在突触串中施加通过偏置时相邻单元的串联电阻以及金属线电阻的IR降的影响。4位权重和宽度调制输入的乘加（MAC）操作可以在单个输入步骤中执行，无需额外的逻辑操作。此外，与具有4位权重的训练后量化（PTQ）相比，研究了量化训练（QT）对分类精度的影响。最后，通过读 - 验证 - 写（RVW）方案获得的与非单元足够低的电流方差分别为MNIST和CIFAR10图像实现了98.14%和89.6%的令人满意的准确率。
