Memristive Device Variability Performance Impact on Neuromorphic Machine Learning Hardware
Memristive Device Variability Performance Impact on Neuromorphic Machine Learning Hardware
复制标题
DOI：
10.1109/igsc51522.2020.9291114
复制
发表时间：
2020-10
期刊：
影响因子：
--
通讯作者：
Andrew J. Ford;R. Jha
中科院分区：
文献类型：
--
作者：
Andrew J. Ford;R. Jha
关于使用新型忆阻设备的机器学习算法的硬件实现的一个重要问题是所提出的架构对高设备可变性的恢复能力。我们发现大多数算法对可变权重更新和初始化具有令人惊讶的高容忍度。我们还提出了一种简单的方法，通过研究准确性与训练时间来验证基于忆阻 RRAM 交叉阵列的单层感知器 (SLP) 神经形态硬件。最后，我们展示了具有中间状态、衰减和高斯变异性的 RRAM 单元的高级模拟。
A vital issue regarding hardware implementations of machine learning algorithms with novel memristive devices is the concern of the proposed architecture's resilience to high device variability. We find that most algorithms have surprisingly high tolerance to variable weight updates and initializations. We also propose a simple method to validate Single Layer Perceptron (SLP) neuromorphic hardware based on memristive RRAM crossbar arrays by studying accuracy vs. training time. Finally, we show high level simulations of an RRAM cell with intermediate states, decay, and Gaussian variability.