Memristor-based Deep Spiking Neural Network with a Computing-In-Memory Architecture
Memristor-based Deep Spiking Neural Network with a Computing-In-Memory Architecture
复制标题
DOI：
10.1109/isqed54688.2022.9806206
复制
发表时间：
2022-04
期刊：
影响因子：
--
通讯作者：
Fabiha Nowshin;Y. Yi
中科院分区：
文献类型：
--
作者：
Fabiha Nowshin;Y. Yi
峰值神经网络（snn）是一种人工神经网络模型，在实现深度学习应用时，在功率和能量方面表现出显着的优势。然而，机器学习应用的数据密集型特性给神经网络的实现带来了一个具有挑战性的问题，如延迟、能源效率和内存瓶颈。因此，我们引入了可扩展的深度SNN来解决延迟和能量效率问题。我们集成了一个内存中计算（CIM）架构，该架构由一个预制的忆阻器交叉棒阵列构建，以减少向量矩阵乘法中的内存带宽，这是深度学习的关键操作。通过对输入信号采用尖峰间间隔（ISI）编码方案，我们证明了我们设计的体系结构的时空信息处理能力。所述忆阻器横条阵列具有增强的散热层，可将所述忆阻器的电阻变化降低约30%。我们进一步开发了一种首次尖峰时间（TTFS）方法来对输出进行分类。所设计的电路和架构可以在数字识别和MNIST数据集上实现非常高的精度。我们的架构可以对手写数字进行分类，而功耗仅为2.9mW，推理速度为2μs/图像。每个突触连接的能量仅为2.51pJ，适合应用于深度学习加速器。
Spiking Neural Networks (SNNs) are artificial neural network models that show significant advantages in terms of power and energy when realizing deep learning applications. However, the data-intensive nature of machine learning applications imposes a challenging problem to neural network implementations in terms of latency, energy efficiency and memory bottleneck. Therefore, we introduce a scalable deep SNN to address the problem of latency and energy efficiency. We integrate a Computing-In-Memory (CIM) architecture built with a fabricated memristor crossbar array to reduce the memory bandwidth in vector-matrix multiplication, a key operation in deep learning. By applying an inter-spike interval (ISI) encoding scheme to the input signals, we demonstrate the spatiotemporal information processing capability of our designed architecture. The memristor crossbar array has an enhanced heat dissipation layer that reduces the resistance variation of the memristors by ~30%. We further develop a time-to-first-spike (TTFS) method to classify the outputs. The designed circuits and architecture can achieve very high accuracies with both digit recognition and the MNIST dataset. Our architecture can classify handwritten digits while consuming merely 2.9mW of power with an inference speed of 2μs/image. Only 2.51pJ of energy per synaptic connection makes it suitable to apply in deep learning accelerators.