Convolutional neural networks (CNNs) offer potentially a better accuracy alternative for conventional deep learning tasks. The hardware implementation of CNN functionalities with conventional CMOS based devices still lags in area and energy efficiency. This has necessitated the investigations of unconventional devices, circuits, and architectures to efficiently mimic the functionality of neurons and synapses for neuromorphic applications. Spin-orbit torque magnetic tunnel junction (SOT-MTJ) device is capable of achieving energy and area efficient rectified linear unit (ReLU) activation functionality. This work utilizes the SOT-MTJ based ReLU for activation and max-pooling in a single unit to eliminate the need of dedicated hardware for pooling layer. Moreover, 2 × 2 multiply-accumulate-activate-pool (MAAP) is implemented by using four activation pairs each of which is fed by the crossbar output. The presented approach has been used to implement various CNN architectures and evaluated for CIFAR-10 image classification. The number of read/write operations reduce significantly by 2X in MAAP based CNN architectures. The results show that the area and energy in MAAP based CNN is improved by at least 25% and 82.9%, respectively, when compared with conventional CNN designs.

中文翻译：

---

CNN 中神经形态函数的节能自旋实现

卷积神经网络 (CNN) 为传统的深度学习任务提供了一种潜在的更准确的替代方案。使用基于传统 CMOS 的设备实现 CNN 功能的硬件在面积和能效方面仍然落后。这就需要对非常规设备、电路和架构进行研究，以有效地模拟神经元和突触的功能，用于神经形态应用。自旋轨道力矩磁隧道结 (SOT-MTJ) 器件能够实现能量和面积高效的整流线性单元 (ReLU) 激活功能。这项工作利用基于 SOT-MTJ 的 ReLU 在单个单元中进行激活和最大池化，以消除对池化层专用硬件的需求。而且，2 × 2 乘法累加激活池 (MAAP) 是通过使用四个激活对实现的，每个激活对由交叉开关输出提供。所提出的方法已用于实现各种 CNN 架构，并针对 CIFAR-10 图像分类进行了评估。在基于 MAAP 的 CNN 架构中，读/写操作的数量显着减少了 2 倍。结果表明，与传统 CNN 设计相比，基于 MAAP 的 CNN 的面积和能量分别提高了至少 25% 和 82.9%。