In-sensor computing implemented by novel neuromorphic devices has been regarded as the potential technology to break the acquisition wall. Moreover, the nonlinear convolution inspired by the biological neural system outperforms the traditional linear convolution. Therefore, realizing the in-sensor nonlinear convolutional processing with the intrinsic nonlinearity of novel neuromorphic devices would be exciting and hardware friendly. Here, a new type of in-sensor nonlinear convolutional processing architecture is proposed based on STT-MTJ (spin transfer torque-magnetic tunnel junction) devices and simple peripheral circuits. The nonlinear dynamic characters of the STT-MTJ device are regulated by bias currents and magnetic fields. Meanwhile, the hybrid STT-MTJ/CMOS (complementary metal oxide semiconductor) circuit exhibits nonlinear dependence of the output voltage on bias currents and magnetic fields. In this way, a nonlinear convolutional unit is developed by the inherent property of hardware. Based on the in-sensor nonlinear convolutional unit, a convolutional network is simulated to perform the classification task, and a high accuracy is realized. This work indicates that the in-sensor nonlinear convolution offers a promising way to develop in-sensor computing at the edge intelligence devices.

中文翻译：

---

基于混合 MTJ/CMOS 阵列的传感器内非线性卷积处理

由新型神经形态设备实现的传感器内计算被认为是打破采集壁垒的潜在技术。此外，受生物神经系统启发的非线性卷积优于传统的线性卷积。因此，利用新型神经形态设备的固有非线性来实现传感器内非线性卷积处理将是令人兴奋的且硬件友好的。这里，提出了一种基于STT-MTJ（自旋转移扭矩-磁隧道结）器件和简单外围电路的新型传感器内非线性卷积处理架构。STT-MTJ器件的非线性动态特性由偏置电流和磁场调节。同时，混合STT-MTJ/CMOS（互补金属氧化物半导体）电路表现出输出电压对偏置电流和磁场的非线性依赖性。这样，利用硬件的固有特性开发了非线性卷积单元。基于传感器内非线性卷积单元，模拟卷积网络执行分类任务，实现了高精度。这项工作表明，传感器内非线性卷积为在边缘智能设备上开发传感器内计算提供了一种有前途的方法。