Analog Circuit Implementation of Neural Networks for In-Sensor Computing

In-sensor computing not only saves hardware resources and power consumption, it also reduces the communication load inside the sensor network. Current in-sensor computation is mostly implemented with digital circuits performing binary operations; however, as sensor signals are analog in nature, it is more efficient to implement in-sensor computation and data processing directly with analog circuits performing analog computation. In this paper, we design the multiplication and accumulation (MAC) functions, which constitute the majority of operations in convolutional neural networks, in analog circuits. The MAC units are fabricated using SMIC 55nm CMOS process and the chips exhibit ideal linearity and accuracy in performing MAC operations. These chips are then used to construct the 3x3 kernel matrix convolution. Not only does the implementation use less transistors and silicon area, without the need of ADC, the direct analog computing is fast and instantaneous. A 3x3 pressure sensor array is used to test the analog kernel convolution. The experiment results demonstrate that the analog computation can accurately and consistently compute different pressure patterns and generate corresponding voltage levels directly from the analog inputs.