Analysis of Calibration Requirements in Image Recognition Tasks with Time-Encoded ANNs

Power and area efficiency of time-encoded neural networks make them highly suitable for smart portable applications. The main limitation here is the high sensitivity against circuit impairments like mismatch effects or PVT variations, which are usually mitigated by means of calibration circuitry. The purpose of this manuscript is to propose a novel time-encoded unit for Artificial Neural Networks, and make an in-depth analysis on how these effects degrade the accuracy in classification tasks. As an application example, image recognition tasks are selected using MNIST dataset. It is checked that the proposed unit does not require calibration to compensate mismatch effects, can be implemented with small devices, and shows accuracy degradation due to PVT variations that can be overcome with simple calibration or by increasing the number of bits used for the inputs and outputs of the unit. With a very simple neural network structure, 94.1 %-accuracy is achieved with a worst case simulated TOPS/W value of 24, and an area of 0.039 mm<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> in a 65 nm CMOS design node, with a data representation of 5 bits in the inputs and outputs of the unit and 4 bits used for the weights.