Efficient Implementation of a Fully Analog Neural Network on a Reconfigurable Platform

This paper investigates the potential of floating gate field-effect transistors (FETs) as primitives for subthreshold computation in analog neural networks. By leveraging the inherent properties of these transistors, we demonstrate their suitability for constructing neural network activation functions, such as sigmoid and rectified linear units (ReLUs), as well as winner-take-all (WTA) circuits for softmax activation. Our end-to-end analog implementation successfully classifies the concentric circles problem, illustrating the advantages of maintaining an analog signal chain throughout the process.