All Hardware-Based Two-Layer Perceptron Implemented in Memristor Crossbar Arrays

Abstract

Implementing synaptic weights using tunable conductance of memristors offers significant power and computing throughput improvements through in-memory analog computing. Hardware-based perceptron implementation with memristor crossbar arrays has attracted increased interest in recent years. However, all the previous memristor-based perceptron demonstrations perform some critical operations such as the activation functions using software, leading to substantial back-and-forth communication between the perceptron and a computer. In this work, we show that by implementing the activation functions between different layers of a perceptron all on hardware, using only analog components, we can avoid those unnecessary communication and improve power efficiency and throughput. We have designed a compact multichannel rectified linear unit activation function and developed a two-layer perceptron using two individual memristor crossbar arrays. We have achieved a 93.63% recognition accuracy in classifying the Modified National Institute of Standards and Technology dataset using the analog neurons, comparable with that for a partially software counterpart, and a much improved power efficiency.

Single-Shot Optical Neural Network

“…An increase in source and detector efficiencies [46] could further reduce the energy consumption by an order of magnitude to the single-femtojoule per MAC regime. DAC and ADC costs may also be eliminated with analog electronic nonlinearities [22,58], as analog output activations from one layer can be directly used as inputs to the next layer. The inputs to the first layer may also be in the analog domain, e.g., as they are read out from a sensor.…”