Neuromorphic circuits on segmented crossbar architectures with enhanced properties

General purpose processors have been used in a wide variety of computational and modeling applications. However, their performance is not always sufficient when simulating neural networks, which are widely applied to signal processing and pattern recognition. In this work, after a systematic study of the computational requirements of such neural networks and an exploration of the available hardware solutions through which the aforementioned applications can be accelerated, a modern neuromorphic circuit structure is proposed with its operation attributed to memristor devices and segmented crossbar architecture. By coupling these two technologies, neuromorphic circuits have been designed with high computational performance versus integration scale and power consumption. An Ex-Situ training paradigm based on the advantageous memristor segmented crossbar is proposed, using the MNIST dataset and resulting at 97% accuracy. At the same time, a novel memristor tuning method on 1D1M configuration has been developed, able to increase the memristor programming speed.