Artificial Electric Synapse of CuI-Based Memristor for Neuromorphic Emotion Recognition and Neural Networks

Abstract

Emotion classification is pivotal for advancing human-computer interaction, where it necessitates efficiently decoding complex dynamic signals. Traditional approaches, however, struggle to capture the temporal dependencies and nonlinear patterns intrinsic to emotional expressions. Herein, a novel CuI-based synaptic memristor is proposed, featuring reliable analog resistive switching and diverse biosynaptic plasticity, including EPSC, PPF, STM/LTM, LTP/LTD, and SRDP. Capitalizing on its nonlinear synaptic modulation capability, the developed neuromorphic reservoir computing system achieves an accuracy of 98.15% in speech emotion recognition on ESD data set, significantly outperforming traditional LSTM models. Moreover, the constructed fully connected neural network, employing its quasi-linear conductance modulation scheme for weight updates, achieves a recognition accuracy of 88.69% on the MNIST data set, a 13% improvement compared to the 75.16% accuracy obtained with nonlinear modulation. These findings validate the effectiveness of the CuI memristor in reservoir computing and neural network architectures, highlighting its potential as a core component of next-generation neuromorphic systems.