A Hybrid Spiking Recurrent Neural Network on Hardware for Efficient Emotion Recognition

In recent years, neuromorphic engineering based on spiking neural networks (SNNs) for real-time and low-power artificial intelligence (AI) tasks has attracted a lot of interest. However, most of the previous implementations on hardware of these algorithms concentrate on traditional feedforward fully-connected/convolutional neural network (CNNs) architectures which are used for vision image processing. Their applications in temporal text tasks using recurrent neural networks (RNNs) is less discussed. In this paper, we point out main difficulties of RNNs implementation on conventional neuromorphic systems and propose a hardware-oriented spiking RNN architecture for emotion recognition, which absorbs the external dynamics of traditional RNN and internal dynamics of SNN. Experimental results on two emotion recognition datasets show our spiking RNNs achieve comparable accuracies with other deep learning models and efficient run-time performance.