Bio-Inspired Spiking Neural Networks for Facial Expression Recognition: Generalisation Investigation

Facial expression recognition is a popular research topic to a wide range of applications in human-computer interaction, social robotics, and affective computing. Various attempts have been made to improve the techniques and accuracies of FER. However, one of the main challenges still persists - how to generalise across different datasets, deal with small data and reduce classifiers bias. In this paper, we explore the application of bio-inspired Spiking Neural Networks (SNN) with unsupervised learning using spike timing dependent plasticity (STDP) for FER. We have evaluated our approach on two publicly available, third-party, facial expression datasets. The results have shown that our approach has achieved consistently high accuracies (92%) in cross-dataset evaluation and exhibited a significant improvement compared with the state-of-the-art CNN and HOG feature extraction techniques. The results suggest that our approach can learn more effective feature representations, which lead to good generalisation across subjects in different ethnic groups with different facial dimensions and characteristics.