Event-Related Potential Reveals Partial Face Cognitive Mechanisms through Machine Learning

Abstract

During the SARS-CoV-2 pandemic, wearing masks became a daily practice, and the cognitive mechanisms of how people correctly recognize a masked face are still questioned. In our previous study, we investigated the electroencephalogram evoked corresponding to the presented images, called the event-related potential, during the partial face cognition task and employed a machine learning model to interpret the cognitive activity. We found that the combination of the xDAWN spatial filter, covariance matrix, tangent space mapping, and support vector machine model performed the best among the six untuned models in the cognitive activity classification of target and non-target faces. However, the model faced difficulty clarifying the significance of each face component. To solve that problem, we implemented the previous method on a bigger dataset but tuned the parameter and achieved the highest accuracy of 0.728 in target classification when C = 0.1. Moreover, we could explain the importance of each face component as we found a sharp drop in accuracy from 0.794 in the full face cognition to 0.695 when the eyes were absent from the face image. These results imply that eyes provide crucial information in face cognition and could be promising in applying neuroscience-based cognitive face preferences.