Eye Movement Detection Using Self-Supervised Deep Learning

Eye movement event detection is crucial to analyzing and understanding human visual information. While deep learning has demonstrated remarkable potential in eye movement detection, the effectiveness of fully supervised learning approaches in practical scenarios is often contingent on access to extensive datasets. Given the challenges of acquiring substantial and meticulously curated eye tracking data, we introduce a self-supervised deep learning technique for the task of detecting eye movement events. In order to take full advantage of the temporal sequential nature of eye tracking data, this work introduces a Contrastive Predictive Coding (CPC) framework in the pre-training of self-supervised to learn efficient representations for the eye movement classification. We adapt and tailor the CPC method to suit the specific needs and limitations inherent in eye movement detection. Subsequently, we carry out experimental assessments using the GazeCom dataset, a large-scale, publicly accessible resource with manual labeling. Our results show that the performance is comparable to or close to that of fully supervised methods. Moreover, our proposed method is able to enhance classification performance especially when limited labeled data is available, demonstrating the effectiveness of self-supervised learning in mitigating the problem of a lack of labeled data for detecting eye movements.