A Technical Odyssey of Self-Supervised Representation Learning for Devanagari-Script-Based P300 Speller

Traditional supervised learning (SL) methods for P300 event-related potential (ERP) detection in P300 spellers require extensive labelled data and often struggle to generalize well across subjects and trials, especially with limited data. Previous efforts using transfer learning and knowledge distillation improved performance but still face high computational complexity and lack transparency. These issues highlight the need to explore new approaches to enhance transferability and reduce uncertainty. To address this, we investigated the effectiveness of representational learning through a self-supervised approach. Our self-supervised learning (SSL) framework, featuring a compact convolutional neural network (CNN) backbone and label-agnostic characteristics, improves the robustness of learned features to variations in ERPs encountered in P300 speller. Experiments on self-recorded data and ablation studies show that the learned representations are robust and effective. Achieving an accuracy of 84%, the downstream classifier trained on the SSL framework performed competitively with traditional supervised methods. Additionally, comparison between features learned with SL and SSL, using t-SNE visualization and correlation coefficient (r = -0.51) analysis, demonstrates that SSL features offer better discrimination between P300 and non-P300.