Optimizing EEGNet for Code-Modulated Visual Evoked Potential Classification

Visual evoked potentials (VEPs) are widely used in EEG-based brain-computer interface (BCI) systems. Among these, code-modulated VEPs (c-VEPs) offer high communication speeds, support a large number of stimulus options and provide robust classification performance through the use of m-sequences. Although convolutional neural networks such as EEGNet have been successfully applied to frequency-based steady-state VEPs (SSVEPs), their performance on c-VEPs remains underexplored. In this study, an ablation analysis is conducted to evaluate architectural and training modifications aimed at improving EEGNet’s performance in c-VEP tasks. Moreover, a simple yet effective multi-cycle c-VEP classification method based on logit aggregation is proposed. Additionally, a calibration method that leverages artificially shifted m-sequences is introduced to construct multiclass training sets from a single sequence, thereby reducing data requirements. The methods are evaluated on three public datasets and one private dataset in a cross-subject setting. Results show that combining layer normalization, mixup augmentation, weight decay, and exponential moving average (EMA) improves EEGNet’s classification accuracy by up to 28.3%. The proposed logit aggregation method improves classification performance by up to 36.6%, achieving near-perfect accuracy depending on the number of cycles. Finally, training the model on artificially shifted training data results in a classification performance comparable to that of real data, demonstrating the potential of these simple modifications to significantly improve c-VEP classification using EEGNet.