Mental workload recognition from EEG signals via semi-supervised autoencoders

Mental workload-the cognitive effort to complete tasks-is vital in fields like system design, healthcare, and human-machine interaction. Supervised learning is often used for EEG-based workload recognition but is limited by scarce labeled data. To address this, we propose semi-supervised autoencoders that combine labeled and abundant unlabeled data. Our model integrates a supervised objective into an unsupervised autoencoder, forming a joint function that minimizes both reconstruction and prediction errors. This enhances discriminative power. To overcome vanishing/exploding gradients, we add skip connections between layers. Tested on two EEG datasets, our framework achieved high accuracy in binary mental workload classification.