A Method of Cross-Subject Transfer Learning for Ultra Short Time SSVEP Classification

The steady-state visual evoked potentials (SSVEP) based brain-computer interfaces (BCIs) require extensive training data for efficient classification, but existing algorithms struggle with ultra short time inputs (less than 0.2 seconds), limiting the feasibility of real-time systems. This paper proposes a novel method CSA-GSDANN consisting of CSA and GSDANN. GSDANN improves SSVEP feature extraction performance in ultra short time input scenarios by applying cross-subject transfer learning techniques, combining a Global Attention Mechanism (GAM) and an optimized SSVEPNet and pre-training method CSA selects the most suitable source subject based on accuracy and aligns it with the target subject to address the inter-subject variability. The proposed CSA-GSDANN method adopts a Domain Adversarial Neural Network (DANN) framework, which integrates an enhanced SSVEPNet algorithm with an attention mechanism to extract features from electroencephalogram (EEG) data within and across subjects. The extracted features undergo domain-adversarial transfer learning. The final stage involves frequency signal classification using a constrained convolutional network. The evaluation of the CSA-GSDANN method on the IMUT dataset containing 12 subjects shows significant improvements. A comparative analysis against eight mainstream deep learning and traditional algorithms demonstrates an average accuracy enhancement of 4.23% and an average Information Transfer Rate (ITR) improvement of 50.482 bits/min compared to state-of-the-art classification algorithms under short time (0.2s) EEG inputs, substantially improving SSVEP classification performance.