TBMSCCN: Two-Branch Multi-Scale Convolutional Correlation Network for Steady-State Visual Evoked Potential...

Area of Science:

• Neuroscience and Artificial Intelligence
• Brain-Computer Interfaces (BCI)
• Signal Processing

Background:

• Steady-state visual evoked potential (SSVEP) based Brain-Computer Interfaces (BCIs) utilize artificial neural networks for improved target recognition.
• Current models often require extensive user calibration trials, leading to poor user experience and suboptimal performance with limited data.
• There is a need for efficient BCI models that perform well with minimal calibration.

Purpose of the Study:

• To propose a novel two-branch multi-scale convolutional correlation network (TBMSCCN) for SSVEP-based BCIs.
• To reduce model training parameters and enhance representation ability and convergence using correlation networks and SSVEP prior knowledge.
• To improve recognition performance in scenarios with limited or no user calibration trials.

Main Methods:

• Designed a multi-scale temporal convolution module for local temporal dependency learning within a parallel two-branch feature extraction framework.
• Constructed a contrastive loss function in the latent feature space to promote intra-class feature consistency and accelerate convergence.
• Employed a group convolution module as a decision layer to decrease network parameters and learn discriminative features between target and non-target stimuli.

Main Results:

• The TBMSCCN method outperformed existing models like TRCA, eTRCA, DNN, Conv-CA, and Bi-SiamCA in individual calibration scenarios, achieving high Information Transfer Rates (ITRs).
• Achieved average ITRs of 378.03 ± 139.18 bit/min on the "Benchmark" dataset and 198.92 ± 111.27 bit/min on the "Beta" dataset.
• Outperformed FBCCA, ttCCA, EEGNet, and TST-CFSR in calibration-free scenarios and demonstrated real-world effectiveness in an online Chinese spelling experiment.

Conclusions:

• The proposed TBMSCCN model offers low parameter count and strong robustness for SSVEP-based BCIs.
• The method effectively addresses the challenge of limited calibration trials, enhancing user experience and BCI performance.
• The TBMSCCN facilitates practical engineering applications of SSVEP-based BCI systems.