Compact Artificial Neural Network Based on Task Attention for Individual SSVEP Recognition With Less Calibration

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Engineering

Background:

• Artificial neural networks (ANNs) show promise for steady-state visual evoked potential (SSVEP) target recognition.
• High trainable parameters in ANNs necessitate extensive calibration data, posing a challenge due to costly EEG collection.
• Overfitting is a concern in individual SSVEP recognition using ANNs.

Purpose of the Study:

• To design a compact ANN for individual SSVEP recognition that avoids overfitting.
• To reduce the number of trainable parameters in ANNs for SSVEP recognition.
• To minimize the need for extensive calibration data in SSVEP detection.

Main Methods:

• Integrated prior knowledge of SSVEP recognition into an attention neural network design.
• Applied an attention layer to translate spatial filtering operations into the ANN structure, reducing layer connections.
• Incorporated SSVEP signal models and shared weights as constraints to condense trainable parameters.

Main Results:

• The proposed compact ANN structure effectively eliminated redundant parameters in simulations.
• Reduced trainable parameters by over 90% compared to deep neural network (DNN) methods and 80% compared to correlation analysis (CA) methods.
• Improved individual recognition performance by at least 57% over DNNs and 7% over CA methods.

Conclusions:

• Integrating task-specific prior knowledge into ANNs enhances effectiveness and efficiency.
• The proposed compact ANN offers reduced parameters, requiring less calibration while maintaining prominent individual SSVEP recognition performance.
• This approach addresses the data acquisition bottleneck in developing personalized SSVEP-based brain-computer interfaces.