BCIwiki (bciwiki.com) — A deep convolutional network called SCR-FTSDNN maintains high target recognition accuracy for steady-state visual evoked potential (SSVEP)-based brain-computer interfaces (BCIs) even with limited calibration data, researchers led by East China University of Science and Technology reported in Cognitive Neurodynamics on October 2, 2026 (PMID 42824274; DOI 10.1007/s11571-026-10565-6).
SSVEP-based BCIs have drawn attention for their high information transfer rate and ability to recognize many targets. The authors note that deep learning methods show promise for SSVEP feature extraction and frequency classification, but their model complexity and reliance on large calibration datasets add to data-collection burden and can tire users, limiting practical use.
SCR-FTSDNN first converts raw EEG into the frequency domain with a Fourier transform, then applies temporal convolution to capture dynamic features. Spatial and channel reconstruction convolution modules strengthen discriminative feature extraction, while dilated convolution widens the receptive fields across frequency, spatial and temporal features, suppresses redundant information and weights key components more heavily.
The team ran systematic experiments on 2 public SSVEP datasets, Benchmark and BETA, comparing the network against several existing methods. The authors say SCR-FTSDNN achieved higher recognition accuracy and robustness, particularly with limited calibration data. The abstract does not report specific figures. Reproducibility code is available on GitHub.
The authors are Xiao Wu, Brendan Z. Allison, Weijie Chen, Xinjie He, Chongfeng Wang and Xingyu Wang, with affiliations including East China University of Science and Technology, the Cognitive Science Department at the University of California, San Diego, the Systems Research Institute of the Polish Academy of Sciences and Nicolaus Copernicus University. All results come from offline dataset tests, according to the paper in Cognitive Neurodynamics.