تحليل پتانسيل‌هاي برانگيخته بينايي حالت پايدار ناشي از تحريك فليكر با استفاده از يادگيري عميق

Steady-state visual evoked potential (SSVEP) brain-computer interfaces (BCIs) achieve some of the highest information transfer rates (ITR) reported for non-invasive neural communication, but the classical spatial-filtering methods that deliver that performance depend on subject-specific calibration data. Deep learning offers a calibration-free alternative, yet published results are hard to compare because studies differ in dataset, channel montage, window length an‎d eva‎luation protocol. This thesis presents a controlled comparison of six SSVEP recognition methods on the 40-class Benchmark SSVEP Dataset (35 subjects). Three classical algorithms are eva‎luated on a nine-channel parieto-occipital montage using leave-one-block-out cross-validation: canonical correlation analysis (CCA) an‎d filter bank CCA (FBCCA), which are calibration-free, an‎d task-related component analysis (TRCA), which requires a per-subject calibration set. Three deep learning models, EEGNet, SSVEPNet an‎d AE-EEGNet, a variant of EEGNet augmented with a squeeze-an‎d-excitation (SE) channel-attention block, are eva‎luated cross-subject using leave-one-subject-out cross-validation on the full 64-channel montage. Rather than presenting a new architecture, the objective is to test empirically whether channel attention improves on EEGNetʹs existing convolutional design under a strict cross-subject protocol. All methods are assessed at matched stimulation window lengths of 0.5 s, 1.0 s an‎d 2.0 s, using classification accuracy an‎d a common Wolpaw ITR formulation. TRCA reaches the highest accuracy of any method tested, 95.4% at the 2.0 s window, against 89.1% for FBCCA an‎d 77.9% for CCA, an‎d is also the only method given calibration data from the test subject. Among the cross-subject deep learning models, SSVEPNet is strongest at every window length, reaching 75.20% at 2.0 s, compared with 54.57% for EEGNet an‎d 54.86% for AE-EEGNet. No statistically significant difference between AE-EEGNet an‎d EEGNet was detected at any window length at the tested configuration (paired t-test p = 0.907, 0.935 an‎d 0.741; |Cohen’s dz| ≤ 0.056), an‎d the 95% confidence intervals bound any benefit of the SE block at no more than about one, one an‎d a half, an‎d two accuracy points at 0.5 s, 1.0 s an‎d 2.0 s respectively. This null result is examined mechanistically rather than set aside, an‎d four falsifiable explanations are offered for why channel attention fails to help in this setting. The comparison indicates that the choice between classical an‎d deep learning methods depends less on raw accuracy than on whether subject-specific calibration data is available, an‎d that a null result reported at this level of statistical an‎d mechanistic detail narrows the design space for future attention-based SSVEP architectures.

احسان روحاني , امير اخوان بي تقصير

محمدرضا احمدزاده , جعفر قيصري