Objective.This study quantifies how the accuracy of convolutional neural networks for electroencephalogram (EEG) classification depends on the amount of training data and evaluates parametric models for extrapolating performance to larger datasets.Approach.We evaluated the classification accuracy of three neural network architectures across three EEG classification tasks, systematically varying the number of subjects and the duration of EEG data per subject. We probed the cross-dataset variability of the learning curves across two datasets. Eight parametric models were assessed for their ability to fit and extrapolate learning curves, focusing on prediction error and uncertainty.Main results. Learning curve characteristics, such as slope, shape, and asymptotic performance, varied substantially between classification tasks but remained consistent across network architectures. Reliable extrapolation using scaling laws required data from several hundred subjects. The benefit from increasing the utilized recording length per subject plateaued after just a few seconds of EEG.Significance.EEG classification is often constrained by the limited availability of labeled data. Our findings provide task-specific learning curves to inform study design and demonstrate how extrapolated learning curves can support cost-benefit analyses for data acquisition in EEG research.
Seibel et al. (Sun,) studied this question.