Transfer learning for GAN-based EEG generation: Fine-tuning a pretrained model for mind-wandering with limited data

Mind-wandering (MW) studies often suffer from extremely small datasets, making it difficult to model condition-specific electroencephalography (EEG) and motivating data augmentation. In these MW studies, EEG is recorded to distinguish between the MW and on-task (OT) state. We examined whether a generative adversarial network (GAN) pretrained on data-rich on-task trials can be adapted to generate physiologically plausible MW EEG under severe data scarcity. To compare the effects of different types of pretraining, we used an event-related potential (ERP)-based Wasserstein GAN with gradient penalty (WGAN-GP) pretrained on OT trials and compared six training approaches: from-scratch training, a no-training baseline that directly reused the OT-pretrained model, full fine-tuning, and three partial fine-tuning variants that updated selected generator and discriminator layers. Models were evaluated using ERP waveforms and power spectra, EEG features (band power, entropy, P3 amplitude and latency, Petrosian fractal dimension, Hjorth complexity), distributional visualization based on principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), kernel density estimation (KDE), distance- and divergence-based metrics, and training time and number of trainable parameters. Full fine-tuning produced the overall closest match to real MW trials across waveform-, and distribution-level measures on the held-out test set, whereas partial fine-tuning achieved competitive feature-level similarity. In contrast, the no-training baseline kept OT-like rather than MW-specific patterns, and from-scratch training failed to learn MW-like characteristics. Therefore, adapting an OT-pretrained ERP-WGAN-GP model, especially with full or partial fine-tuning approaches, is more effective for MW EEG trial generation than training from scratch or reusing the source model without adaptation.