The small-data regime in deep BCI

The defining constraint of deep learning for neural data is that labeled trials are few — typically tens to a few thousand per subject, bounded by how long a participant can perform a task — while deep networks are data-hungry. This produces a persistent, well-documented reality check: on many EEG benchmarks, compact architectures (EEGNet, ShallowConvNet) and non-deep pipelines (Riemannian tangent-space plus logistic regression, shrinkage LDA) are competitive with or superior to large deep models, and headline deep-learning gains often trace to cross-subject data, augmentation or pretraining rather than raw model capacity.

The practical implications are to prefer strong inductive biases (frequency and spatial factorization), heavy regularization, and cross-validation discipline that avoids leakage, and to treat any large network's apparent superiority as unproven until it is tested against these simple baselines on held-out subjects. The field's current frontier — self-supervision and foundation models — is precisely an attempt to escape this regime by importing unlabeled data.