Bias–Variance Tradeoff

A decomposition of expected generalization error into squared bias (error from a model too rigid to capture the true function), variance (error from sensitivity to the particular training sample), and irreducible noise. Simple, strongly-biased models (LDA, few features) have low variance and generalize from tiny calibration sets; flexible models (deep nets, many features) have low bias but high variance and need far more data.

BCI lives at the small-sample end: calibration data are expensive (minutes of a fatiguing task), signals are noisy and nonstationary, so the variance term dominates. This is why heavily-biased, regularized linear methods remain competitive, and why the field's effort goes into controlling variance (shrinkage, cross-validation, priors) rather than chasing lower bias.