L2 regularization is also called?

L2 regularization is also called ridge, after ridge regression: least squares with an L2 penalty, minimizing ‖y − Xw‖² + λ Σ wᵢ². Neural networks call it weight decay; numerical analysis calls it Tikhonov regularization. It shrinks every weight toward zero but, unlike lasso, almost never sets one exactly to zero.

Arthur Hoerl and Robert Kennard introduced ridge regression in Technometrics in 1970. Its closed form, ŵ = (XᵀX + λI)⁻¹ Xᵀy, adds λ to the diagonal so the matrix stays invertible even with collinear features. Example: if a feature appears twice, least squares cannot choose between weights of 2 and 0 or 102 and −100, which fit equally well. Ridge picks the equal split, a little under 1 each, because that has the smallest penalty.

A common follow-up: are weight decay and L2 the same? With plain SGD they are equivalent; with Adam they are not, which is why AdamW applies the decay separately. Scale features first, leave the intercept unpenalized and choose λ by cross-validation.