The lasso
Tibshirani proposed penalising the sum of absolute regression coefficients, which forces some of them to become exactly zero.
Why it matters
Feature selection stopped being a separate step: the model discards what it does not need during training.
Unlike ridge regression, which only shrinks coefficients, an absolute-value penalty gives a sparse solution. That is regularisation and interpretability at once: one can see which features the model considers necessary. The method later took hold where there are more features than observations, in genomics and text; the 1996 paper does not treat that case, and Tibshirani in 2011 notes that least squares is undefined when p > N, which was not a hot topic in 1996.