Which regularization technique is most known for performing implicit feature selection by shrinking some model coefficients to exactly zero?

L1 (Lasso) regularization adds a penalty proportional to the absolute value of the weights ( λΣ |w ᵢ |). The geometry of this penalty (a diamond-shaped constraint region in weight-space, with sharp corners on the axes) makes it likely that the optimal solution lands exactly on an axis, forcing some coefficients to exactly zero — effectively removing those features from the model. L2 regularization's circular constraint region, by contrast, almost never intersects the axes exactly, so it shrinks weights without eliminating them.