New AI Trick Could Make Healthcare Predictions Faster and Cheaper

A new paper proposes a neighboring early-stopping rule that adaptively selects regularization for kernel ridge regression with random features—without needing to know the underlying smoothness or capacity parameters. The method compares only adjacent estimators on a uniform inverse-regularization grid, cutting the number of discrepancy checks compared to standard all-pairs procedures, and computes everything directly in the random feature space without forming the full kernel matrix. The authors prove the selected estimator reaches the oracle polynomial learning rate up to logarithmic factors, under standard source and capacity conditions, and works for both well-specified and partially misspecified settings. Simulations and real-data experiments demonstrate its prediction performance and computational behavior.