
The optimal ridge regularization level and the corresponding optimal risk can exhibit surprising behavior, such as negative regularization and non-monotonic risk profiles, especially in the out-of-distribution setting where the test distribution deviates from the training distribution.


coremsg

optimal-ridge-regularization-for-out-of-distribution-prediction-characterizing-the-behavior-of-regularization-and-risk


Optimal Ridge Regularization for Out-of-Distribution Prediction: Characterizing the Behavior of Regularization and Risk
