Double descent for least-squares interpolation on contaminated data: A simulation study
Researchers have investigated the "double descent" phenomenon in linear regression models when the training data is contaminated with outliers. Their simulation study compared the standard least-squares interpolation estimator with several robust alternatives. The findings indicate that even with contaminated data, highly overparametrized models can still exhibit double descent, leading to superior generalization performance compared to robust methods. AI
IMPACT This research explores the behavior of overparametrized models with noisy data, potentially informing the design of more robust machine learning systems.