<ul><li>Researchers have conducted a study on transfer learning to optimize downstream performance.</li><li>They introduced a simple linear model that utilizes a pretrained feature transform.</li><li>The researchers derived the exact asymptotics of the downstream risk and its fine-grained bias-variance decomposition.</li><li>The study revealed that the optimal featurization is naturally sparse and undergoes a phase transition from hard selection to soft selection of relevant features.</li></ul>

Understanding Optimal Feature Transfer via a Fine-Grained Bias-Variance Analysis

Discover more