Tight Bounds for Logistic Regression with Large Stepsize Gradient Descent in Low Dimension
Researchers have developed tighter bounds for logistic regression using gradient descent in low-dimensional settings. The study focuses on binary classification with separable data, analyzing the optimization problem with a budget of T iterations. The findings provide an improved rate by analyzing the time it takes for gradient descent to transition from unstable to stable states, offering a fine-grained analysis of its oscillatory dynamics. AI