Coursera Flash Sale
40% Off Coursera Plus for 3 Months!
Grab it
This talk by Jingfeng Wu from UC Berkeley explores the regularization effects of early stopping in gradient descent for overparameterized logistic regression. Learn how gradient descent iterates behave in overparameterized settings, diverging in norm while converging in direction to the maximum â„“2-margin solution. Discover why early-stopped gradient descent achieves vanishing excess logistic risk and remains well-calibrated, while gradient descent at convergence becomes statistically inconsistent with diverging risk. Understand the significant statistical advantage of early stopping, which requires only polynomial sample complexity to achieve low excess zero-one risk, compared to the exponential samples needed for interpolating estimators. Examine the nonasymptotic connections between early-stopped gradient descent and explicit â„“2-regularization, providing theoretical insights into deep learning optimization dynamics and their statistical implications.
