Contributed Talk: Benign Overfitting without Linearity: Neural Network Classifiers Trained by Gradient Descent for Noisy Linear Data

Speaker: Spencer Frei, UC Berkeley
Talk title: Benign Overfitting without Linearity: Neural Network Classifiers Trained by Gradient Descent for Noisy Linear Data

Time: Tuesday, April 5, 5:05pm-5:30pm (ET)

Abstract:
We characterize benign overfitting behavior in two-layer neural networks trained by gradient descent on the logistic loss when the data comes from well-separated class-conditional log-concave distributions. In contrast to previous work on benign overfitting that require linear or kernel-based predictors, our analysis holds in a setting where both the model and learning dynamics are fundamentally nonlinear.

Joint work with Niladri Chatterji and Peter Bartlett.

Return to workshop schedule