MixupE: Understanding and Improving Mixup from Directional Derivative Perspective - Oral Session 1

Explore a 28-minute conference talk from the Uncertainty in Artificial Intelligence (UAI) 2023 Oral Session that delves into "MixupE: Understanding and Improving Mixup from Directional Derivative Perspective." Gain insights into the popular data augmentation technique called Mixup and its impact on training deep neural networks. Learn about the researchers' analysis of Mixup, revealing its implicit regularization of infinitely many directional derivatives of all orders. Discover the proposed improved version of Mixup, theoretically justified to deliver better generalization performance. Examine the experimental results across various domains, including images, tabular data, speech, and graphs, showcasing the effectiveness of the proposed method. Understand how the new approach improves upon vanilla Mixup, demonstrating a 0.8% increase in ImageNet top-1 accuracy. Access the presentation slides to follow along with the detailed explanations and visual aids provided by the researchers.