Geometric Insights on Deep Learning Models

This lecture explores the success of deep learning models in various tasks, focusing on image classification. It delves into the structure of convolutional neural networks, the challenges of classification, and the vulnerability of classifiers to perturbations. The instructor discusses the concept of adversarial perturbations, universal perturbations, and the geometric transformations that impact deep networks' behavior. The lecture also covers the importance of robustness and interpretability in deep learning systems, presenting tools and algorithms to analyze the geometry of deep networks and improve their performance.