Title: Deep Learning Theory: An Overview

Abstract : Deep learning has revolutionized modern machine learning, achieving impressive results in practice across diverse domains. Yet its theoretical foundations remain an active area of research. This talk explores three key aspects of (supervised) deep learning theory: approximation, optimization, and generalization. We first examine the expressive power of neural networks, highlighting their ability to approximate complex functions. Next, we discuss optimization and the challenges of gradient-based methods in high-dimensional, non-convex landscapes. Finally, we address generalization, investigating why deep networks perform well on unseen data despite overparameterization.