A Comprehensive Theory of Deep Learning
A Comprehensive Theory of Deep Learning Table of Contents Preface Motivation The Need for Theory Scope and Audience 1. Introduction: Demystifying Deep Learning The Evolution of Machine Learning Perceived Mysteries and Myths Framing the Theoretical Challenge 2. The Geometric Field Perspective Neural Networks as High-Dimensional Dynamical Fields Attractors, Phase Transitions, and Topological Structures Semantic Clouds and Representation Manifolds 3. Generalization: From Soft Biases to Universality Inductive Biases: Hard vs. Soft PAC-Bayes, Compression, and Classical Bounds Double Descent, Benign Overfitting, and Their Interpretation Universality and Mode Connectivity 4. Representation Learning and Semantic Structure Adaptive Bases and Feature Construction Semantic Cloud Geometry: Modularity, Entanglement, Disentanglement The Role of Finsler Manifolds and Lattice Resonance Analogies, Transfer, and Compositionalit...