QIT Press - International Journal of Artificial Intelligence and Machine Learning Research and Development (QITP-IJAIMLRD)
Advancing Generalization in Deep Neural Networks through Theoretically Grounded Regularization and Geometric Optimization
Abstract
Deep neural networks have demonstrated powerful representational capabilities, yet their ability to generalize remains a central challenge in modern machine learning. This paper investigates the theoretical underpinnings of generalization by integrating principles of regularization and geometric optimization. We explore how norm-based constraints, implicit regularization via optimization paths, and flat minima contribute to better generalization bounds. By synthesizing insights from learning theory and empirical deep learning practices, we provide a principled perspective on scalable and robust model design.
Keywords
Document Preview
Details