Vol. 8, Special Issue 3 (2019)

Author(s):
Ajay Kumar, Yashwant Mali and Anubhav Kumar

Abstract:
Deep Learning (DL) models have demonstrated remarkable success across diverse domains, revolutionizing the landscape of artificial intelligence. One pivotal aspect influencing the performance of these models is the quality and quantity of training data. In recent years, Data Augmentation (DA) has emerged as a crucial strategy to enhance the robustness and generalization capabilities of DL models. This review paper delves into the theoretical foundations underpinning the role of Data Augmentation in the training of deep neural networks. The exploration commences with an elucidation of the fundamental principles of Data Augmentation, a technique involving the generation of synthetic data by applying various transformations to existing samples. A critical analysis of the theoretical frameworks governing the impact of augmented data on model training reveals profound insights into the regularization effects, enabling DL models to resist overfitting and better adapt to diverse real-world scenarios. Furthermore, this review investigates the interplay between Data Augmentation and the optimization landscape of deep learning. We delve into the theoretical constructs governing how augmented data influences the convergence behavior of optimization algorithms, shedding light on the intricate dynamics that shape the learning process. Keywords such as regularization, overfitting prevention, and optimization dynamics are central to understanding the nuanced relationships between Data Augmentation and model training.