IPF-RDA: An Information-Preserving Framework for Robust Data Augmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Data augmentation is crucial for deep model generalization but can introduce distribution shifts and noise.
• These issues limit the potential and performance of deep learning networks.

Purpose of the Study:

• To propose a novel information-preserving framework, IPF-RDA, to enhance the robustness of data augmentation.
• To address the limitations of current data augmentation techniques by preserving critical information and ensuring adaptive diversity.

Main Methods:

• Developed a class-discriminative information estimation algorithm to identify vulnerable data points and their importance scores.
• Introduced an information-preserving scheme to retain critical information in augmented samples adaptively.
• Categorized data augmentation methods and integrated them into the IPF-RDA framework.

Main Results:

• IPF-RDA consistently improves the performance of state-of-the-art data augmentation methods.
• The framework enhances the robustness and unleashes the full potential of data augmentation techniques.
• Demonstrated significant performance improvements across diverse datasets (CIFAR-10/100, Tiny-ImageNet, CUHK03, Market1501, Oxford Flower, MNIST) and deep models.

Conclusions:

• IPF-RDA is a simple yet effective framework for enhancing data augmentation robustness in deep learning.
• The proposed method improves generalization performance and scalability of deep models.
• IPF-RDA offers a promising approach to overcome limitations associated with data augmentation in AI.