AdaAugment: A Tuning-Free and Adaptive Approach to Enhance Data Augmentation.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Data augmentation (DA) is crucial for enhancing deep model generalization.
• Current DA methods often use fixed or random augmentation magnitudes, leading to potential misalignment with model training status.
• This misalignment can increase risks of underfitting and overfitting.

Purpose of the Study:

• To introduce AdaAugment, a novel, tuning-free adaptive data augmentation method.
• To dynamically adjust augmentation magnitudes for individual training samples using real-time network feedback.
• To mitigate underfitting and overfitting by aligning augmented data with model training progress.

Main Methods:

• AdaAugment employs a dual-model architecture: a policy network and a target network.
• The policy network adaptively adjusts augmentation magnitudes via reinforcement learning.
• The policy and target networks are jointly optimized, with the target network training on adaptively augmented samples.

Main Results:

• AdaAugment consistently outperforms state-of-the-art data augmentation methods.
• The method demonstrates superior effectiveness across benchmark datasets and deep architectures.
• AdaAugment maintains remarkable computational efficiency during training.

Conclusions:

• AdaAugment offers an effective and efficient solution for adaptive data augmentation.
• The proposed method successfully addresses limitations of fixed/random augmentation strategies.
• AdaAugment enhances deep model generalization by intelligently adapting augmentation to training dynamics.