An exploration of three approaches to automating data augmentation for machine learning

Abstract

Data Augmentation is an effective way to improve the generalization of deep learning models. However, current augmentation methods primarily rely on manual operations, such as flipping and cropping for image data. These methods are often designed based on human expertise or trial and error. Meanwhile, Automated Data Augmentation (AutoDA) is a promising research direction that treats the data augmentation process as a learning task and discovers the most effective ways to augment the data. In this thesis, we examine recent AutoDA research and introduce four new methods to enhance AutoDA using composition-based, mixing-based, and generation-based approaches. Existing composition-based AutoDA methods apply a fixed policy to the entire dataset and mainly focus on image classification tasks. In response, we propose AdaAug to learn class-and instance-adaptive augmentation policies and MODALS to learn modality-agnostic latent space augmentation strategies. For mixing-based AutoDA, we introduce TRANSFORMMIX to learn transformation and mixing augmentation strategies from data. Compared to previous mixing-based methods, TRANSFORMMIX takes into account the saliency information of the input images and produces more compelling mixed images that contain accurate and important information for the target tasks. Regarding generation-based AutoDA, we present AUTOGENDA to address imbalanced classification tasks. Specifically, AUTOGENDA identifies and transfers label-invariant changes across data classes through image captions and text-guided generative models. We also propose an automated search strategy to adapt the AUTOGENDA augmentation to each data class, leading to better generalization.

Date of Award	2025
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Dit Yan YEUNG (Supervisor)