Learning to Compose Domain-Specific Transformations for Data Augmentation.

Area of Science:

• Machine Learning
• Computer Vision
• Natural Language Processing

Background:

• Data augmentation is crucial for enhancing machine learning model generalization by expanding training datasets.
• Manual creation and optimization of complex data transformation sequences for state-of-the-art augmentation are time-consuming and labor-intensive.
• Existing methods often struggle with arbitrary or non-deterministic transformations and require labeled data for tuning.

Purpose of the Study:

• To develop an automated method for learning sophisticated data augmentation strategies.
• To reduce the manual effort required in constructing effective data transformation compositions.
• To enable data augmentation for any downstream discriminative model using learned transformation sequences.

Main Methods:

• A generative adversarial approach is employed to learn a sequence model over user-defined transformation functions.
• The method utilizes unlabeled data for training, making it broadly applicable.
• It accommodates arbitrary, non-deterministic transformations and is robust to potential user input errors.

Main Results:

• Significant performance improvements were observed across various datasets and tasks.
• Achieved a 4.0 accuracy point increase on CIFAR-10 (image classification).
• Demonstrated a 1.4 F1 point improvement on the ACE relation extraction task (text).
• Showcased a 3.4 accuracy point gain on a medical imaging dataset using domain-specific transformations.

Conclusions:

• The proposed automated data augmentation method effectively learns complex transformation sequences.
• This approach surpasses standard heuristic augmentation techniques in performance across diverse data modalities.
• The learned transformation model offers a flexible and powerful tool for enhancing machine learning model training.