Overview

Data augmentation creates modified or synthetic training examples so models can learn from limited, imbalanced, or varied data. Common approaches include geometric and appearance transforms, Mixup and CutMix, graph and feature perturbations, and examples generated with diffusion models, language models, or simulation. Recent work increasingly adapts augmentation to a model’s learning state and to domain-specific constraints in settings such as medical imaging, speech, and robotics. The central tension is to add useful diversity without changing labels or important structure; effectiveness depends on the task and should be tested for out-of-domain performance, calibration, and synthetic-data quality as well as accuracy.

Latest papers

CardsList

273 papers29 papers in the last four weeks, up 16% on the four weeks before. 0.2% of all new papers.

  1. Exploring specialization and sensitivity of convolutional neural networks under simultaneous image augmentations

    Mar 5, 2025Pavel Kharyuk, Sergey Matveev, Ivan OseledetsNeural Network InterpretabilityConvolutional Neural Networks

  2. Source-Free Domain Adaptive Object Detection with Semantics Compensation

    Oct 7, 2024Song Tang, Jiuzheng Yang, Mao Ye +3Source-Free Domain AdaptationData Augmentation

  3. VITA: A Multi-Source Vicinal Transfer Augmentation Method for Out-of-Distribution Generalization

    Apr 25, 2022Minghui Chen, Cheng Wen, Feng Zheng +2Synthetic Data AugmentationImage Corruption Robustness