Researchers have proposed a new method for data augmentation called Learning-State-Aware Dynamic Generative Data Augmentation (LSADA). LSADA aims to improve the performance of image classification models on small-scale datasets by adaptively adjusting the strength and type of augmentations applied to each sample. The approach involves constructing a learning state for each sample based on its current loss and loss-decrease rate, which is then used to determine the augmentation strength. Additionally, LSADA introduces a decoupled data augmentation and diffusion fusion strategy that applies transformations to class-relevant regions while generating diverse class-irrelevant regions. Experiments on nine public datasets show that LSADA outperforms existing methods by an average of 4.5% on natural image datasets and 2.5% on medical image datasets. --- Why it matters: This matters because it addresses the limitations of current data augmentation methods, which often struggle to adapt to small-scale datasets. By providing a more adaptive and dynamic approach, LSADA can improve the performance of image classification models in these scenarios. Source: https://arxiv.org/abs/2608.18907