Reproduce “The Effectiveness of Data Augmentation in Image Classification using Deep Learning” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Luis Perez, Jason Wang
Researchers at Stanford University explored the efficacy of traditional, GAN-based, and a novel neural augmentation method for image classification in data-constrained settings. Neural augmentation, optimized by classification loss alone, achieved 0.915 accuracy on the 'Dogs vs. Goldfish' dataset, while traditional methods reached 0.775 on 'Dogs vs. Cats', consistently outperforming non-augmented baselines.
Key insights
- Data augmentation consistently improved classification accuracy across most tasks and methods, underscoring its general utility in combating data scarcity.
- Neural augmentation, particularly when optimized solely by the classification loss, demonstrated superior performance on visually distinct datasets and learned to generate regularized, feature-blended images.
- Traditional affine transformations remain highly effective, proving to be the most performant method for tasks involving visually similar classes, like 'Dogs vs. Cats'.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.