Reproduce “Generative Adversarial Minority Oversampling” on OpenResearch

About this paper

Sankha Subhra Mullick, Shounak Datta, Swagatam Das

A method called Generative Adversarial Minority Oversampling (GAMO) mitigates class imbalance in deep learning by employing a three-player adversarial game to generate targeted, hard synthetic minority samples. This approach consistently yields higher average class-specific accuracy and geometric mean scores across diverse datasets and generates more realistic images with lower Fréchet Inception Distance compared to existing techniques.

Key insights

• Improving deep learning classification for imbalanced data benefits most from synthetic samples generated strategically near class boundaries and peripheries.
• A convex generator, combined with adversarial training against both a classifier and a discriminator, can effectively produce high-fidelity, "hard" minority samples.
• The explicit discriminator (D) is essential for enforcing distribution fidelity, especially for complex or non-convex class distributions, preventing the generation of unrealistic or out-of-distribution samples.

Summary by alphaXiv.