Original title: On the Reliability of Generative Augmentation: A Wasserstein-Based Theoretical and Empirical Study
Chathurika S Abeykoon, Mathias Nthiani Muia, Mallory Goldstein · September 2026
When machine learning models train on uneven data, one category often has far fewer examples than the rest. To fix this, researchers use artificial intelligence to generate fake extra samples. However, nobody fully understood how this affects the model's accuracy on new data. The researchers built a mathematical framework to study this process. They treated adding fake data as mixing two sets of information: real and generated. They showed that errors depend on how much fake data is added and how closely it matches real data. Testing these ideas, they found that a model called Conditional WGAN-GP matched real data better than an older method called Conditional GAN. But better matching did not always lead to better test accuracy. Older, simpler techniques often performed just as well, showing that fake data quality cannot be judged by accuracy alone.
You're making a mixed fruit salad, but you have too many apples and not enough strawberries. You substitute some apple slices with strawberry-flavored gummy candies to even out the mix. In this work, adding gummy candies is like generating synthetic data for rare classes. The issue is that even if the gummies taste similar, they change the texture and overall quality. The paper shows that the reliability of this substitution depends on how closely the gummies mimic real strawberries, not just on having a balanced mix, and sometimes simply using more of the available fruit works better.
When training AI for rare events like disease detection, generating synthetic examples seems helpful. This work shows that the quality of those fake examples matters more than just having more data, which could change how we build reliable systems.
The abstract doesn't specify under what conditions generative augmentation does improve performance, as it focuses on cases where it doesn't always help. It also limits evaluation to Conditional GAN and WGAN-GP models, so results might not generalize to other generative approaches.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
On the Reliability of Generative Augmentation: A Wasserstein-Based Theoretical and Empirical Study — Chathurika S Abeykoon, Mathias Nthiani Muia, Mallory Goldstein
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.