Megadose The AI curve, compressed.

Generative Data Augmentation Without External Data: An Empirical and Theoretical Study on Deep Networks

· ArXiv · AI/CL/LG ·
Traditional augmentation matched most of the largest synthetic-data gains at the same number of optimizer steps.

The study tested four classifier backbones and four generators across six datasets, including two medical datasets, with generators trained only on each classifier’s training data. Synthetic samples helped most when classifiers detected little difference between real and generated data and the real-only classifier had room to improve. Larger detected differences were associated with smaller gains or worse accuracy. The authors’ theoretical bound balances the benefit of more training samples against the mismatch between real and generated distributions, but does not explain the size of the largest gains.

ArXiv · AI/CL/LG's note

score 4

Categories: Research