How to train your model dynamically using adversarial data

Hugging Face · 4 years ago · 14

Dynamic adversarial data collection (DADC) involves having humans create examples designed to fool current models, then retraining the model on these adversarial samples in repeated cycles. The approach was demonstrated on MNIST digit recognition, where a model initially achieved 89% accuracy on standard test data but failed on diverse human handwriting. By iteratively collecting human-generated adversarial examples and retraining, models improve generalization and become more aligned with real-world performance rather than saturating on static benchmarks.