New research identifies 'catastrophic overfitting' in AI training and proposes a fix

Researchers have identified a phenomenon called catastrophic overfitting (CO) in fast adversarial training (FAT) of neural networks. CO occurs when robustness to multi-step attacks rapidly degrades, even as the model continues to perform well on single-step attacks used during training. The study reveals that this is due to a subtle 'weight drift' that the standard training loss is blind to. An exponential moving average (EMA) of the weights, however, remains robust. The paper proposes a simple fix: resetting the weights to their EMA when validation robustness drops, which effectively eliminates CO and achieves state-of-the-art accuracy on benchmarks like CIFAR-100 with reduced training cost. AI

IMPACT Identifies a critical training failure mode and offers a simple, effective mitigation strategy that improves robustness and efficiency.

RANK_REASON The cluster contains a research paper detailing a new phenomenon and proposed solution in AI training.

Read on arXiv cs.LG →

• arXiv
• AutoAttack
• catastrophic overfitting
• CIFAR-100
• DagsHub
• Fast Adversarial Training
• Hugging Face
• AutoAttack accuracy
• coupling-based base method
• exponential moving average
• multi-step attacks
• single-step attack
• training loss

• paper
• model release

AI-generated summary · Google Gemini · from 2 sources. How we write summaries →