y0news
← Feed
Back to feed
🧠 AI NeutralImportance 7/10

Robust Fine-Tuning from Non-Robust Pretrained Models: Mitigating Suboptimal Transfer With Epsilon-Scheduling

arXiv – CS AI|Jonas Ngnaw\'e, Maxime Heuillet, Sabyasachi Sahoo, Yann Pequignot, Ola Ahmad, Audrey Durand, Fr\'ed\'eric Precioso, Christian Gagn\'e||5 views
🤖AI Summary

Researchers identified that fine-tuning non-robust pretrained AI models with robust objectives can lead to poor performance, termed 'suboptimal transfer.' They propose Epsilon-Scheduling, a novel training technique that adjusts perturbation strength during training to improve both task adaptation and adversarial robustness.

Key Takeaways
  • Fine-tuning non-robust pretrained models with robust objectives often leads to suboptimal transfer and poor performance.
  • The researchers introduced Epsilon-Scheduling, a training heuristic that schedules perturbation strength to promote optimal transfer.
  • A new metric called 'expected robustness' was developed to better evaluate the accuracy-robustness trade-off across different models.
  • Extensive experiments across six pretrained models and five datasets showed consistent improvements in expected robustness.
  • The findings address a significant knowledge gap in robust fine-tuning from widely available non-robust pretrained models.
Read Original →via arXiv – CS AI
Act on this with AI
Stay ahead of the market.
Connect your wallet to an AI agent. It reads balances, proposes swaps and bridges across 15 chains — you keep full control of your keys.
Connect Wallet to AI →How it works
Related Articles