New PAC-Bayesian framework explains adversarial training overfitting

Researchers have developed a new PAC-Bayesian analytical framework to understand the phenomenon of robust overfitting in adversarial training. By modeling adversarial training with momentum SGD as a discrete-time dynamical system, the framework provides time-resolved robust generalization bounds. This approach connects a model's robust generalization performance to factors like learning rate, local loss geometry, and mini-batch stochastic gradients, offering insights into the underlying mechanisms of robust overfitting and suggesting methods to improve generalization. AI

IMPACT Provides a theoretical framework to improve adversarial training robustness and mitigate overfitting in machine learning models.

RANK_REASON The cluster contains a research paper detailing a new analytical framework for understanding machine learning dynamics. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• alphaXiv
• arXiv
• CatalyzeX
• DagsHub
• Gotit.pub
• Hugging Face
• IArxiv
• Influence Flower
• PAC-bayesian learning
• ScienceCast
• SGD
• Yuelin Xu

• paper
• safety

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →