New theory extends Gaussian universality and CGMT to dependent data settings

Researchers have extended the principles of Gaussian universality and the convex Gaussian min-max theorem (CGMT) to dependent data settings. This work demonstrates that Gaussian universality remains applicable to high-dimensional logistic regression even with block dependence, m-dependence, and certain mixing conditions. Additionally, a new CGMT framework has been developed to account for correlations in both covariates and observations. These advancements allow for a better understanding of the impact of data augmentation practices on asymptotic risk in deep learning. AI

IMPACT Provides a theoretical foundation for understanding data augmentation in deep learning under dependent data conditions.

RANK_REASON The cluster contains an academic paper detailing theoretical advancements in statistics. [lever_c_demoted from research: ic=1 ai=0.7]

Read on arXiv stat.ML →

• convex Gaussian min-max theorem (CGMT)
• data augmentation
• deep learning
• Gaussian universality
• high-dimensional logistic regression
• Matthew Esmaili Mallory

• paper
• other

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