This paper analyzes the generalization error of min-norm interpolators in transfer learning, particularly when limited test samples are available during training. The research characterizes the bias and variance under covariate and model shifts, revealing that adding data can sometimes be detrimental at low signal-to-noise ratios. The findings suggest that transfer learning is beneficial at higher SNR levels if the shift-to-signal ratio remains below a specific threshold. The study also introduces novel technical tools in random matrix theory and universality analysis. AI
IMPACT Provides theoretical insights into model generalization, potentially informing future algorithm design.
RANK_REASON The cluster contains an academic paper on a theoretical machine learning topic. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
