Reproduce “On the Variance of Neural Network Training with respect to Test Sets and Distributions” on OpenResearch

About this paper

Keller Jordan

This work distinguishes between test-set variance and underlying distribution-wise variance in neural network training, demonstrating that converged models exhibit minimal true performance variability despite test-set fluctuations. It introduces an unbiased estimator for distribution-wise variance and a simple formula that accurately predicts test-set variance for binary classification.

Key insights

• Standard neural network trainings, despite exhibiting significant test-set variance, show minimal variance in performance on the underlying test-distributions (e.g., 0.033% standard deviation for CIFAR-10).
• Trained networks make approximately independent errors on their test sets as training converges, contributing to predictable test-set variance.
• Test-set variance is a predictable consequence of neural network training's class-calibration property, allowing for its accurate estimation via a simple formula for binary classification.

Summary by alphaXiv.