Measurement uncertainty and machine learningAnnotated2025

Why Machine Learning Models Systematically Underestimate Extreme Values

Y.-S. Ting

The Open Journal of Astrophysics

Show overview ▸

Measurement noise in the inputs can compress predictions from error-ignorant regression. More training data improves precision without removing this bias; correlated spectral features help, but their protection weakens with noise, temperature variation and lower resolution.

Open PDFLoading page…
Why Machine Learning Models Systematically Underestimate Extreme Values — annotated page 1

Annotify adds explanatory annotations and fades the original page background. Source paper: CC BY 4.0. Added annotations: CC BY 4.0.

Open annotated PDF