Universal Rates of Empirical Risk Minimization

Steve Hanneke, Mingyue Xu · 2024

Read Universal Rates of Empirical Risk Minimization by Steve Hanneke, Mingyue Xu (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-02810.txt.

Description

Steve Hanneke and Mingyue Xu study in 2024 how fast a classic learning rule improves as it sees more data, under a broad model of learning. The rule, empirical risk minimization, picks the option with the lowest error on the training sample. It underlies many common algorithms and the classic theory of learnability.

Classic theory fails to explain the observed shape of learning curves, the fall of error with sample size. A newer universal model of learning separates two kinds of best possible rates. The gap between them for this classic rule was not understood. Hanneke and Xu study the case where a perfect option exists. Their results clarify when the rule learns at the best universal rate. That sharpens what teachers of the theory can promise about data needs.

Catalog record

CREATORSteve Hanneke, Mingyue Xu
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE156 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-02810

Browse the full library · About the Tower of Babel