Yapay Zeka · 8 dk okuma · 17 Nisan 2026
Formalizing How Much Data Proves a Learning Model Right
Researchers formalize identifying information—the bits needed to confirm or reject a hypothesis—bridging information theory with practical sample complexity.
Kaynak: arxiv/cs.LG · Derek S. Prijatelj (University of Notre Dame), Timothy J. Ireland (Independent Researcher), Walter J. Scheirer (University of Notre Dame) · orijinali aç ↗ ↗
A formal framework quantifies how many observations are needed to verify or falsify a hypothesis in machine learning.
- — Identifying information measures bits that confirm or reject a hypothesis as the true data-generating process.
- — Sample complexity—how many observations are required—connects to information-theoretic properties of hypothesis identification.
- — Framework spans deterministic processes through ergodic stationary stochastic processes, unifying finite-sample and asymptotic analysis.
- — Indicator functions over hypothesis sets formalize novelty detection and misspecified model identification.
- — PAC-Bayes learner sample complexity distribution is computable from prior probability moments over finite hypothesis sets.
- — Bridges algorithmic information theory with probabilistic frameworks, answering when a learner has sufficient evidence.
Sık sorulanlar
- Identifying information refers to the bits of data that either confirm or reject a hypothesis about the true data-generating process. It quantifies how much evidence is needed to distinguish the correct model from incorrect alternatives. The framework formalizes this using information theory, connecting it to sample complexity—the number of observations required to make that determination with confidence.