Understanding Deep Learning with Statistical Relevance

Bibliographic Data

ID10705246
AuthorsTim Räz (0000-0002-8464-4190, University of Bern, corresponding author)
Year2022
Volume89
Issue1
Pages20-41
Publication date2022-01-01
Open AccessYes
TypeARTICLE
VenuePhilosophy of Science (JOURNAL)
Journal identifiersISSN: 0031-8248 • E-ISSN: 1539-767X
PublisherCambridge University Press (CUP) (PUBLISHER)
DOI10.1017/psa.2021.12
OpenAlexW3107238979
LanguageEN
Citations received7
References cited12

This paper argues that a notion of statistical explanation, based on Salmon’s statistical relevance model, can help us better understand deep neural networks. It is proved that homogeneous partitions, the core notion of Salmon’s model, are equivalent to minimal sufficient statistics, an important notion from statistical inference. This establishes a link to deep neural networks via the so-called Information Bottleneck method, an information-theoretic framework, according to which deep neural networks implicitly solve an optimization problem that generalizes minimal sufficient statistics. The resulting notion of statistical explanation is general, mathematical, and subcausal

Artificial neural network · Bottleneck · Core (optical fiber) · Deep learning · Deep neural networks · Inference · Machine learning · Relevance (law) · Statistical hypothesis testing · Statistical inference · Statistical model · Statistics · Artificial Intelligence · Bayesian Modeling and Causal Inference · Computer Science · Explainable Artificial Intelligence (XAI · Mathematics · Neural Networks and Applications · Theoretical Computer Science

Tools

Open DOIOpen Access