Räz, Tim (2020) Understanding Deep Learning With Statistical Relevance. [Preprint]

Preview

Text SR_philsci_archive.pdf Download (618kB) | Preview

Abstract

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 sub-causal.

---

Export/Citation:

EndNote | BibTeX | Dublin Core | ASCII/Text Citation (Chicago) | HTML Citation | OpenURL

Social Networking:

Share |

---

Item Type:	Preprint
Creators:	Creators Email ORCID Räz, Tim tim.raez@gmail.com
Additional Information:	Forthcoming in Philosophy of Science
Keywords:	statistics, explanation, deep neural networks, information bottleneck
Subjects:	General Issues > Data Specific Sciences > Mathematics > Explanation Specific Sciences > Computer Science General Issues > Explanation Specific Sciences > Artificial Intelligence > Machine Learning Specific Sciences > Probability/Statistics
Depositing User:	Tim Räz
Date Deposited:	04 Nov 2020 17:22
Last Modified:	04 Nov 2020 17:22
Item ID:	18356
Subjects:	General Issues > Data Specific Sciences > Mathematics > Explanation Specific Sciences > Computer Science General Issues > Explanation Specific Sciences > Artificial Intelligence > Machine Learning Specific Sciences > Probability/Statistics
Date:	2020
URI:	https://philsci-archive.pitt.edu/id/eprint/18356

Monthly Views for the past 3 years

Monthly Downloads for the past 3 years

Plum Analytics

Actions (login required)

View Item