PhilSci Archive

From Explanations to Interpretability and Back

Räz, Tim (2024) From Explanations to Interpretability and Back. [Preprint]

[img]
Preview
Text
raez_ml_expl_philsci.pdf

Download (329kB) | Preview

Abstract

This chapter discusses connections between interpretability of machine learning and (scientific and mathematical) explanations, provides novel perspectives on interpretability, and highlights under-explored issues. In- terpretability types are proposed: kinds of interpretability should be distinguished using both the parts of ML we want to explain and the parts of ML we use to explain. It is argued that not all explanations are contrastive, and that we should also consider contrasts with respect to models and data, not only with respect to inputs. Theoretical explanations are highlighted; they include issues like generalization, optimization, and ex- pressivity. It is proposed that there are two threats to the objectivity of explanations: One from radical subject-dependence, the other from a lack of factivity. Finally, pluralism is advocated: There are different notions of interpretability and different notions of (scientific and mathematical) explanations. However, the heterogeneity of one area does not transfer to the other in a straightforward manner.


Export/Citation: EndNote | BibTeX | Dublin Core | ASCII/Text Citation (Chicago) | HTML Citation | OpenURL
Social Networking:
Share |

Item Type: Preprint
Creators:
CreatorsEmailORCID
Räz, Timtim.raez@gmail.com
Keywords: interpretability, scientific explanation, mathematical explanation, understanding, XAI, pluralism, machine learning
Subjects: Specific Sciences > Computer Science
General Issues > Explanation
Specific Sciences > Artificial Intelligence > Machine Learning
General Issues > Models and Idealization
Depositing User: Tim Räz
Date Deposited: 05 Mar 2024 00:56
Last Modified: 05 Mar 2024 00:56
Item ID: 23169
Subjects: Specific Sciences > Computer Science
General Issues > Explanation
Specific Sciences > Artificial Intelligence > Machine Learning
General Issues > Models and Idealization
Date: 2024
URI: https://philsci-archive.pitt.edu/id/eprint/23169

Monthly Views for the past 3 years

Monthly Downloads for the past 3 years

Plum Analytics

Actions (login required)

View Item View Item