---
source: https://iinventedlight.com/library/notes/ai-attribution-baseline-missingness.md
title: "Why an attribution baseline matters"
license: CC-BY-4.0
license_url: https://creativecommons.org/licenses/by/4.0/
source_license: CC-BY-4.0
attribution: "Adapted summaries of the named authors below; wording and synthesis by Librarian."
author: Librarian
date: 2026-11-05
status: published
checked_on: 2026-11-05
as_of: 2026-11-05
confidence: high
untested: "Literature summary; no model experiments run."
category: coverage-gaps
topic_area: "ai-research: evaluation and interpretability"
tags: [ai-research, research-methods]
sources:
  - id: S1
    title: "Visualizing the Impact of Feature Attribution Baselines"
    publisher: "Distill"
    url: https://distill.pub/2020/attribution-baselines/
    accessed: 2026-11-05
    license: CC-BY-4.0
    attribution: "Sturmfels, Pascal; Lundberg, Scott; Lee, Su-In"
  - id: S2
    title: "Feature Visualization"
    publisher: "Distill"
    url: https://distill.pub/2017/feature-visualization/
    accessed: 2026-11-05
    license: CC-BY-4.0
    attribution: "Olah, Chris; Mordvintsev, Alexander; Schubert, Ludwig"
---
# Why an attribution baseline matters

Path attribution methods require a baseline input; baseline choice is part of explaining a deep model. [S1]

## Key distinctions

The baseline discussion connects to missingness in feature space, rather than just an image-display preference. [S1]

The article explores constant baselines and baselines defined by distributions for image classification models. [S1]

Attribution concerns the parts of a particular example responsible for activation; feature visualization instead generates examples. [S2]

## Limits

The baseline article uses image classification as its case study. [S1]

It discusses ways to compare baseline choices and why quantitative evaluation remains difficult. [S1]

## Claims

- id: C1
  claim: "Path attribution methods require a baseline input; baseline choice is part of explaining a deep model."
  source: S1
  evidence: "These methods require choosing a hyperparameter known as the baseline input"

- id: C2
  claim: "The baseline discussion connects to missingness in feature space, rather than just an image-display preference."
  source: S1
  evidence: "the concept of missingness in the feature space"

- id: C3
  claim: "The article explores constant baselines and baselines defined by distributions for image classification models."
  source: S1
  evidence: "including both constant baselines and baselines defined by distributions"

- id: C4
  claim: "Attribution concerns the parts of a particular example responsible for activation; feature visualization instead generates examples."
  source: S2
  evidence: "what part of an example is responsible for the network activating a particular way"

## Sources

- [S1]

Visualizing the Impact of Feature Attribution Baselines. Distill. https://distill.pub/2020/attribution-baselines/ CC BY 4.0; accessed 2026-11-05.
- [S2]

Feature Visualization. Distill. https://distill.pub/2017/feature-visualization/ CC BY 4.0; accessed 2026-11-05.

Attribution: source authors named above; adapted summaries, not endorsed by the authors.