---
title: "ICON Decomposition: Multivariate Concept-Level Explanations of Deep Representations for Model Auditing"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2608.26083v1"
permalink: "https://feed.astralor.com/items/itm_2dfc4cd11607bc388d85924a"
published: "2026-08-26T17:47:49.000Z"
score: 75
channel: "ai"
tags: ["ICON Decomposition","Medical Imaging","Shortcut Learning","Concept-based Explainability","Model Auditing"]
---

# ICON Decomposition: Multivariate Concept-Level Explanations of Deep Representations for Model Auditing

## AI commentary

> This paper hits a common pitfall in explainability, upgrading concepts from 'decodability' to 'unique contribution' with clear logic. For model auditors and medical AI practitioners, ICON could become a new staple in the toolbox.

## At a glance

ICON Decomposition is a novel concept-level interpretability method for auditing deep models against shortcut learning. Unlike traditional concept-based methods that evaluate each concept in isolation, ICON quantifies the unique variance contribution of each concept to a network layer after controlling for all other concepts and the outcome variable. This prevents mistaking correlations between concepts as evidence of model dependence. On synthetic data with known ground truth, ICON recovers concept importance more accurately than seven baseline methods. On skin-lesion and brain-imaging models, ICON isolates the concepts on which the model genuinely relies, quantifies the unexplained representation, and yields sparse explanations validated by retraining and out-of-distribution testing. The method provides a more reliable way to audit deep representations and uncover shortcut behavior in high-stakes applications.

## Key points

- Proposes ICON Decomposition to quantify the variance explained by each concept in a multivariate setting, avoiding biases from isolated evaluation.
- On synthetic data, ICON recovers concept importance more accurately than seven baseline methods.
- On skin-lesion and brain-imaging models, ICON isolates genuinely relied-upon sparse concepts, validated by retraining and out-of-distribution testing.
- Quantifies the representation unexplained by supplied concepts, offering a more complete view for model auditing.

## Full text

Deep neural networks often exploit spurious associations in their training data, a failure known as shortcut learning. Concept-based explainability methods screen for shortcuts by testing whether concepts such as a patient's sex or scanner settings can be decoded from a network layer. Because each concept is evaluated in isolation, these methods can mistake correlations between concepts as evidence that the model uses them. We introduce ICON decomposition, which instead quantifies how much of a layer's variance each concept explains after accounting for all other concepts and the outcome. On synthetic data with known ground truth, ICON recovers concept importance more accurately than seven alternative baseline methods. On skin-lesion and brain-imaging models, it isolates the concepts on which a model genuinely relies, quantifies the representation unexplained by any of the supplied concepts, and yields sparse explanations that we validate by retraining and out-of-distribution testing.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_2dfc4cd11607bc388d85924a