Portrait of Dr. Avinash Kori

Dr. Avinash Kori

Chief Scientist @ blurgs.ai
Co-founder & Chief Scientist, Blurgs AI

Visual debates: explaining image classifiers through dialogue

3 minute read

In brief

  • Current explanation methods rely heavily on static pixel heatmaps that do not reflect human explanatory reasoning.
  • Visual debates model explanations as an interactive adversarial dialogue between two agents arguing for competing classes.
  • Arguments correspond to extracted visual concepts and evidence that attack and defend rival hypotheses.

This retrospective explainer is dated to the October 2024 Discovery Science conference. It discusses Explaining image classifiers with visual debates by Avinash Kori, Ben Glocker and Francesca Toni.

When a doctor consults a radiologist, or a juror listens to an expert witness, explanation rarely takes the form of a static highlight marker drawn across a photograph. Human reasoning is communicative and dialectical: one party presents evidence for a diagnosis, a second party raises a counter-argument or points out an alternative explanation, and the debate proceeds until a consensus emerges.

Standard explainable AI (XAI) for computer vision, by contrast, overwhelmingly relies on static attribution maps (such as Grad-CAM or saliency maps). These maps highlight which pixels or image regions correlated with a classifier’s score, but they cannot answer basic conversational questions: Why this class instead of that one? Which feature rebuts the alternative?

This paper introduces visual debates, an interactive framework that reframes image explanation as an adversarial debate game between two agents arguing for and against the classifier’s outputs.

Why pixel heatmaps are not explanations

A saliency map identifies regions where small perturbations change the output logit. However:

  1. Lack of contrastive semantics: Highlighting a dog’s snout does not explain why the model rejected the label “wolf” or “fox.”
  2. No logical structure: A heatmap cannot express conditional rules, such as: “Feature $A$ supports the diagnosis only if feature $B$ is absent.”
  3. Passive consumption: The human user is given an unverified visual hint rather than an interactive argument that can be challenged.

In a visual debate, two computational agents ($P_1$ and $P_2$) represent rival hypotheses. Player $P_1$ argues that the image depicts class $c_1$, while Player $P_2$ defends class $c_2$. Each player selects arguments grounded in the latent feature representations learned by the vision model.

Dialectical argumentation over visual evidence

The debate operates over an argumentation graph $\mathcal{G} = (\mathcal{A}, \mathcal{R})$, where $\mathcal{A}$ is a finite set of arguments derived from image features, and $\mathcal{R} \subseteq \mathcal{A} \times \mathcal{A}$ is an attack relation.

An argument $a \in \mathcal{A}$ is a tuple:

\[a = \langle z_i, c, w \rangle,\]

where $z_i$ represents a latent visual concept extracted from the classifier’s intermediate layers, $c$ is the target class being supported, and $w \in \mathbb{R}$ indicates the strength or predictive relevance of the feature.

When Player 1 introduces argument $a_1$ supporting class $c_1$, Player 2 can respond by offering counter-argument $a_2$ such that $(a_2, a_1) \in \mathcal{R}$. The attack succeeds if the strength of $a_2$ undermines the evidentiary validity of $a_1$ under the current image context:

\[\mathrm{Att}(a_2, a_1) \iff \mathrm{sim}(z_1, z_2) > \tau \quad \text{and} \quad c_1 \neq c_2,\]

where $\mathrm{sim}(\cdot, \cdot)$ measures semantic feature compatibility and $\tau$ is an acceptance threshold.

Resolving debates through dialectical semantics

The debate proceeds in turns until no further undefeated arguments can be introduced. The acceptability of each claim is then evaluated using Dung’s abstract argumentation semantics:

  1. A set of arguments $S \subseteq \mathcal{A}$ is conflict-free if no argument in $S$ attacks another argument in $S$.
  2. An argument $a \in \mathcal{A}$ is defended by $S$ if for every $b \in \mathcal{A}$ that attacks $a$, there exists some $d \in S$ that attacks $b$.
  3. The preferred extension represents a maximal self-consistent set of defended arguments justifying the winning classification.

By following this dialectical trajectory, the user sees not just a glowing blob on an image, but the exact chain of claims, attacks, and counter-claims that drove the classifier to its conclusion.

Paper and sources