Saved in:
Bibliographic Details
Main Authors: Zaher, Eslam, Trzaskowski, Maciej, Nguyen, Quan, Roosta, Fred
Format: Preprint
Published: 2026
Subjects:
Online Access:https://arxiv.org/abs/2601.18678
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912934019465216
author Zaher, Eslam
Trzaskowski, Maciej
Nguyen, Quan
Roosta, Fred
author_facet Zaher, Eslam
Trzaskowski, Maciej
Nguyen, Quan
Roosta, Fred
contents Latent-space optimization methods for counterfactual explanations - framed as minimal semantic perturbations that change model predictions - inherit the ambiguity of Wachter et al.'s objective: the choice of distance metric dictates whether perturbations are meaningful or adversarial. Existing approaches adopt flat or misaligned geometries, leading to off-manifold artifacts, semantic drift, or adversarial collapse. We introduce Perceptual Counterfactual Geodesics (PCG), a method that constructs counterfactuals by tracing geodesics under a perceptually Riemannian metric induced from robust vision features. This geometry aligns with human perception and penalizes brittle directions, enabling smooth, on-manifold, semantically valid transitions. Experiments on three vision datasets show that PCG outperforms baselines and reveals failure modes hidden under standard metrics.
format Preprint
id arxiv_https___arxiv_org_abs_2601_18678
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle Counterfactual Explanations on Robust Perceptual Geodesics
Zaher, Eslam
Trzaskowski, Maciej
Nguyen, Quan
Roosta, Fred
Machine Learning
Computer Vision and Pattern Recognition
Human-Computer Interaction
Differential Geometry
Latent-space optimization methods for counterfactual explanations - framed as minimal semantic perturbations that change model predictions - inherit the ambiguity of Wachter et al.'s objective: the choice of distance metric dictates whether perturbations are meaningful or adversarial. Existing approaches adopt flat or misaligned geometries, leading to off-manifold artifacts, semantic drift, or adversarial collapse. We introduce Perceptual Counterfactual Geodesics (PCG), a method that constructs counterfactuals by tracing geodesics under a perceptually Riemannian metric induced from robust vision features. This geometry aligns with human perception and penalizes brittle directions, enabling smooth, on-manifold, semantically valid transitions. Experiments on three vision datasets show that PCG outperforms baselines and reveals failure modes hidden under standard metrics.
title Counterfactual Explanations on Robust Perceptual Geodesics
topic Machine Learning
Computer Vision and Pattern Recognition
Human-Computer Interaction
Differential Geometry
url https://arxiv.org/abs/2601.18678