Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Zarlenga, Mateo Espinosa, Dominici, Gabriele, Barbiero, Pietro, Shams, Zohreh, Jamnik, Mateja
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909720130879488
author Zarlenga, Mateo Espinosa
Dominici, Gabriele
Barbiero, Pietro
Shams, Zohreh
Jamnik, Mateja
author_facet Zarlenga, Mateo Espinosa
Dominici, Gabriele
Barbiero, Pietro
Shams, Zohreh
Jamnik, Mateja
contents In this paper, we investigate how concept-based models (CMs) respond to out-of-distribution (OOD) inputs. CMs are interpretable neural architectures that first predict a set of high-level concepts (e.g., stripes, black) and then predict a task label from those concepts. In particular, we study the impact of concept interventions (i.e., operations where a human expert corrects a CM's mispredicted concepts at test time) on CMs' task predictions when inputs are OOD. Our analysis reveals a weakness in current state-of-the-art CMs, which we term leakage poisoning, that prevents them from properly improving their accuracy when intervened on for OOD inputs. To address this, we introduce MixCEM, a new CM that learns to dynamically exploit leaked information missing from its concepts only when this information is in-distribution. Our results across tasks with and without complete sets of concept annotations demonstrate that MixCEMs outperform strong baselines by significantly improving their accuracy for both in-distribution and OOD samples in the presence and absence of concept interventions.
format Preprint
id arxiv_https___arxiv_org_abs_2504_17921
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts
Zarlenga, Mateo Espinosa
Dominici, Gabriele
Barbiero, Pietro
Shams, Zohreh
Jamnik, Mateja
Machine Learning
Artificial Intelligence
Cryptography and Security
Human-Computer Interaction
In this paper, we investigate how concept-based models (CMs) respond to out-of-distribution (OOD) inputs. CMs are interpretable neural architectures that first predict a set of high-level concepts (e.g., stripes, black) and then predict a task label from those concepts. In particular, we study the impact of concept interventions (i.e., operations where a human expert corrects a CM's mispredicted concepts at test time) on CMs' task predictions when inputs are OOD. Our analysis reveals a weakness in current state-of-the-art CMs, which we term leakage poisoning, that prevents them from properly improving their accuracy when intervened on for OOD inputs. To address this, we introduce MixCEM, a new CM that learns to dynamically exploit leaked information missing from its concepts only when this information is in-distribution. Our results across tasks with and without complete sets of concept annotations demonstrate that MixCEMs outperform strong baselines by significantly improving their accuracy for both in-distribution and OOD samples in the presence and absence of concept interventions.
title Avoiding Leakage Poisoning: Concept Interventions Under Distribution Shifts
topic Machine Learning
Artificial Intelligence
Cryptography and Security
Human-Computer Interaction
url https://arxiv.org/abs/2504.17921