Mitigating Shortcut Learning with Diffusion Counterfactuals and Diverse Ensembles

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Scimeca, Luca, Rubinstein, Alexander, Teney, Damien, Oh, Seong Joon, Bengio, Yoshua
Formato: Preprint
Publicado: 2023
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866912306076581888
author Scimeca, Luca
Rubinstein, Alexander
Teney, Damien
Oh, Seong Joon
Bengio, Yoshua
author_facet Scimeca, Luca
Rubinstein, Alexander
Teney, Damien
Oh, Seong Joon
Bengio, Yoshua
contents Spurious correlations in the data, where multiple cues are predictive of the target labels, often lead to a phenomenon known as shortcut learning, where a model relies on erroneous, easy-to-learn cues while ignoring reliable ones. In this work, we propose DiffDiv an ensemble diversification framework exploiting Diffusion Probabilistic Models (DPMs) to mitigate this form of bias. We show that at particular training intervals, DPMs can generate images with novel feature combinations, even when trained on samples displaying correlated input features. We leverage this crucial property to generate synthetic counterfactuals to increase model diversity via ensemble disagreement. We show that DPM-guided diversification is sufficient to remove dependence on shortcut cues, without a need for additional supervised signals. We further empirically quantify its efficacy on several diversification objectives, and finally show improved generalization and diversification on par with prior work that relies on auxiliary data collection.
format Preprint
id arxiv_https___arxiv_org_abs_2311_16176
institution arXiv
publishDate 2023
record_format arxiv
spellingShingle Mitigating Shortcut Learning with Diffusion Counterfactuals and Diverse Ensembles
Scimeca, Luca
Rubinstein, Alexander
Teney, Damien
Oh, Seong Joon
Bengio, Yoshua
Machine Learning
Artificial Intelligence
Computer Vision and Pattern Recognition
Spurious correlations in the data, where multiple cues are predictive of the target labels, often lead to a phenomenon known as shortcut learning, where a model relies on erroneous, easy-to-learn cues while ignoring reliable ones. In this work, we propose DiffDiv an ensemble diversification framework exploiting Diffusion Probabilistic Models (DPMs) to mitigate this form of bias. We show that at particular training intervals, DPMs can generate images with novel feature combinations, even when trained on samples displaying correlated input features. We leverage this crucial property to generate synthetic counterfactuals to increase model diversity via ensemble disagreement. We show that DPM-guided diversification is sufficient to remove dependence on shortcut cues, without a need for additional supervised signals. We further empirically quantify its efficacy on several diversification objectives, and finally show improved generalization and diversification on par with prior work that relies on auxiliary data collection.
title Mitigating Shortcut Learning with Diffusion Counterfactuals and Diverse Ensembles
topic Machine Learning
Artificial Intelligence
Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2311.16176