Model-Agnostic Visual Explanations via Approximate Bilinear Models

Onderzoeksoutput: Conference paper

130 Downloads (Pure)


This paper proposes InteractionLIME: a model-agnostic attribution technique to explain deep models predictions in terms of feature interactions. Specifically, we regress a bilinear form to approximate the output of two-input models, by sampling perturbations of both inputs simultaneously. Upon training, we retrieve a global explanation and a set of feature partitioning maps via the singular value decomposition of the learned interaction matrix of the bilinear model. We demonstrate InteractionLIME on vision and text-vision contrastive models, using visual examples and quantitative evaluation metrics. Our results show that the bilinear model successfully retrieves important interacting features from both inputs, while strongly reducing the occurrence of incomplete or asymmetric explanations produced by a linear model.
Originele taal-2English
Titel2023 IEEE International Conference on Image Processing
Aantal pagina's5
ISBN van elektronische versie978-1-7281-9835-4
ISBN van geprinte versie978-1-7281-9836-1
StatusPublished - 11 jul 2023
Evenement2023 IEEE International Conference on Image Processing - Kuala Lumpur Convention Center (KLCC), Kuala Lumpur, Malaysia
Duur: 8 okt 202311 okt 2023

Publicatie series

NaamProceedings - International Conference on Image Processing, ICIP
ISSN van geprinte versie1522-4880


Conference2023 IEEE International Conference on Image Processing
Verkorte titelICIP2023
StadKuala Lumpur
Internet adres

Bibliografische nota

Publisher Copyright:
© 2023 IEEE.


Duik in de onderzoeksthema's van 'Model-Agnostic Visual Explanations via Approximate Bilinear Models'. Samen vormen ze een unieke vingerafdruk.

Citeer dit