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.
Joukovsky, B, Sammani, F & Deligiannis, N 2023, Model-Agnostic Visual Explanations via Approximate Bilinear Models. in 2023 IEEE International Conference on Image Processing. Proceedings - International Conference on Image Processing, ICIP, IEEE, pp. 1770-1774, 2023 IEEE International Conference on Image Processing, Kuala Lumpur, Malaysia, 8/10/23. https://doi.org/10.1109/ICIP49359.2023.10222440
Joukovsky, B., Sammani, F., & Deligiannis, N. (2023). Model-Agnostic Visual Explanations via Approximate Bilinear Models. In 2023 IEEE International Conference on Image Processing (pp. 1770-1774). (Proceedings - International Conference on Image Processing, ICIP). IEEE. https://doi.org/10.1109/ICIP49359.2023.10222440
@inproceedings{47ef400d48d142fc826ee0859509c66d,
title = "Model-Agnostic Visual Explanations via Approximate Bilinear Models",
abstract = "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.",
author = "Boris Joukovsky and Fawaz Sammani and Nikos Deligiannis",
note = "Funding Information: This research received funding from the Flemish Government under the âOnderzoeksprogramma Artifici{\"e}le Intelligentie (AI) Vlaanderenâ programme, and from the FWO (Grants 1SB5721N and G0A4720N), Belgium. Publisher Copyright: {\textcopyright} 2023 IEEE.; 2023 IEEE International Conference on Image Processing, ICIP2023 ; Conference date: 08-10-2023 Through 11-10-2023",
year = "2023",
month = jul,
day = "11",
doi = "10.1109/ICIP49359.2023.10222440",
language = "English",
isbn = "978-1-7281-9836-1",
series = "Proceedings - International Conference on Image Processing, ICIP",
publisher = "IEEE",
pages = "1770--1774",
booktitle = "2023 IEEE International Conference on Image Processing",
url = "https://2023.ieeeicip.org/",
}