Invertible generative models for inverse problems: mitigating representation error and dataset bias

Muhammad Asim, Max Daniels, Oscar Leong, Ali Ahmed, Paul Hand
Proceedings of the 37th International Conference on Machine Learning, PMLR 119:399-409, 2020.

Abstract

Trained generative models have shown remarkable performance as priors for inverse problems in imaging – for example, Generative Adversarial Network priors permit recovery of test images from 5-10x fewer measurements than sparsity priors. Unfortunately, these models may be unable to represent any particular image because of architectural choices, mode collapse, and bias in the training dataset. In this paper, we demonstrate that invertible neural networks, which have zero representation error by design, can be effective natural signal priors at inverse problems such as denoising, compressive sensing, and inpainting. Given a trained generative model, we study the empirical risk formulation of the desired inverse problem under a regularization that promotes high likelihood images, either directly by penalization or algorithmically by initialization. For compressive sensing, invertible priors can yield higher accuracy than sparsity priors across almost all undersampling ratios, and due to their lack of representation error, invertible priors can yield better reconstructions than GAN priors for images that have rare features of variation within the biased training set, including out-of-distribution natural images. We additionally compare performance for compressive sensing to unlearned methods, such as the deep decoder, and we establish theoretical bounds on expected recovery error in the case of a linear invertible model.

Cite this Paper


BibTeX
@InProceedings{pmlr-v119-asim20a, title = {Invertible generative models for inverse problems: mitigating representation error and dataset bias}, author = {Asim, Muhammad and Daniels, Max and Leong, Oscar and Ahmed, Ali and Hand, Paul}, booktitle = {Proceedings of the 37th International Conference on Machine Learning}, pages = {399--409}, year = {2020}, editor = {III, Hal Daumé and Singh, Aarti}, volume = {119}, series = {Proceedings of Machine Learning Research}, month = {13--18 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v119/asim20a/asim20a.pdf}, url = {http://proceedings.mlr.press/v119/asim20a.html}, abstract = {Trained generative models have shown remarkable performance as priors for inverse problems in imaging – for example, Generative Adversarial Network priors permit recovery of test images from 5-10x fewer measurements than sparsity priors. Unfortunately, these models may be unable to represent any particular image because of architectural choices, mode collapse, and bias in the training dataset. In this paper, we demonstrate that invertible neural networks, which have zero representation error by design, can be effective natural signal priors at inverse problems such as denoising, compressive sensing, and inpainting. Given a trained generative model, we study the empirical risk formulation of the desired inverse problem under a regularization that promotes high likelihood images, either directly by penalization or algorithmically by initialization. For compressive sensing, invertible priors can yield higher accuracy than sparsity priors across almost all undersampling ratios, and due to their lack of representation error, invertible priors can yield better reconstructions than GAN priors for images that have rare features of variation within the biased training set, including out-of-distribution natural images. We additionally compare performance for compressive sensing to unlearned methods, such as the deep decoder, and we establish theoretical bounds on expected recovery error in the case of a linear invertible model.} }
Endnote
%0 Conference Paper %T Invertible generative models for inverse problems: mitigating representation error and dataset bias %A Muhammad Asim %A Max Daniels %A Oscar Leong %A Ali Ahmed %A Paul Hand %B Proceedings of the 37th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2020 %E Hal Daumé III %E Aarti Singh %F pmlr-v119-asim20a %I PMLR %P 399--409 %U http://proceedings.mlr.press/v119/asim20a.html %V 119 %X Trained generative models have shown remarkable performance as priors for inverse problems in imaging – for example, Generative Adversarial Network priors permit recovery of test images from 5-10x fewer measurements than sparsity priors. Unfortunately, these models may be unable to represent any particular image because of architectural choices, mode collapse, and bias in the training dataset. In this paper, we demonstrate that invertible neural networks, which have zero representation error by design, can be effective natural signal priors at inverse problems such as denoising, compressive sensing, and inpainting. Given a trained generative model, we study the empirical risk formulation of the desired inverse problem under a regularization that promotes high likelihood images, either directly by penalization or algorithmically by initialization. For compressive sensing, invertible priors can yield higher accuracy than sparsity priors across almost all undersampling ratios, and due to their lack of representation error, invertible priors can yield better reconstructions than GAN priors for images that have rare features of variation within the biased training set, including out-of-distribution natural images. We additionally compare performance for compressive sensing to unlearned methods, such as the deep decoder, and we establish theoretical bounds on expected recovery error in the case of a linear invertible model.
APA
Asim, M., Daniels, M., Leong, O., Ahmed, A. & Hand, P.. (2020). Invertible generative models for inverse problems: mitigating representation error and dataset bias. Proceedings of the 37th International Conference on Machine Learning, in Proceedings of Machine Learning Research 119:399-409 Available from http://proceedings.mlr.press/v119/asim20a.html.

Related Material