Building firmly nonexpansive convolutional neural networks

Matthieu Terris, Audrey Repetti, Jean-Christophe Pesquet, Yves Wiaux

Research output: Contribution to conferencePaper

Abstract

Building nonexpansive Convolutional Neural Networks (CNNs) is a challenging problem that has recently gained a lot of attention from the image processing community. In particular, it appears to be the key to obtain convergent Plug-and-Play algorithms. This problem consists in controlling the Lipschitz constant of the convolutional layers. Interestingly, this problem has also been investigated for Generative Adversarial Networks to improve robustness to adversarial perturbations. However, to the best of our knowledge, no efficient method has been developed yet to build nonexpansive CNNs. In this paper, we develop an optimization algorithm that can be incorporated in the training of a network to ensure the nonexpansiveness of its convolutional layers. This is shown to allow us to build firmly nonexpansive CNNs. We apply the proposed approach to train a CNN for an image denoising task and show its effectiveness through simulations.
Original languageEnglish
Number of pages5
Publication statusSubmitted - Oct 2019
EventICASSP 2020 - Barcelona, Spain
Duration: 4 May 20208 May 2020
https://2020.ieeeicassp.org

Conference

ConferenceICASSP 2020
CountrySpain
CityBarcelona
Period4/05/208/05/20
Internet address

Fingerprint

Neural networks
Image denoising
Image processing

Cite this

Terris, M., Repetti, A., Pesquet, J-C., & Wiaux, Y. (2019). Building firmly nonexpansive convolutional neural networks. Paper presented at ICASSP 2020, Barcelona, Spain.
Terris, Matthieu ; Repetti, Audrey ; Pesquet, Jean-Christophe ; Wiaux, Yves. / Building firmly nonexpansive convolutional neural networks. Paper presented at ICASSP 2020, Barcelona, Spain.5 p.
@conference{e63c63c34aa64fc2b32ee7eba26dbdd3,
title = "Building firmly nonexpansive convolutional neural networks",
abstract = "Building nonexpansive Convolutional Neural Networks (CNNs) is a challenging problem that has recently gained a lot of attention from the image processing community. In particular, it appears to be the key to obtain convergent Plug-and-Play algorithms. This problem consists in controlling the Lipschitz constant of the convolutional layers. Interestingly, this problem has also been investigated for Generative Adversarial Networks to improve robustness to adversarial perturbations. However, to the best of our knowledge, no efficient method has been developed yet to build nonexpansive CNNs. In this paper, we develop an optimization algorithm that can be incorporated in the training of a network to ensure the nonexpansiveness of its convolutional layers. This is shown to allow us to build firmly nonexpansive CNNs. We apply the proposed approach to train a CNN for an image denoising task and show its effectiveness through simulations.",
author = "Matthieu Terris and Audrey Repetti and Jean-Christophe Pesquet and Yves Wiaux",
year = "2019",
month = "10",
language = "English",
note = "ICASSP 2020 ; Conference date: 04-05-2020 Through 08-05-2020",
url = "https://2020.ieeeicassp.org",

}

Terris, M, Repetti, A, Pesquet, J-C & Wiaux, Y 2019, 'Building firmly nonexpansive convolutional neural networks' Paper presented at ICASSP 2020, Barcelona, Spain, 4/05/20 - 8/05/20, .

Building firmly nonexpansive convolutional neural networks. / Terris, Matthieu; Repetti, Audrey; Pesquet, Jean-Christophe; Wiaux, Yves.

2019. Paper presented at ICASSP 2020, Barcelona, Spain.

Research output: Contribution to conferencePaper

TY - CONF

T1 - Building firmly nonexpansive convolutional neural networks

AU - Terris, Matthieu

AU - Repetti, Audrey

AU - Pesquet, Jean-Christophe

AU - Wiaux, Yves

PY - 2019/10

Y1 - 2019/10

N2 - Building nonexpansive Convolutional Neural Networks (CNNs) is a challenging problem that has recently gained a lot of attention from the image processing community. In particular, it appears to be the key to obtain convergent Plug-and-Play algorithms. This problem consists in controlling the Lipschitz constant of the convolutional layers. Interestingly, this problem has also been investigated for Generative Adversarial Networks to improve robustness to adversarial perturbations. However, to the best of our knowledge, no efficient method has been developed yet to build nonexpansive CNNs. In this paper, we develop an optimization algorithm that can be incorporated in the training of a network to ensure the nonexpansiveness of its convolutional layers. This is shown to allow us to build firmly nonexpansive CNNs. We apply the proposed approach to train a CNN for an image denoising task and show its effectiveness through simulations.

AB - Building nonexpansive Convolutional Neural Networks (CNNs) is a challenging problem that has recently gained a lot of attention from the image processing community. In particular, it appears to be the key to obtain convergent Plug-and-Play algorithms. This problem consists in controlling the Lipschitz constant of the convolutional layers. Interestingly, this problem has also been investigated for Generative Adversarial Networks to improve robustness to adversarial perturbations. However, to the best of our knowledge, no efficient method has been developed yet to build nonexpansive CNNs. In this paper, we develop an optimization algorithm that can be incorporated in the training of a network to ensure the nonexpansiveness of its convolutional layers. This is shown to allow us to build firmly nonexpansive CNNs. We apply the proposed approach to train a CNN for an image denoising task and show its effectiveness through simulations.

M3 - Paper

ER -

Terris M, Repetti A, Pesquet J-C, Wiaux Y. Building firmly nonexpansive convolutional neural networks. 2019. Paper presented at ICASSP 2020, Barcelona, Spain.