PhD offer (M/W): self-supervised learning for non-linear inverse problems with applications to phase retrieval

22 Sep 2023
Job Information

Organisation/Company

CNRS

Department

Laboratoire de Physique

Research Field

Engineering
Computer science
Mathematics

Researcher Profile

First Stage Researcher (R1)

Country

France

Application Deadline

16 Oct 2023 – 23:59 (UTC)

Type of Contract

Temporary

Job Status

Full-time

Hours Per Week

35

Offer Starting Date

1 Nov 2023

Is the job funded through the EU Research Framework Programme?

Not funded by an EU programme

Is the Job related to staff position within a Research Infrastructure?

No

Offer Description

The project will be fully funded by the ANR “Unsupervised Learning for Non-Linear Inverse Problems (UNLIP)” project. The student will be part of the physics laboratory (https://www.ens-lyon.fr/PHYSIQUE ) at ENS Lyon, France. The student will benefit from a stimulating environment of experts in machine learning (https://team.inria.fr/ockham/ ), signal processing (https://www.ens-lyon.fr/PHYSIQUE/teams/signaux-systemes-physique ) and physics. The laboratory also organizes weekly seminars given by international experts (https://www.ens-lyon.fr/PHYSIQUE/seminars/machine-learning-and-signal-p… ).

Reconstruction algorithms play a fundamental role in modern imaging systems by converting indirect, noisy, and incomplete measurements into interpretable clean images. Modern reconstruction methods leverage knowledge about the set of plausible images to improve the quality of the reconstructions. Throughout the years, various mathematical models have been proposed to capture the set of plausible images, with wavelets and total variation amongst the most popular ones. In recent years, these model-based methods have been replaced by data-driven ones, which incorporate information about the set of plausible reconstructions directly from data.

The predominant data-driven approach for designing image reconstruction algorithms consists of training a deep neural network (DNN) using a dataset of pairs of signals and measurements. DNNs have achieved state-of-the-art performance in a wide range of imaging inverse problems, from medical imaging to computational photography [1]. However, widespread deployment of deep learning-based solutions in scientific and medical imaging applications has been limited as i) it is often very expensive or even impossible to obtain large datasets of ground-truth signals for supervised training, and ii) supervised DNNs can fail to reconstruct structures which do not appear in the ground-truth training examples [2].

Recent advances in unsupervised learning have highlighted the possibility of learning to reconstruct images using measurement data alone [3-7], thus removing the need for ground-truth references as training data. This new generation of measurement-driven computational imaging has been successfully applied to many important problems where large amounts of ground-truth data are hard to obtain, such as accelerated magnetic resonance imaging (MRI) [5], microscopy [2], and sparse view computed tomography [5]. Despite these promising advances, most existing unsupervised approaches are limited to linear problems and cannot be applied to a large number of real-world applications where the measurement process is inherently non-linear, i.e., when measurements are quantized or phaseless (e.g., phase retrieval [8]), or when the linear operator is not fully known (e.g., blind calibration/deconvolution).

The PhD project will focus on the phase retrieval problem [8], where only phaseless measurements are available.
In this setting, obtaining large and diverse training datasets of signal and measurement pairs is particularly expensive (e.g., via holography), thus most DNN-based solutions use synthetic datasets for supervised training.

The goals of the project are to i) propose new practical deep learning-based unsupervised learning algorithms that learn incomplete and phaseless measurement data, ii) study theoretical guarantees for learning from measurement data that complement existing signal recovery theorems [8] (which characterize the conditions for which the reconstruction mapping is well-behaved), and iii) contribute to open source libraries (e.g., contributing to the deep inverse library (https://github.com/deepinv/deepinv )).

[1] G. Ongie, A. Jalal, C. A. Metzler, R. G. Baraniuk, A. G. Dimakis, and R. Willett, “Deep learning techniques for inverse problems in imaging,” IEEE Journal on Selected Areas in Information Theory, vol. 1, no. 1, pp. 39–56, 2020
[2] C. Belthangady and L. A. Royer, “Applications, promises, and pitfalls of deep learning for fluorescence image reconstruction,” Nature Methods, vol. 16, no. 12, pp. 1215–1225, 2019.
[3] J. Lehtinen, J. Munkberg, J. Hasselgren, S. Laine, T. Karras, M. Aittala, T. Aila, et al., “Noise2Noise,” in International Conference on Machine Learning (ICML), PMLR, 2018.
[4] J. Batson and L. Royer, “Noise2self: Blind denoising by self-supervision,” in Int. Conf. on Mach. Learning (ICML), 2019.
[5] D. Chen, Julián Tachella, and M. Davies, “Equivariant imaging: Learning beyond the range space,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pp. 4379–4388, 2021.
[6] Julián Tachella, D. Chen, and M. Davies, “Unsupervised learning from incomplete measurements for inverse problems,” in Advances in Neural Information Processing Systems (NeurIPS), 2022
[7] Julián Tachella, D. Chen, and M. Davies, “Sensing theorems for learning from incomplete measurements,” Journal of Machine Learning Research, vol. 24, no. 39, pp. 1–45, 2023
[8] J. Dong, L. Valzania, A. Maillard, T.-a. Pham, S. Gigan, and M. Unser, “Phase retrieval: From computational imaging to machine learning. A tutorial,” IEEE Signal Processing Magazine, vol. 40, no. 1, pp. 45–57, 2023

Requirements

Research Field

Engineering

Education Level

PhD or equivalent

Research Field

Computer science

Education Level

PhD or equivalent

Research Field

Mathematics

Education Level

PhD or equivalent

Languages

FRENCH

Level

Basic

Research Field

Engineering

Years of Research Experience

None

Research Field

Computer science

Years of Research Experience

None

Research Field

Mathematics

Years of Research Experience

None

Additional Information

Website for additional job details

https://emploi.cnrs.fr/Offres/Doctorant/UMR5672-JULTAC-001/Default.aspx

Work Location(s)

Number of offers available

1

Company/Institute

Laboratoire de Physique

Country

France

City

LYON 07

Where to apply

Website

https://emploi.cnrs.fr/Candidat/Offre/UMR5672-JULTAC-001/Candidater.aspx

Contact

City

LYON 07

Website

http://www.ens-lyon.fr/PHYSIQUE/

STATUS: EXPIRED