Job offer

Desired Skills: The candidate should hold a PhD in machine learning, computer vision, signal or image processing, applied mathematics, or a related field, and have a strong publication record. A good background in mathematics (optimization, probability, statistics) and strong programming skills in Python and PyTorch are expected. Prior experience with inverse problems, generative models or optimal transport will be appreciated.

Languages FRENCH

• Subsidized meals
• Partial reimbursement of public transport costs
• Leave: 7 weeks of annual leave + 10 extra days off due to RTT (statutory reduction in working hours) + possibility of exceptional leave (sick children, moving home, etc.)
• Possibility of teleworking (90 days / year) and flexible organization of working hours
• Social, cultural and sports events and activities
• Access to vocational training
• Social security coverage under conditions

2927 € gross salary / month Selection process

Applications must be submitted online via the Inria website. Processing of applications submitted via other channels is not guaranteed.

Information for Applicants: Please note that all public officials are required to comply with ethical rights and obligations

The post-doctoral researcher will be part of the Thoth team, which is specialized in computer vision and machine learning, and will be supervised by Julien Mairal. Thoth is one of the most recognized European research teams in machine learning and computer vision, is part of the Laboratoire Jean Kuntzmann at Université Grenoble Alpes, participates in several European projects (including ERCs), and has many international academic and industrial collaborations. The post-doc will join an active group of PhD students and post-docs working on generative modeling, optimization and inverse problems.

Location: The post-doc will work at the Inria centre at Université Grenoble Alpes. Grenoble is a lively city which hosts many foreign students and researchers. Located in the heart of the French Alps, its direct surroundings offer great outdoor recreation including skiing, cycling, and hiking. Grenoble is well-connected to Lyon and Geneva airports, and Paris can be reached from Grenoble in 3h by train.

The post-doc is expected to conduct high-quality research leading to publications in top-tier venues in machine learning, computer vision or signal processing, and will also have the oppor- tunity to interact with and co-supervise PhD students of the team.

Research Context: Inverse problems are ubiquitous in scientific and engineering domains that involve image acquisition, from computational photography to microscopy, astrophysics and medical imaging. The quantity of interest is observed through a forward process (blur, noise, subsampling, sensor effects) and must be recovered from these degraded measurements. Modern solvers, such as plug-and-play algorithms or methods based on diffusion and flow- based generative models, combine a model of the acquisition process with a powerful learned prior on clean signals, and achieve impressive results when the forward operator is known. In practice, however, the forward operator is rarely known precisely. Calibrating a real imaging system typically requires dedicated hardware or pairs of clean and degraded images, which are expensive or even impossible to acquire, for instance for a real camera lens. A recent line of work in the Thoth team has shown that the forward operator can instead be learned from unpaired data, by matching distributions both on the measurement side and on the reconstruction side, through a differentiable plug-and-play solver parameterized by the operator being learned. This opens the door to imaging systems that calibrate themselves from data, and to reconstruction methods that remain accurate on real-world acquisitions.

Objectives: The goal of the post-doc is to develop principled and scalable methods for learning imaging systems and solving inverse problems when supervision is scarce or unavailable. The research directions are deliberately open and will be shaped together with the candidate according to their interests and expertise. They include, for instance:

-ˆ extending unpaired operator learning to richer forward models (spatially varying, nonlinear, or with unknown noise statistics) and to new acquisition modalities; -ˆ jointly learning the forward operator and the reconstruction method, possibly leveraging modern generative models (diffusion models, flow matching) as priors; -ˆ understanding the theoretical aspects of the problem, such as identifiability of the operator from unpaired data and the interplay between calibration and reconstruction quality;

-ˆ designing models that can be trained across many devices or degradations and adapt quickly to new ones; ˆ- applying these methods to real-world imaging problems, in computational photography or in scientific imaging.