Context
This project will develop a novel framework for solving inverse problems in imaging, focusing on the joint reconstruction of a sequence of high-quality images from multiple degraded or incomplete observations. These inverse problems are typically addressed using neural network models that explicitly respect the underlying variational structure, enhancing both interpretability and adaptability. However, most existing approaches rely on discretized formulations that are sensitive to resolution changes, resulting in poor generalization across scales. To overcome this, we propose to operate directly in function spaces, ensuring discretization invariance and enabling models trained at one resolution to generalize seamlessly to others. In this context, we will define and explore an extension of unfolded proximal networks—where the network architecture is derived from the variational formulation of the inverse problem—into function spaces, leading to unfolded proximal neural operators.
The postdoctoral researcher will:
- Define and implement deep unfolded proximal neural operators.
- Develop self-supervised training strategies suitable for scenarios without ground-truth data.
- Provide theoretical guarantees by deriving generalization estimates.
- Analyze real data from experiments using holographic microscopy.
Candidate Profile
- PhD in applied mathematics, machine learning, signal/image processing, or related fields
- Strong background in optimization and inverse problems
- Experience with proximal methods, statistical learning, or operator theory is a plus
- Good programming skills in Python; experience with PyTorch preferred
- High proficiency in written and spoken English
More Information
You can download the full postdoctoral position description here: 2025 Postdoctoral Offer – PROXIMA at LabHC (PDF)
