ICTEAM > ISPGroup > PostDocLearn2Sense


Open postdoctoral position:

Learn2Sense: Learning Strategies for Computational Sensing

Discipline keywords: computational imaging, non-linear inverse problems, deep learning, convex and non-convex optimization, learning physical models, generative priors, compressive sensing, lensless endoscopy, astronomy

Location: Mathematical Engineering Department (INMA), ISPGroup, UCLouvain, Belgium.

Mobility criterion: This position is open to all nationalities, including Belgian, provided the candidate has spent less than 12 months over the last 3 years in Belgium.


The research group of Prof. Laurent Jacques in the Image and Signal Processing Group (ISPGroup) of UCLouvain of Louvain-la-Neuve in Belgium opens one position for a postdoctoral researcher to work on the new project "Learn2Sense: Learning Strategies for Computational Sensing" funded by the Belgian Fund for Scientific Research - FNRS.

Learn2Sense Project

Learn2Sense aims at creating physics-driven learning strategies for solving inverse problems encountered in modern computational imaging (CI) techniques.

Currently, CI is mainly driven by two co-existing approaches. First, the "analytical method" (AM) is based on our ability to: (AM1) develop an accurate forward model of the sensing process, i.e., from the interaction between the object and a probing signal (e.g., an electromagnetic field, a pressure wave) to the collection of the disturbed probing signal by the sensing device; (AM2) determine an accurate prior model on the structure of the object signal (such as sparsity or low-rank representation models); (AM3) efficiently compute the forward sensing model, i.e., to accurately reproduce the observed data if some representation of the object signal is known; and (AM4) to finally solve the related inverse problem --- often ill-posed or under-determined --- by restricting it to plausible solutions thanks to the prior model. As a result, AM-based computational sensing often targets simple, intrinsically linear or linearized forward models (as in tomographic applications) combined with linear prior models so that one can solve the related inverse problem with, e.g., convex optimization.

Second, the more recent “learning method” (LM) directly learns to invert the sensing model; it proposes to bypass AM1-AM4 thanks to the supervised training of deep neural networks. Thus, LM relies on the collection of large datasets coupling (known) object signals with their corresponding observations. However, while achieving unprecedented results in, e.g., image denoising, deconvolution or superresolution, these methods reach good accuracy only for sufficiently large learning datasets and are prone to instability under slight sensing model corruption. In addition, they do not provide interpretable models, e.g., allowing us to improve our knowledge of the acquisition physics or the sensing system.

This project lies at the frontier between the AM and LM approaches, reducing both the requirements drawn in AM1-AM4 and the need for large datasets. In particular, this project will pursue the following research directions:

  • (D1) Learning of complex linear forward models that are found in specific optical computational imaging applications (e.g., in complex optical systems leading to spatially-varying convolutive kernels or more general integral operators);
  • (D2) Physics-driven learning of non-linear sensing models from hierarchical structures inspired by neural (NN) networks, such as those induced by light propagation in scattering media where NN layers are associated with physical depth or approximation levels.
  • (D3) Quantifying how deep generative priors (e.g., variational autoencoders and generative adversarial networks) improve the regularization of specific inverse problems of computational sensing, as non-linear alternatives to sparse and low-rank models.

These research directions will be fed both by a theoretical analysis of the proposed approaches, and by the design of novel computational sensing solutions for tomographic lensless endoscopy with ultrathin multicore optical fibers, as well as the analysis of new calibration and deconvolution techniques for astronomical imaging (such as direct telescopic imaging of exoplanets).

The project is associated with an existing collaborative network with local and international researchers specialized in deep learning, computational imaging, optics, astronomy and biology, namely, C. De Vleeschouwer and J. Lee (UCLouvain), O. Absil (ULiege), and H. Rigneault (Fresnel Institute, France).

Research description

In the general context described above, the postdoctoral researcher will more particularly focus his/her research for 2 years on one or several of the following topics:

  • Physics-driven learning: Design of physics-driven neural-networks for linear and non-linear computational imaging forward models (e.g., based on iterative approximation schemes, such as Neumann or Born series, on factorization of linear models into a product of structured matrices)
  • Deep generative priors (DGP): Leveraging the power of recent DGPs for regularizing the inversion of the imaging process; studying both theoretically and numerically how the compression power of these priors, i.e., their capability to summarize the object signal to a few parameters, drives the sensing rate (also called sample complexity) of a given computational sensing application beyond what is achievable with linear, sparsity-based priors.
  • Fast and stable inversion algorithms: Developing a practical inversion algorithm with reduced computational complexity by leveraging the iterative, neural-network structures of both the forward models and DGPs; stochastic gradient descent and stochastic proximal gradient methods will be considered in this research axis with applications to specific (non-linear) inverse problem settings.

Applicant's profile:


  • PhD in applied mathematics, electrical engineering, physics, applied statistics or related.


  • Specialization in at least one of the following topics: computational imaging, inverse problem solving, (deep) learning techniques, or compressive sensing.
  • Knowledge in at least one of the following topics is a plus: stochastic gradient descent, proximal methods, measure concentration, high-dimensional data processing, convex optimization.

Numerical and programming skills

  • Excellent programming skills in python;
  • Programming skills in common machine learning modules (e.g., PyTorch, Keras,TensorFlow, JAX) as well as in parallel/GPU computing is a plus;

Scientific and collaborative skills

  • Excellent communications skills, both written and oral, in English.
  • Knowing French *is not* required (the research group is international)

More information about the project can be obtained upon request.

We offer:

  • A research position in a dynamic environment, working on leading-edge theories and applications with international contacts;
  • A research team composed of one professor and 5 PhD students working on topics related to Learn2Sense; The candidate will also be involved in the guidance of these junior researchers.
  • A 24-month position funded by the Belgian NSF (FNRS) with a comfortable salary relatively to the cost of life in Belgium.
  • The funding is a scholarship, and Visa will be needed for a non-EU researcher.


First application deadline on Friday, February 14th, 2020.

Applicants are requested to send the following documents (all in pdf):

  1. a detailed CV;
  2. a list of publications;
  3. a brief statement of research interests (<= 2 pages) explaining why the candidate is interested in working in the research topics described above and how the open position is connected to his/her PhD background;
  4. two letters of reference.

Please send applications by email to:

Selection procedure

  • Pre-selection of candidates based on their application files;
  • Interview of the shortlisted candidates (remotely);
  • The successful applicant will be hired anytime before June 1, 2020, in agreement with the candidate.

Last updated March 03, 2020, at 07:41 AM