skip to content

Generative models, parameter learning and sparsity

30th October 2017 to 3rd November 2017
Michael Hintermüller Humboldt-Universität zu Berlin
Simon Arridge University College London
Martin Burger Universität Münster
Alfred Hero University of Michigan
Nick Kingsbury Trinity College, Cambridge
Gabriel Peyre CNRS - Ecole Normale Superieure Paris
Guillermo Sapiro Duke University
Carola Schönlieb University of Cambridge

Workshop Theme

A key issue in image reconstruction, and in inverse problems as a whole, is the correct choice of image priors (or regularisation functionals) and data models (or fidelity terms) in a variational or Bayesian reconstruction model. Depending on the setup of the model, very different qualitative image reconstruction results are obtained. A setup of a variational imaging approach is influenced by the type of image one aims to reconstruct, as well as the way the image or data is acquired. The knowledge of the image properties -- such as the regularity of the image or present scales of image structures -- and the capability of modelling them, are crucial for an accurate setup of the image prior and as such for faithfully reconstructing the image contents. The image prior can have various forms, such as a regularisation term or a basis in which the image should be expanded. Sparsity plays a central role here. Sparsity promoting regularization is a widespread and very popular approach to solve inverse problems. Standard SPR methods like total variation (TV) or l1 regularization have been shown to be powerful tools to recover inverse problems solutions from a reduced amount of noisy measurements. Nevertheless, despite their ability of capturing important features such as discontinuities, these model-based regularizations are also well known to produce artefacts, such as the staircasing effect, if the measured data does not fit the corresponding model. An ideal SPR for a given application should be tailor-made, and reconstruct solutions one would expect rather than to create best fits to a standardized model.
The mechanism of the data acquisition process embodies the data model. This model explains how the data is related to the underlying image, containing information about the noise distribution, the amount of under-sampling and the physics of the image acquisition technique. Several strategies for deriving an optimal choice for an image enhancement approach have been considered in the literature. More heuristic approaches derive the model setup from the physics behind the acquisition process. Statistically grounded approaches are more data driven in the sense that they estimate or learn the noise and structure from the data itself. Adaptive regularisation approaches for instance are capable of adjusting the parameter values locally taking into account the noise level and the local scale of structures in the image. Moreover, machine learning methods, e.g. dictionary learning, are very powerful techniques to determine the correct basis in which an image should be reconstructed. Recent approaches in the community also propose to learn the imaging model by bilevel optimisation techniques. For gaining more insight into the reconstruction abilities of regularisers their analysis via singular vectors has also proven valuable in some recent works in the community.

Workshop Speakers include:

  • Francis Bach
  • Thomas Pock    
  • Rebecca Willett
  • Bob Plemmons
  • Pierre Weiss      
  • Raymond Chan
  • Xavier Bresson
  • Joan Bruna
  • Richard Baraniuk             
  • Jeff Calder         
  • Julianne Chung
  • Julie Delon         
  • Mario Figueiredo            
  • Eldad Haber       
  • Anders Hansen
  • Lior Horesh        
  • Peyman Milanfar            
  • Marcelo Peyrera             
  • Lars Ruthotto    
  • Pierre Vanergheynst     
  • Ozan Oektem
  • Gitta Kutyniok

Deadline for applications: 3rd September2017

Apply now

Please note members of Cambridge University are welcome to turn up and sign in as a non-registered attendee on the day(s) during the workshop and attend the lecture(s). Please note that we cannot provide you with any support including name badge, meals or accommodation.

In addition to visiting the INI, there are multiple ways in which you can participate remotely.



Registration Only
  • Registration Package: £227
  • Student Registration Package: £177

The Registration Package includes admission to all seminars, lunches and refreshments on the days that lectures take place (Monday - Friday), wine reception and formal dinner, but does not include other meals or accommodation.

Formal Dinner Only
  • Formal Dinner: £50

Participants on the Registration Package, including organisers and speakers, are automatically included in this event. For all remaining participants who would like to attend the above charge will apply.


Unfortunately we do not have any accommodation to offer so all successful applicants will need to source their own accommodation. Please see the Hotels Combined website for a list of local hotels and guesthouses.



Lunch will be served at Wolfson Court in the Cafeteria from 12:30 to 13:30 on days that lectures take place.

  • and Registration Package participants should present their badge as payment for their meal
  • Those issued with a blue Institute door entrance card can add money onto the card via the Porters' Lodge at Wolfson Court
  • Other participants must purchase their meal using their dining card via the Porters' Lodge (forms can be found on the registration desk or at the Porters' Lodge)

Evening Meal

Participants are free to make their own arrangements for dinner.

Formal Dinner

The Formal Dinner will take place on Thursday 2nd November at Corpus Christi College. Participants on the Registration Package, including organisers and speakers, are automatically included in this event.

University of Cambridge Research Councils UK
    Clay Mathematics Institute London Mathematical Society NM Rothschild and Sons