Projects Co-ordinator

Prof. Mark Plumbley, Centre for Vision, Speech and Signal Processing, University of Surrey, UK.

Projects Administrator

Dr Helen Cooper, Centre for Vision, Speech and Signal Processing, University of Surrey, UK.

Contact via Project Mailboxes

SpaRTaN-MacSeNet Workshop on Sparse Representations and Compressed Sensing

23rd March 2018, INRIA, Paris, France

This one-day workshop will bring together researchers working in the area of sparse representations and compressed sensing to find out about the latest developments in theory and applications of these approaches, and to explore directions for future research.

The concept of sparse representations deals with systems of linear equations where only a small number of the coefficients are non-zero. The technique of compressed sensing aims to efficiently sense and reconstruct a signal from few measurements, typically by exploiting the sparse structure of the underlying representation. These techniques have proved very popular over the last decade or so, with new theoretical developments, and successful applications in areas such as hyperspectral imaging, brain imaging, audio signal processing and graph signal processing.

This one-day workshop, organized by the SpaRTaN and MacSeNet Initial/Innovative Training Networks*, will include invited keynote talks by Karin Schnass (Universität Innsbruck, Austria) and Jean-Luc Starck (CEA-Saclay, France), oral presentations and posters. The talks and posters will include theoretical advances in sparse representations, dictionary learning and compressed sensing, as well as advances in areas such as brain imaging and MRI, hyperspectral imaging, audio and visual signal processing, inverse imaging problems, and graph-structured signals.

The workshop will take place at INRIA, 2 rue Simone Iff, 75012 Paris, 10:00 - 16:30 and there will be an opportunity for discussions to continue after the end of the formal workshop. Please register your intention to attend via Eventbrite.

PhD students and Early Career Researchers wishing to bring along a poster of their work for the poster session are encouraged to contact with a brief abstract of their work. Posters do not need to be novel: this is an opportunity to showcase work and discuss it with others in the field.

* European Union's Seventh Framework Programme (FP7-PEOPLE-2013-ITN) under grant agreement n° 607290 SpaRTaN and H2020 Framework Programme (H2020-MSCA-ITN-2014) under grant agreement n° 642685 MacSeNet


The workshop is being held at INRIA, 2 Rue Simone IFF, 75012 Paris, France


Please register your intention to attend via Eventbrite so that we can arrange catering and prepare your name badge.

Poster Session

PhD students and Early Career Researchers wishing to bring along a poster of their work for the poster session are encouraged to contact with a brief abstract of their work. Posters do not need to be novel: this is an opportunity to showcase work and discuss it with others in the field.

Design and Layout

The boards are portrait with size : 1m wide by 2m tall so they will fit A0 portrait (or a bit larger).

There are lots of resources on the web for how best to design a poster I've collected a few here but do take a look around yourself and if you find any other nice tutorials please let us know so we can share them:

University of Leicester's 20 min tutorial
Berkley's how to design a poster slides
A nice blog post covering the same as the above for those who like the texty approach

Friday 23rd March


Registration & Coffee



Francis Bach (INRIA) & Mark Plumbley (Uni. Surrey)


Keynote: Dictionary Learning: from local to global and adaptive

Karin Schnass (Universität Innsbruck, Austria)

In this talk we first give relaxed conditions that guarantee one iteration of an alternating dictionary learning scheme to contract an estimate for the desired generating dictionary towards this generating dictionary.

Conversely we will provide examples of dictionaries not equal to the generating dictionary that are stable fixed points of the alternating scheme. Based on these characterisations we then propose a (cheap) replacement strategy for alternate dictionary learning to avoid local minima. Finally we will discuss how the replacement strategy can be used to automatically determine the dictionary size and sparsity level.

Karin Schnass holds an MSc in Mathematics from the University of Vienna (AT), 2004, and a PhD in Computer, Communication and Information Sciences from EPFL (CH), 2009. Following a postdoc at RICAM Linz (AT) and two maternity leaves, she spent 2 years as Schroedinger fellow at the University of Sassari (IT). Since 2015 she is assistant professor at the University of Innsbruck, where she is heading the FWF-START-project ‘Optimisation Principles, Models & Algorithms for Dictionary Learning' (up to 6 yrs, 1.2M€). She is an expert on (theoretical) dictionary learning.


Oral Session 1


Efficient Algorithms For Large-Scale Datasets

Milad Niknejad (Instituto de Telecomunicações), Damien Scieur (INRIA), Francis Bach (INRIA)

Recent advances in data science has led to huge-size datasets of different kinds such as ImageNet. Having a massive amount of examples can noticeably help the improvement of machine learning algorithms. However, traditional methods for machine learning are often not scalable on these large-scale datasets. In this presentation, we address these issue by presenting two methods. In the first part of our talk, we present a general method which can be implemented on any large-scale dataset. In the second part, we focus on a specific application of image restoration with large-scale dataset of image patches.


SPARTAN advances in Compressed Sensing for MRI/Hyperspectral Imaging

Wajiha Bano (Uni. Edinburgh), Konstantinos Pitas (EPFL), Lina Zhuang (Instituto de Telecomunicações), Zhongwei Xu (Noiseless Imaging), Mike Davies (Uni. Edinburgh)

Compressed sensing (CS) has emerged as an area that opens new perspectives in signal acquisition and processing. It has provided an alternative approach to the traditional sampling theory attempting to reduce the required number of samples for successful signal reconstruction. In practice, CS aims to provide reduction in sensing resources, transmission and storage capacities and to facilitate signal processing in the circumstances when certain data are unavailable. Compressed Sensing and its associated sparsity models have now influenced aspects of signal processing from signal acquisition, signal/image analysis and even learning in neural networks. In this presentation we will give a rapid summary of the work that we have done in this WorkPackage progressing the state of the art from MRI and hyperspectral imaging to 3D modelling and compressing Deep Neural Network representations.


Audio Visual Sparse Representation

Alfredo Zermini (Uni. Surrey), Cian O'Brien (Uni. Surrey), Rodrigo Pena (EPFL), Lucas Borges (Noiseless Imaging), Benjamin Ricaud (EPFL), Pierre Vandergheynst (EPFL)


Lunch & Posters


Oral Session 2


Novel theory and algorithms for machine learning and signal processing

Tatiana Shpakova (INRIA), Dmitry Babichev (INRIA), Junqi Tang (Uni. Edinburgh), Francis Bach (INRIA)

In this talk we briefly present our work, which is dedicated to investigate new theory and algorithms for robust and efficient machine learning and signal processing. We begin with the discussion of the learning and inference in probabilistic models, which rely on various effective optimization methods. Then, we are going to talk about methods on efficient stochastic and big data optimization algorithms which are highly-demanded in a wide range of applications. All of these domains (learning, inference, stochastic and big data optimization) are tackled by our adjacent work packages which are more application-focused.


MacSeNet advances in Brain Imaging and Analysis

Arnold Benjamin (Uni. Edinburgh), Dhritiman Das (Technische Universität München), Cagdas Ulas (Technische Universität München), Manuel Morante (Computer Technology Institute, Athens), Christos Chatzichristos (Computer Technology Institute, Athens), Mike Davies (Uni. Edinburgh)

Brain imaging, particularly through Magnetic Resonance Imaging (MRI) has had significant impact on the study and diagnosis of the Brain. A typical MRI examination must be able to address issues such as clinical time constraints and local motion (i.e. due to breathing, beating heart etc.) because its performance is constrained by a slow data acquisition process which limits the imaging speed and makes it susceptible to motion artefacts. This has led to the development of several techniques that focus on accelerating MRI data acquisition, as well as fusing such information with other fast time modalities such as EEG for functional imaging. While much recent work has focused on compressed sensing and parallel imaging based acceleration, there are emerging techniques that have a more data driven and machine learning flavour. These aim to process data and extract the salient information via appropriate data models (tensors, manifolds, decision trees, etc.). This presentation will give a quick overview of the work that we have done in this WorkPackage developing new machine learning and data driven techniques for Quantitative and dynamic MRI, MR spectroscopy, and functional brain imaging.


Beyond Linearity and Convexity in Imaging Inverse Problems

Joshin Krishnan (Instituto de Telecomunicações), Marina Ljubenovic (Instituto de Telecomunicações), Nasser Eslahi (Tampere University of Technology), Cristóvão Cruz (Noiseless Imaging), Mario Figueiredo (Instituto de Telecomunicações)

Linearity and convexity are usual assumptions in image inverse problems (IIPs). But many practical imaging scenarios cannot be efficiently modelled in a linear or a convex framework. This work focuses on pushing research in IIPs beyond such assumptions. The objectives of this workpackage are optimization methods for non-convex formulations of non-convex IIPs, algorithms for non-convex regularization, theoretical guarantees and applications to phase imaging and image super-resolution. This talk discusses four main research directions namely 1) phase imaging via sparse coding in the complex domain, 2) patch-based, non-local and dictionary-based methods for blind image deblurring, 3) improved sparse signal recovery via adaptive correlated noise model and 4) non-local HOSVD methods for denoising and super-resolution imaging.


Audio Machine Sensing

Lucas Rencker (Uni. Surrey), Iwona Sobieraj (Uni. Surrey), Stylianos Mimilakis (Fraunhofer IDMT), Gerald Schuller (Fraunhofer IDMT)

Audio machine sensing aims at analysing and understanding sounds around us, such as music, speech, or environmental sounds. This talk will address some of the recent advances in audio machine sensing, for audio-related problems such as source separation, audio restoration or audio event detection, using techniques such as sparse decomposition, matrix factorization or deep learning.


Beyond Traditional Signals

Youngjoo Seo (EPFL), Volodymyr Miz (EPFL), Benjamin Ricaud (EPFL), Pierre Vandergheynst (EPFL)

Traditional signals such as 1-d(audio), 2-d(image) and 3-d(video) possess a well-defined structure (spatial and temporal). However, many datasets that are nowadays collected and need to be analyzed, for example texts and documents or events or activity on (social) networks, are unstructured. In addition, the average size of datasets is rapidly increasing and this requires adapted tools able to process large amount of data to retrieve information. The purpose of this work package is to formulate new models and algorithms to process data beyond the tradition framework. One direction for these new approaches is to introduce graphs which are generic data representations that create a structure on the data. The graph structure makes it possible to use traditional algorithms on unstructured data and open new doors for data analysis. A second direction is to go beyond traditional signal processing by creating new signal processing methods inspired by the emerging field of machine learning. Deep learning is now the best approach for pattern recognition in images and outperform all the previous / traditional computer vision algorithms. We investigate such approaches on structured as well as unstructured signals and datasets.


Coffee & Posters


Keynote: A Compressed Sensing perspective on astronomical image acquisition and processing

Jean-Luc Starck (CEA-Saclay, France)

We will present how Compressed Sensing (CS) and sparse recovery idea impact the way astronomers process existing data set, or design new instruments. Interferometric images or gammay ray images acquired through coded masks are two typical examples where data can be interpreted through the CS perspective. We will then focus on the radio-astronomy case. We will present a new sparse recovery method for LOFAR and SKA, which allows to perform jointly a deconvolution and a blind source separation.

Jean-Luc Starck is Director of Research and head of the CosmoStat laboratory at the Institute of Research into the Fundamental Laws of the Universe, Service d'Astrophysique, CEA-Saclay, France. Jean-Luc Starck has a Ph.D from Nice Observatory and an Habilitation from University Paris XI. He was a visitor at the European Southern Observatory in 1993, at UCLA in 2004 and at Stanford’s Department of Statistics in 2000 and 2005. Since 1994, he is a tenured researcher at CEA. He created in 2010 the CosmoStat laboratory and is strongly involved in the Euclid ESA space mission. He received the EADS prize of the French Academy of Science in 2011.

He leads Cosmostat, an interdisciplinary research group at the interface between cosmology and statistical methods with a focus on industry-academia partnership. He has organized 16 conferences, and was keynote, invited or seminar speaker over fifty times in the last five years. Over the last 10 years, he has been involved as Co-I or PI in the management of 8 million euros of grants from national, European and international sources, including a senior ERC. He has published over 200 refereed papers in astrophysics, cosmology, signal processing and applied mathematics, and he is also author of three books.


Closing Remarks

Mark Plumbley (Uni. Surrey, UK)


Discussions continue.