Thematic quarters on Artificial Intelligence for Signal and Image Processing
From 12/07/2021 to 23/07/2021
●Philippe Ciuciu (CEA NeuroSpin)
●Frédéric Pascal (L2S / CentraleSupélec)
●Charles Soussen (L2S / CentraleSupélec)
●Bertrand Thirion (INRIA Paris-Saclay)
●Alexandre Gramfort (INRIA Paris-Saclay)
●Emilie Chouzenoux (CVN / INRIA Paris-Saclay)
●Matthieu Kowalski (L2S / Université Paris-Sud)
●Agnès Desolneux (CMLA / CNRS)
●Anna Kazeykina (Université Paris-Sud / Dépt. de Mathématiques
The goal of the program is to gather researchers from the fields of signal and image processing and machine learning and to bridge the gap between both communities. The recent years have witnessed a tremendous interest in Artificial Intelligence and especially Deep Learning for massive and heterogeneous data analysis. Learning and extracting information from multivariate and possibly heterogeneous data is nowadays a major challenge. The deep learning approach has gained a lot of popularity because it yields very promising empirical performance compared to state-of-the-art approaches. For instance, in medical image analysis, supervised learning has become extremely popular, due to the availability of large annotated data making it possible to learn a statistical model that can then be used to analyse novel unseen data. Recent progress has been achieved to propose semi-supervised approaches in which the task of annotation is alleviated. Although the deep learning approaches often bring impressive results thanks to the convolutional neural net architecture, the reports in the literature are often empirical, and there are many open questions related to:
● the theoretical understanding of the deep learning approach;
● learning the structure of deep networks;
● deep network ablation and model compression;
● the empirical definition of the regimes where deep learning approaches outperform traditional approaches;
● quantification of uncertainties of the learning process: stability with respect to small variations in the training data, robustness or insensitivity to noise present in the training data, in connection with the concept of adversarial learning;
● the reduction of computing time with deep learning.
These questions are closely related to the knowledge of researchers in the fields of robust statistics and inverse problems widely used in signal and image processing. Indeed, robust statistics approaches aim at generalising classical statistical analyses to various cases, notably when the data are heterogeneous or contain outliers. Those methods are also adapted to the case of missing or incomplete data. Recent advances handled cases of large dimensional data sets, fitting more practical problems.
Nowadays, robust statistics could offer a convenient methodology to bring some explainability to deep learning methods together with helping to theoretically guarantee the robustness of algorithms. The inverse problem community gathers experts in numerical optimisation for solving large scale inverse problems (e.g., tomographic reconstruction in medical imaging or electro/magneto-encephalography for brain activity analysis or computational imaging problems in electron microscopy to cite a few) as well as experts in computational statistics, whether it relies on bootstrap and permutation methods or on Bayesian inference for uncertainty quantification. The latter approach relies on the probabilistic modelling (using possibly hidden variables) of the signal or image of interest and allows one to deliver “error-bars” on the reconstructed signals and images and on hyperparameters as well.
The AI approach for data analysis has become very popular in many application fields including health and biological engineering, non-destructive testing of materials, seismology, hyperspectral image analysis in astrophysics, to name a few. Since the proposed program in 2021 is only 2-week long (with a desire to propose a second 2-week edition in 2022) and essentially oriented towards methodological aspects of statistical signal processing, covering many application fields is not reasonable. Therefore, we will focus on medical data analysis in neuroscience and computational imaging for health application as showcase applications of the methodological issues outlined above. The evaluation of the signal and image methods on common biomedical data sets is an important objective. We target to develop common open-source software between some participants, and possibly data-challenges between student attendees.
The scientific organising team share strong knowledge in theory and methods in statistical signal processing and machine learning, and also in biomedical data analysis.
P. Ciuciu is Research Director at NeuroSpin, the largest high field MRI center dedicated to cognitive and clinical neuroscience in France. His research interests are inter-disciplinary ranging from signal and image processing (Compressed Sensing for MR Imaging) to functional brain imaging (fMRI, MEG) for applications to cognitive neuroscience. Since 2017 he has been leading the transversal research project, called “Toward 500 µm fMRI” at NeuroSpin.
F. Pascal is Full Professor at L2S / CentraleSupélec. Since Jan. 2017, he has been the head of the “Signals and Statistics” group of L2S. He is also the coordinator of the data science activities at CentraleSupélec. From Sept. 2017, he has been appointed as the Program Coordinator in the Executive Committee of the DATAIA institute. His research 3 interests include estimation, detection and classification for statistical signal processing and applications in radar and image processing.
C. Soussen is Full Professor at L2S / CentraleSupélec. His research interests are in inverse problems and sparse approximation. He is the principal investigator of the French ANR project BECOSE (Beyond Compressive Sensing: Sparse approximation algorithms for ill-conditioned inverse problems, 2016-2020). B. Thirion is Research Director at Inria and is the head of the Parietal team at INRIA Paris Saclay. His research focuses on mathematical methods for statistical modelling of brain function using brain imaging data, with a particular interest in machine learning techniques. He contributes applications to human cognitive neuroscience and software development. Since Jan. 2019, he has been the director of the DATAIA institute.
Structure of the program, on a week-by-week basis
The expected participation per week expected is about 30 people, but this number could range between a minimum of 15 and a maximum of 50.
Since the program lasts 2 weeks, all 30 participants are expected to stay during the full 2-week duration according to the minimum 2 week policy of the Call for Proposals.
The organisation of the 2 weeks is summarised in the following tables:
The sessions appearing in blue are carried out in the small offices of the IPa building, with a team of reduced size (from 2 to 6 participants) per team. The other sessions are plenaries. In the first week, most morning sessions will be devoted to tutorials on theory and methods of machine learning and signal processing, which is a prerequisite for the team work that will take place during the 2 weeks. It is expected that the participants will not all have the same background, so these tutorials are crucial. Because the field of machine learning has been growing rapidly, it is very likely that all participants will not have the same background, hence the requirement to organise a number of comprehensive tutorials. The afternoon sessions will start by 1-hour pitchs (Monday, Tuesday) during which some of the participants will briefly introduce their main aim. This will be followed by spontaneous team working sessions in separate rooms. These working sessions will take the following forms depending on the groups: 1. meeting of colleagues working on theoretical aspects of signal processing and machine learning, such as conception of new algorithms or theoretical analysis of algorithms; 2. coding sprints for open-source software development; 3. journal clubs aiming to survey and critically evaluate articles of the literature. A poster session will take place on Wednesday afternoon for the PhD students and postdocs (and any other participant willing to present a poster) and will be followed by a social event on Wednesday evening. Thursday will be fully dedicated to working groups without tutorials in the morning. On Wednesday and Friday, scrums will be organised in early afternoon, where the different teams will shortly debrief the work they have already achieved. A (non-exhaustive) list of possible topics of tutorial talks can be found below: 1. An introduction to robust statistics 2. An introduction to inverse problem theory and algorithms: variational and Bayesian approaches, pros and cons of the two! 3. Deep learning in inverse problems 4. Deep learning in computational imaging 5. New trends for multidimensional signal analysis in neuroscience 6. Convolutional sparse coding with applications to medical imaging
The second week is organised similarly to the first one up to a few adaptations. The 2 tutorials in the morning are replaced by a Scientific Presentation (SP) in which a participant will either:
● present to the audience a team work that has been done in the first week
● present more in-depth an open topic of interest for other participants
The week will end with a global synthesis of the team works, their achievements, and the perspective of ongoing research and open-source software development for the future months. Because we cannot dedicate one full month to the IPa program with respect to other duties and activities (teaching and other tight research responsibilities), we chose to propose a 2-week program. We realise that 2-week may be thought of as a short duration, but we expect that there will be interesting outcomes in terms of new methods, new softwares and applied results as well. We expect that the results at the end of the program will allow us to propose a new program in 2022 (in response to a future call of the IPa) on this topic around the same core group of participants than in the program of 2021.