Imaging & machine learning meeting

January 29 - 31 2020 - Laboratoire I3S, Sophia-Antipolis

This is the website of the meeting Imaging & machine learning which will be held at the Laboratoire d'Informatique, Signaux et Systèmes de Sophia Antipolis (I3S) of Sophia-Antipolis, France, in the period January 29-31 2020.

Topics

The purpose of this meeting is to foster the creation of new collaborations between the Morpheme project team and the Machine Learning Genoa Centre on research topics ranging from theoretical and applied machine learning, computer vision, biological image analysis, inverse problems etc.

The meeting will start with a tutorial on Optimisation for machine learning lectured by Prof. Silvia Villa (DIMA) and Prof. Lorenzo Rosasco (DIBRIS).

Location & transports

The event will take place in the Amphithéâtre Forum and in the Conference Room of the laboratory I3S of Sophia-Antipolis.

  • The laboratory I3S can be reached from Antibes (Place de Gaulle/Dugommier) by bus 100 (closest stop: Templiers) or by the bus-tram A (closest stop: INRIA) and from Nice by bus 230 (closest stop: INRIA).
  • Several car parking spots are available outside the laboratory.

Tutorial Jan. 29 - Amphithéâtre Forum, Campus SophiaTech

!!! PLEASE NOTE THE ROOM CHANGE !!!

A three-hour tutorial on Optimisation for machine learning will be organised on Wednesday January 29 starting from 14:00 in the Amphithéâtre Forum, Campus SophiaTech (click here for directions). The course will be lectured by Prof. Silvia Villa (DIMA) and Prof. Lorenzo Rosasco (DIBRIS).

PROGRAM

  • 14:00 - 15:00: Statistical learning theory: empirical VS expected risk.
  • 15:00-16:00: Optimisation to solve regularised empirical risk minimisation: gradient, proximal gradient .
  • 16:00-16:30: Coffee break
  • 16:30-17:30: Iterative regularisation approach

The registrations to the tutorial are now closed. Confirmed participants have been informed of their successful registration by a confirmation e-mail.

Workshop Jan. 30-31 - Conference Room 007 I3S, Les Algorithmes, Bât. Euclide B

The workshop will be held in the conference room 007 of the I3S laboratory, Les Algorithmes, Bât. Euclide B (click here for directions).

THURSDAY 30 SCHEDULE (click to unroll)

  • 9:30-10:00: Welcome coffee
  • 10:00-10:15: Opening
  • 10:15-10:50: Xavier Descombes (Morpheme), Morpheme fundamentals and a few examples
Abstract: In this talk, I will first briefly present the main topics addressed in Morpheme team. Then, I will illustrate two exemples from my own work. First, I will consider object detection using marked point processes that can be seen as an extension of the Markov Random Field approach where the configuration space consists of random sets of random number of parametric objects. Second, I will show preliminary results to obtain median graphs within a simulated annealing scheme.Abstract: In this talk I will present the BMB initiative, a three year project funded by the Regione Liguria and the University of Genova, starting in May 2020. BMB, through the creation of a Center for BioMedical Big Data, aims at making P-Medicine (Personalized, Precision, Preventive, Predictive, Participated, and Proactive Medicine) reality. The spectrum of the problems we have to tackle is wide and includes legal and ethical aspects, issues related to data privacy, protection, and security, and the extensive effort for collecting the biomedical and health data flowing in the data lake, data of heterogenous type and often lacking standard. BMB will clearly leverage on existing partial solutions by inheriting the considerable amount of data already available in digital form. At the same time BMB will set in motion the full pipeline needed to manage, process, analyse, and visualize the available data by recruting data engineers and scientists trained to work on biomedical data. In the end we expect BMB to provide quantitative support to health policy regulators and constitute the basis for upscaling data-driven biomedical research. From day zero BMB seeks collaboration with similar initiatives and is open to establish connections with interested parties.
  • 11:25-11:45: Anca Ioana Grapa (Morpheme), Optimal Transport vs Many-to-many assignment for Graph Matching
Abstract: Graph matching for shape comparison or network analysis is a challenging issue in machine learning and computer vision. Generally, this problem is formulated as an assignment task, where we seek the optimal matching between the vertices that minimizes the difference between the graphs. We compare a standard approach to perform graph matching, to a slightly-adapted version of regularized optimal transport, initially conceived to obtain the Gromov-Wassersein distance between structured objects (e.g. graphs) with probability masses associated to the nodes. We adapt the latter formulation to undirected and unlabeled graphs of different dimensions, by adding dummy vertices to cast the problem into an assignment framework. The experiments are performed on randomly generated graphs onto which different spatial transformations are applied. The results are compared with respect to the matching cost and execution time, showcasing the different limitations and/or advantages of using these techniques for the comparison of graph networks. This comparison was also intended to serve as a basis for the development of a numerical quantitative model to characterize and compare biological networks (Extracellular Matrix variants) from confocal miscroscopy images.
  • 11:45-12:05: Matteo Moro (DIBRIS), A Deep Learning and Computer Vision-based Approach for Marker-less Human Motion Understanding
Abstract: Measuring and understanding human motion is crucial in several domains, ranging from neuroscience, to rehabilitation and sports biomechanics. Nowadays the study of human motion is commonly done through marker-based techniques and motion capture systems. If, from one side, these methods are precise and reliable, from the other they present some disadvantages, in particular they are expensive, invasive and time consuming. For these reasons in the last years the research of cheaper and more ecological markerless techniques had made a lot of progress. In particular we are focusing our study in the design and implementation of computer vision and machine learning algorithms which can be applied for marker-less human motion analysis. This type of analysis may facilitate the extraction of features that give qualitative and quantitative information about human motion and that can be used also for classification tasks. We first focus on a tool that allows a reliable semantic features detection in RGB videos. In this talk the logic behind this method will be presented together with examples of application and preliminary results.
  • 12:05-14:00: LUNCH BREAK
  • 14:00-14:35: Matteo Santacesaria (DIMA), Infinite-dimensional inverse problems with a finite number of measurements
Abstract: After a brief introduction to the Computational Harmonic Analysis & Machine Learning group of MaLGa, I will present a general stability and uniqueness result for inverse problems in infinite dimensional spaces when only a finite dimensional approximation of the measurements is available. This finds application to severely ill-posed problems such as electrical impedance tomography and inverse scattering. Related results on infinite-dimensional compressed sensing will be discussed as well.
  • 14:35-14:55: Matteo Monti (DIMA), Unitarization of the Radon transform on homogeneous trees
Abstract: During 20th century, the problem of inverting the Radon transform has been deeply studied because of its several applications. A classical use in the continuous case (e.g. two-dimensional signals) is the medical computed tomography. Recently, the Radon transform on discrete setups (e.g. graphs) also began to be treated and network tomography has been developed with the same philosophy of the algorithms used in CT scans. We analyze Radon on a discrete manifold: a homogeneous tree X. On X the set H of all the horocycles is defined, they are the analog of hyperplanes in Euclidean spaces. Given a signal f on X, its Radon transform Rf is defined at a horocycle h as the sum of values of f at vertices lying on h. We solves the unitarization problem: we show the existence of a pseudo-differential operator such that its postcomposition with R extends to a unitary operator Q from L²(X) to L²(H). Furthermore, such Q is an intertwining operator for the representations of the automorphism group of X on L²(X) and the one on L²(H), which are therefore unitary equivalent.
  • 14:55-15:30: Alin Achim (University of Bristol, visiting Morpheme), Computational Ultrasound Imaging – Some Inverse Problems
Abstract: This talk will focus on a range of inverse problems with application in both medical and biological ultrasound imaging. First, we will introduce a method for detecting straight lines (B-lines) in medical ultrasound images of the lungs for children with severe kidney conditions who undertake dialysis. We addressed this as a sparse estimation problem using both convex and non-convex optimisation techniques based on the Radon transform and sparse regularisation. This breaks into subproblems which are solved using the alternating direction method of multipliers (ADMM), thereby achieving line detection and deconvolution simultaneously. In the second part, we will describe a framework for compressive sensing (CS) data acquisition and reconstruction in quantitative acoustic microscopy (QAM). We designed and implemented an innovative technique, whereby the approximate message passing (AMP) method is adapted to account for QAM data statistics.
  • 15:30-16:00: COFFEE BREAK
  • 16:00-16:35: Fabienne de Graeve (Morpheme), An image-based high-throughput screen to identify modifiers of RNP granules
Abstract: The purpose is to study the molecular bases underlying the assembly and regulation of RNA granules using the highly conserved IMP-containing granules as a paradigm. Specifically, we propose to perform an unbiased genome-wide RNAi screen on Drosophila cultured cells to identify mutant conditions in which the organization and/or distribution of IMP-containing granules is altered. To quantitatively and statistically analyze mutant conditions, we combine high-throughput microscopy with the development of a computational pipeline optimized for automatic analysis and classification of cells. The main steps of this pipeline are: segmentation and classification of cell nuclei into living or dead on DAPI images, segmentation of living cell cytoplasms on GFP images, detection of IMP granules inside the segmented cytoplasms, and (upcoming work) statistical analysis of the granule distributions inside the cytoplasms depending on the conditions of the cell culture.
  • 16:35-16:55: Rudan Xiao (Morpheme), Analysis and classification of cellular and vascular markers in histological images: application to kidney cancer
Abstract: Kidney cancer is one of the most common malignancies worldwide. Over 85-90% of kidney cancers arise in the renal parenchyma, most of which are renal cell carcinoma (RCCs). The prevalence of clear cell RCC (CCRCC) and papillary RCC (PRCC) is the first and second, respectively, which is the focus of this study. We are interested in developing a system with high sensitivity and specificity to classify non-tumor tissue versus RCC and CCRCC versus PRCC within RCC. The main work includes: 1. Construction of tumor and non-tumor image databases. 2. Preprocess images and use Gabor filter to extract the vascular network of the tumor area. 3. On the one hand, the vascular network features are used directly for classification by using machine learning methods. On the other hand, we will construct two interacting graph models, the first is directly defined by the vascular network, the nodes of the second graph are defined by the nucleus, and the edges are defined by the spatial relationship between the cells, its correlation will show some relationships between tumor cells and vascular networks, and then use machine learning methods to classify based on these graphical features. 4. Verification will be performed in the database of Nice University Hospital.
  • 16:55-17:30: Francesca Odone (DIBRIS), Image and video analysis and the Shearlet transform
Abstract: Shearlets are a recent directional multi-scale framework for signal analysis, able to capture efficiently the anisotropic information in multivariate signals, which have been shown effective to enhance discontinuities such as edges and corners at multiple scales. In our recent work we have extented this analysis considering blob-like features in the image domain and space-time interest point on the image sequence domain. We have derived algorithms that incorporate interesing properties and allow us to address effectively problems such as the presence of noise and compression artifacts. In the talk I will present our work from theory to algorithms and show some applications of these algorithms to the real world, discussing the problems we are facing and the possible future research directions.

FRIDAY 31 SCHEDULE (click to unroll)

  • 9:00-9:35: Laure Blanc-Féraud (Morpheme), L0 sparse optimisation for super-resolution in fluorescence microscopy.
Abstract: In this talk we are interested in super-resolution fluorescence microscopy by Single Molecule Localization Microscopy (SMLM) methods. One way to reconstruct super-resolved images from SMLM acquisitions is to model the observations by a linear system which matrix has more columns (number of unknowns) than rows (number of observations) but which can be efficiently regularized by a sparse term involving the counting pseudo-norm l0. The reconstruction then consists in solving a minimization of a criterion composed of a least square (l2) term plus a l0 term. Such criterion is non continuous, non convex and its optimization is an NP hard problem. We have introduced an exact continuous relaxation which allows to use standard optimization algorithms. We have shown that global minimizers are preserved whereas some local minimizers are removed by this exact relaxation. Results are shown on super-resolution fluorescent biological images by SMLM and compare with standard IHT algorithm.
  • 9:35-9:55: Arne Henrik Bechensteen (Morpheme), New $\ell_2-\ell_0$ algorithms for single-molecule localization microscopy
Abstract: Among the many super-resolution techniques for microscopy, single-molecule localization microscopy methods are widely used. This technique raises the difficult question of precisely localizing fluorophores from a blurred, under-resolved, and noisy acquisition. In this work, we focus on the grid-based approach in the context of a high density of fluorophores formalized by a $\ell_2$ least-square term and sparsity term modeled with $\ell_0$ pseudo-norm. We consider both the constrained formulation and the penalized formulation. Based on recent results, we formulate the $\ell_0$ pseudo-norm as a convex minimization problem. This is done by introducing an auxiliary variable. An exact biconvex reformulation of the $\ell_2-\ell_0$ constrained and penalized problems is proposed with a minimization algorithm. The algorithms, named CoBic (Constrained Biconvex) and PeBic (Penalized Biconvex) are applied to the problem of single-molecule localization microscopy and we compare the results with other recently proposed methods.
  • 9:55-10:30: Daniele Calandriello (LCSL), Scaling machine learning algorithms with adaptive random embeddings
Abstract: Large feature spaces, and in the limit non-parametric and kernel models, can often be beneficial when learning, providing a rich representations necessary to accurately model the data. However many machine learning (ML) algorithms cannot be applied to complex high-dimensional problems due to their poor scalability when the number of input features grow. In this presentation we will present an overview of novel adaptive embedding approaches to compress high dimensional data into a more manageable low-dimensional representation. At the same time, we show that this compression does not hinder the learning process, which provably continues to converge at statistically optimal rates. Finally, we show that these embedding approaches are quite flexible and find application in various applications such as clustering, black-box/Bayesian optimization, iterative convex optimization and more.
  • 10:30-10:50: Nicolò Pagliana (DIMA), Not so Fast: Learning with Accelerated Methods
Abstract: We study learning properties of accelerated gradient descent methods for linear least-squares in Hilbert spaces in terms of corresponding learning error bounds. Our results show that acceleration can provides faster bias decay than gradient descent, but also suffers of a more unstable behavior. As a result acceleration cannot be in general expected to improve learning accuracy with respect to gradient descent, but rather to achieve the same accuracy with reduced computations. However these behaviors are not always seen in practice, in this direction we show that a misspecified model can significantly slow down the performances of those algorithms.
  • 10:50-11:15: COFFEE BREAK
  • 11:15-12:15: Concluding remarks, discussion & round table
  • 12:15-14:00: LUNCH BREAK & GREETINGS

Organisation

  • Luca Calatroni (calatroni 'at' i3s.unice.fr): Chargé de recherche CNRS, UCA, INRIA.
  • Matteo Santacesaria (matteo.santacesaria 'at' unige.it): Assistant Professor, DIMA, University of Genova.

Support

The event is supported by the CNRS, the Academy "Systèmes complexes" of UCA (IDEX JEDI) and the laboratory I3S.