All Years Seminars
[INMA] 2020-05-19 (13:30) : Two 20-minute talks given by the students (1/2)
At MS Teams
Section 1: Random Projections, Compressive Sensing and Compressive Classification
Speaker :
Mattéo Couplet
Abstract : About a decade ago, groundbreaking papers revealed the possibility of faithfully recovering high-dimensional signals from some incomplete information about them. They gave rise to the field of compressive sensing, which is now in a mature state and whose foundations rely on an elegant mathematical theory. This talk will start with an overview of the field. Then, the compressive classification task will be addressed, for which it can be shown that even fewer compressive measurements are required than for signal reconstruction, using ideas from high-dimensional convex geometry.
Section 2: Proximal methods
Speaker :
Calbert, Julien
Abstract : In this talk, we will discuss a particular class of optimization algorithms: proximal algorithms. These algorithms are well suited to solve convex, non-smooth, constrained and large-scale optimization problems. They are based on the successive resolution of a convex optimization problem called proximal operator. These methods are of great interest because they work under general conditions, they can be fast if a closed form can be derived for the proximal operator and under certain conditions they can be solved in a distributed way. In addition, proximal methods can be seen as an additional level of abstraction to the classical optimization methods. They form a theoretical framework for demonstrating simultaneously the convergence of several optimization methods such as the gradient method or the projected gradient method. Finally, we will see several applications in the field of signal processing.
[INMA] 2020-05-12 (14:00) : Prediction and uncertainty in the control of visual tracking
At MS Teams
Speaker :
Lefèvre, Philippe
Abstract :
Vision plays a key role in behaviour. In humans, visual acuity is non-homogeneous across the visual field and is characterized by a foveal zone of very high resolution. Therefore, for human observers, the quality of visual inputs is heavily influenced by their ability to detect and fixate parts of the visual scene that are of particular interest. When everything is stationary if the environment, this is performed with saccades that are fast orienting eye movements allowing to scan the visual scene. However, in a dynamic environments, objects can move with respect to the observer, requiring to track moving objects with smooth pursuit eye movements.
Saccades and smooth pursuit eye movements are two different modes of oculomotor control. However, behavioural and neurophysiological data demonstrated that both types of eye movements work in synergy for visual tracking. This suggests that saccades and pursuit are two outcomes of a single sensorimotor process that aims at orienting the visual axis. More recently, it has been hypothesized that oculomotor behaviors integrate sensory (visual) and prior information to overcome sensory-motor delays and noise. After much debate about this process, reliability-based integration has recently been proposed and several models of smooth pursuit now include recurrent Bayesian integration or Kalman filtering.
In this talk, I will present some recent experimental and modeling work demonstrating the key role played by prediction and uncertainty in the synergy between saccades and smooth pursuit in the control of visual tracking.
[INMA] 2020-05-05 (14:00) : Identifiability and Detection of Multiset Correlation Structure
At MS Teams
Speaker :
Marrinan, Tim
Abstract :
Relationships between random vectors with deterministic but unknown linear mixing are commonly assessed via canonical correlations. Detecting this correlation structure allows us to infer unseen associations between the underlying variables. For example, this type of analysis has been used to identify conditions that lead to a rise in sea surface temperature and to localize regions of brain activity used for visual recognition tasks. For pairs of random vectors, general conditions have been discovered under which the correlation structure is identifiable, but for collections of more than two random vectors such general conditions remain evasive. In this lecture we describe conditions under which a subset of multiset correlation structures can be identified, and demonstrate how the correlation structures can be detected from the augmented coherence matrix of the data.
[INMA] 2020-04-28 (14:00) : Challenges in Open Multi-Agent Systems Subject to Arrivals and Departures
At MS Teams
Speaker :
Hendrickx, Julien
Abstract :
Scalability and robustness to agent losses are often cited as advantages of multi-agent systems, but almost all theoretical results apply to system with fixed compositions. We consider open systems, that agents continuously leave and join during the process considered. We discuss the general challenges to analyze and design algorithms for such systems.
Challenges in the analysis come from the fact that each arrival or departure implies a discontinuity and a change of the system state dimension. Moreover, these repeated events forbid any asymptotic convergence.
On the design side, arrivals perturb the system and may also modify the algorithms objective. Departures result in information loss, or conversely, in a persistent influence of information that is no longer relevant. Moreover, correction mechanisms designed to cope with a small number of arrivals or departures may fail or even have a counterproductive and destabilizing effect, when these events keep taking place.
We focus in particular on averaging, decentralized estimation, computation of the maximum value and decentralized optimization. We also present some fundamental performance limitations in open systems.
[INMA] 2020-04-21 (14:00) : Deep Learning in Computer Vision
At MS Teams
Speaker :
De Vleeschouwer, Christophe
Abstract :
Deep learning and more specifically Convolutional Neural Networks (CNNs) are now predominant in computer vision. In this seminar, we survey a variety of image processing problems and present how they are formulated in terms of CNNs optimization, using image-wise or pixel-wise predictions. We then introduce some of the practical issues encountered when training and deploying neural networks in real-life, which reveals that a major shortcoming of neural networks lies in the poor understanding we have about their inner processes and generalization capabilities. Since a solid mathematical framework is still lacking to explain neural networks performance, we summarize a few recent experimental results revealing the tight connection between network pruning and generalization. Those results suggest to look for an explanation in the way neurons split data in diverse and representative binary classes.
Seminars
Subscribe (ICS)
INMA Contact Info
Mathematical Engineering (INMA)
L4.05.01
Avenue Georges Lemaître, 4
+32 10 47 80 36
secretaire-inma@uclouvain.be
Mon – Fri 9:00A.M. – 5:00P.M.
JOIN US