ISP Seminars

Previous Seminars

An outlook on deep reinforcement learning
(invited talk) Dr Vincent François-Lavet (McGill University, Canada & Mila)
6 September 2019
Deep reinforcement learning is the combination of reinforcement learning (RL) and deep learning. This field of research has been able to solve a wide range of complex decision-making tasks that were previously out of reach for a machine. In this talk, I will provide an introduction to deep reinforcement learning models, algorithms and techniques. I will particularly focus on the aspects related to generalization of the learned policy to slightly different situations than the ones encountered during training.
Ressources :

Euler building (room A.002) -- Friday, 6 September 2019 at 11:00 (45 min.)

Bit-Based Learning Machines
Sercan Aygün
4 September 2019
Stochastic computing (SC) is a re-emerging computation methodology uses the hardware-based approaches in the presence of bit-streams. In recent years, the stochastic computing is being used from communication systems to the neural networks including hardware design issues, device testing & modelling, image processing and finally the learning machines. SC is basically advantageous for the area and power consumption criterion on the hardware, however, especially for the applied areas such as computer vision and neural systems, they require investigation and deep analysis. In this talk, stochastic computing basics on the neural network systems will be given to extend it into multilayer neural networks including the binary network case. Preserving streams through all of the network layers preserves robustness to the noise. Basic operations such as multiply-and-accumulate are handled with simple logic elements. Hardware structure both in well-known conventional binary neural networks and in the stream-based neural networks is very similar. Thus, bit stream injected XNOR gates and counters are used to obtain noise-resistant architecture in a same way.

Nyquist Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 4 September 2019 at 11:00 (45 min.)

The lensless endoscope: a playground for acquisition schemes based on compressed sensing principles
Stéphanie Guérit
5 June 2019
Can you conceive of imaging brain activity at a cellular level using an extra thin probe built with an optical fiber? Researchers from Institut Fresnel in Marseille are currently working on designing such a device, called lensless endoscope (LE). During this talk, I will present a joint work initiated one year ago. The aim of this collaboration is to design new acquisition scheme inspired by compressed sensing theory. Traditionally, data are acquired using raster scanning method: each discrete localization of the field of view is illuminated sequentially with a focused light beam, and the fluorescence signal is then collected back in the fiber. This produces satisfactory images that does not need post-processing but require a time-consuming calibration step and a high number of measurements. The proposed acquisition framework is based on two key elements: (i) the ability to produce pseudo-random illumination patterns by removing the calibration step and (ii) the memory effect of the fiber, i.e., the ability to shift the illumination pattern in the field of view thanks to mirror galvanometer. It allows us to reach reconstruction quality similar to the traditional framework but using far less measurements. I will describe this framework, the considered sampling strategies and some numerical considerations (e.g., the use of cross-validation to set the main parameter of the reconstruction algorithm). Finally, I will show and comment some results on both synthetic and experimental data.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 5 June 2019 at 11:00 (45 min.)

Low-complexity signal processing for direct imaging of stellar systems
Benoît Pairet
8 May 2019
Direct imaging of stellar systems is an emerging and multi-disciplinary field of research. Stellar systems consist of faint objects lying in close vicinity of a very bright object, namely the host star. Hence, obtaining a faithful image of such a system not only requires state-of-the-art imaging hardware and dedicated observation strategy but also heavily relies on tailored post-processing techniques.

Nyquist Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 8 May 2019 at 11:00 (45 min.)

Compressive Learning meets privacy
Vincent Schellekens
3 April 2019
Compressive Learning (CL) is a framework where a target learning task (e.g., clustering or density fitting) is not performed on the whole dataset of signals, but on a heavily compressed representation of it (called sketch), enabling training with reduced time and memory resources. Because the sketch only keeps track of general tendencies (i.e., generalized moments) of the dataset while discarding individual data records, previous work argued that CL should protect the privacy of the users that contributed to the dataset, but without providing formal arguments to back up this claim. This work aims to formalize this observation.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 3 April 2019 at 11:00 (45 min.)

Quantity over Quality: Dithered Quantization for Compressive Radar Systems
Thomas Feuillen
20 March 2019
In this talk, we investigate a trade-off between the number of radar observations (or measurements) and their resolution in the context of radar range estimation. To this end, we introduce a novel estimation scheme that can deal with strongly quantized received signals, going as low as 1-bit per signal sample. We leverage for this a dithered quantized compressive sensing framework that can be applied to classic radar processing and hardware. This allows us to remove ambiguous scenarios prohibiting correct range estimation from (undithered) quantized base-band radar signal. Two range estimation algorithms are studied: Projected Back Projection (PBP) and Quantized Iterative Hard Thresholding (QIHT). The effectiveness of the reconstruction methods combined with the dithering strategy is shown through Monte Carlo simulations. Furthermore we show that: (i), in dithered quantization, the accuracy of target range estimation improves when the bit-rate (i.e., the total number of measured bits) increases, whereas the accuracy of other undithered schemes saturate in this case; and (ii), for fixed, low bit-rate scenarios, severely quantized dithered schemes exhibit better performances than their full resolution counterparts. These observations are confirmed using real measurements obtained in a controlled environment, demonstrating the feasibility of the method in real ranging applications.

Nyquist Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 20 March 2019 at 11:00 (45 min.)

Post-processing techniques for exoplanet imaging.
Faustine Cantalloube (MPIA, Heidelberg, Germany)
25 October 2018
Imaging exoplanets provides unique insights to our knowledge about planet formation and evolution. However this is an extremely challenging task: the angular separation between the star and the planet is of a few hundred milli-arcseconds and the brightness ratio is of a few billion. To tackle this, a very specific instrument is needed but still the images we obtain are corrupted by the so-called “speckle noise” within which planets are hidden. Adapted post-processing is therefore necessary to unveil the planets in the images. The current post-processing techniques used have limitations that we need to address to image more exoplanets but also the environment in which they evolve to, in the end, understand better our universe.

Nyquist Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 25 October 2018 at 14:15 (45 min.)

Approximate digital system design with hardware-based stochastic computing for computer vision
Sercan Ayguns, PhD student (Visitor from İstanbul Technical University, Maslak, İstanbul, Turkey)
13 September 2018
Stochastic Computing (SC) is an emerging research topic as a new calculation paradigm for numerical systems and for application-specific structures such as hardware depended implementation of image processing and learning science. The stochastic calculation, although based on the 1960s, has increased considerably in years for its importance and applicability. The approach on this computing paradigm is to approximate results using fewer circuitry treating either transfer function or the algorithm itself. The computing paradigm is treated in three stages from bit streams to a logic system itself and from the system to the output bit streams. The design methodology is also in the scope of nanoscale devices to measure the probabilistic behavior of the atomic structures. Moreover, test and verification of conventional logic systems are controlled for the fault and error analysis using the SC approach. SC is recently is applied to image processing algorithms to be realized in hardware with less power consumption and area. Image pre-processing such as noise reduction, or edge detection are some examples that are handled as application of SC like image compression, too. As a Ph.D. research, it is being researched how to apply this paradigm into some of several machine learning structures like deep learning for the treatment of images.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 13 September 2018 at 15:30 (45 min.)

High resolution microCT, and its past and future challenges in the biomedical field
Prof. Greet Kerckhofs (UCLouvain/iMMC, KULeuven)
12 June 2018
Biological tissues are 3D structures with complex spatial heterogeneity, for which traditional 2D imaging techniques such as standard histomorphometry are insufficient to comprehensively characterize them or assess their quality. Advanced 3D imaging is one of the enabling technologies that is of increasing importance in the biomedical field to assess tissue quality, and to provide better knowledge on the mechanisms behind tissue formation and regeneration. In this seminar, I will provide an overview of the process of implementing high resolution microCT as advanced 3D imaging tool for different biomedical applications, as well as of its optimization, both in terms of hardware (side equipment and contrast agents) and software (for image acquisition, reconstruction and analysis) specifically for different biomedical applications. I will highlight what we have achieved over the past 15 years within the KU Leuven microCT labs, but I will also discuss which the currently remaining challenges are and how we will aim to tackle (some of) them in the future.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Tuesday, 12 June 2018 at 15:00 (45 min.)

Fisheye stereovision to model 3D urban environments in the context of a GNSS positioning application
Dr Julien Moreau (IFSTTAR, France)
28 May 2018
This research deals with 3D modelling from an embedded fisheye vision system, used for a GNSS (Global Navigation Satellite Systems) application as part of CAPLOC project. Satellite signal propagation in urban area implies reflections on structures, impairing localisation’s accuracy and availability. The project purpose is to define an omnidirectional vision system able to provide information on urban 3D structure and to demonstrate that it allows to improve localisation. This presentation addresses problems of (1) self-calibration, (2) matching between images, (3) 3D reconstruction; each algorithm is assessed on computer-generated and real fisheye images. Moreover, it describes a way to correct GNSS signals reflections from a 3D point cloud to improve positioning. Calibration is handled by a two-steps process: the 9-point algorithm fitted to “equisolid” model coupled with a RANSAC, followed by a Levenberg-Marquardt optimisation refinement. Dense matching is done by dynamic programming along conjugated epipolar curves. Distortions are not rectified in order to neither degrade visual content nor to decrease accuracy. In the binocular case it is possible to estimate full-scale coordinates. In the monocular case, we do it by adding odometer information.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Monday, 28 May 2018 at 13:00 (45 min.)