ISP Seminars

Previous Seminars

Fitting on manifolds with Bézier functions
Pierre-Yves Gousenbourger
16 March 2017
Fitting and interpolation are well known topics on the Euclidean space. However, when it turns out that the data points are manifold valued (understand: when the data point belongs to a certain manifold), most of the algorithms are no more applicable because even the notion of distance is not as simple as on the Euclidean space.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 16 March 2017 at 16:15 (45 min.)

Deep learning and structured output problems
Soufiane Belharbi (INSA, Rouen, France)
8 December 2016
Deep neural networks have shown to be efficient models for learning complex mapping functions. It is known that adding more layers improves the performance but it comes with the cost of adding more parameters which requires more training labeled data. Pre-training techniques have shown to be helpful in training deeper networks by exploiting unlabeled data. However, in practice, this technique requires a lot of effort to setup the raised hyper-parameters. We present in this talk a regularization scheme to alleviate this problem. We extend our approach for structured output problems.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 8 December 2016 at 16:15 (45 min.)

An Introduction to Deep Learning
Simon Carbonnelle
1 December 2016
In the last 5 years, Deep Learning has revolutionized many fields such as computer vision, speech processing and language modelling. Some works have also shown that Deep Learning methods can be combined with external memories and reason over complex data structures such as graphs. Based on these successes, Deep Learning is considered by many as the most promising approach for Artificial Intelligence. The goal of this talk is to give you a broad, high-level overview of Deep Learning. More specifically, the following three questions will be addressed: What is Deep Learning and how does it work? What is Deep Learning currently capable of? And finally why does it work so well?
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 1 December 2016 at 16:15 (45 min.)

A 1-bit quantized compressive scheme for RADAR: From the hardware implementation to the signal model
Thomas Feuillen
20 October 2016
This seminar studies a 1-bit Quantized Compressive Sensing (QCS) scheme of a radar signal receiver that could enable novel systems with reduced complexities or cheaper design compared to high resolution strategies. In particular, the range of a sparse set of targets is estimated with a dense or randomized Stepped Frequency Modulation (SFM) with an acquisition process limited to only two 1-bit Analog to Digital Converters (ADCs) per antenna, one for each "I" and "Q" channels. This seminar will start by reviewing the basic principles of radar and the challenges related to the implementation. The signal model and its link to compressive sensing are then introduced. Second, we develop a complex variant of the Binary Iterative Hard Thresholding algorithm, or CBIHT, for the estimation of a sparse set of targets from complex QCS radar observations. Next, we show that the proposed 1-bit QCS framework reaches estimation errors similar to those obtained by both the common Maximum Likelihood Estimation (MLE) approach and the Iterative Hard Thresholding (IHT) algorithm when applied on full resolution radar observations. Finally, the feasibility of the QCS approach is studied over two different experimental setups in real conditions, with various sizes and quality.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 20 October 2016 at 16:15 (45 min.)

Restricted Range Space Property Based Theory for 1-Bit Compressive Sensing
Chunlei Xu
28 September 2016
Plenty works have been devoted to the study of compressive sensing over the past decades. Such a promising development of compressive sensing has a great impact on many aspects of signal and image processing. One of the key mathematical problems addressed in compressive sensing is how to reconstruct a sparse signal from linear/nonlinear measurements via a decoding algorithm. In the classic compressive sensing, it is well known that to exactly reconstruct the sparsest signal from a limited number of linear measurements is possible when the sensing matrix admits certain properties.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Wednesday, 28 September 2016 at 11:00 (45 min.)

Are classifiers really robust to deformations in the data?
Alhussein Fawzi, LTS4, EPFL, Switzerland (invited talk)
6 September 2016
The robustness of classifiers to small perturbations of the datapoints is a highly desirable property when the classifier is deployed in real and possibly hostile environments. Despite achieving excellent performance on recent visual benchmarks, state-of-the-art classifiers are surprisingly unstable to small perturbations of the data.

TVNUM Seminar Room (not Shannon!!), Place du Levant 2, Stévin Building, 1st floor -- Tuesday, 6 September 2016 at 11:00 (45 min.)

Inner Ear modelling with MRI
Dr Jérôme Plumat (University of Auckland, New Zealand) (invited talk)
30 August 2016
The inner ear (IE) is one of the most controlled organ, it is mostly fluid dominated and the barrier between blood and inner compartment is very tight. In this talk we will present recent results to quantify the permeability of the blood labyrinth barrier (BLB) and fluid mechanism parameters using the Dynamic Contrast Enhancement MRI (DCE-MRI). This safe imaging technic helps us to understand the IE, the transferts between blood and inner compartments as well as the propagation inside the cochlea. Also, we are able to estimate the BLB permeability and comparing control with patients with Ménière disease in order to understand more about this IE disorder. DCE-MRI models and methods will be discussed in this particular context.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Tuesday, 30 August 2016 at 11:00 (45 min.)

Cytomine for collaborative and semantic analysis of multi-gigapixel images
Raphaël Marée, Montefiore Institute, University of Liège (invited talk)
9 June 2016
Cytomine (http://www.cytomine.be/) is an open-source, rich internet application, for remote visualization of high-resolution images (à la Google Maps), collaborative and semantic annotation of regions of interest using user-defined ontologies, and semi-automated image analysis using machine learning.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 9 June 2016 at 14:00 (45 min.)

Low Rank and Group-Average Sparsity Driven Convex Optimization for Direct Exoplanets Imaging
Benoît Pairet
26 May 2016
Direct exoplanets imaging is a challenging task for two main reasons. First, the host star is several order of magnitude brighter than exoplanets. Second, the great distance between us and the star system makes the exoplanets-star angular distance very small. Furthermore, imperfections on the optics along with atmospheric disturbances create "speckles" that look like planets. Ten years ago, astronomers introduced Angular Differential Imaging (ADI) that uses the rotation of the Earth to add diversity to the data. Signal processing then allows to separate the starlight from the planet signal. After that, a detection signal to noise ratio (SNR) is computed. We will present a sparsity driven convex optimization program along with a new SNR map that beat the state of the art methods for small angular separation.

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 26 May 2016 at 14:00 (45 min.)

Dimensionality Reduction with t-SNE for data visualization
Simon Carbonnelle
12 May 2016
The t-SNE method addresses the problem of reducing the dimensionality of data to 2 or 3 dimensions while keeping as much local and global structure as possible. The main purpose of this dimensionality reduction technique is enabling human interpretation of datasets via 2D or 3D plots. t-SNE is based on the Stochastic Neighbor Embedding (SNE) method. The main idea of those methods is to model the pairwise similarity of data points in the initial data space and in the low dimensional target space through conditional or joint probability distributions. The coordinates of the data points in the new low-dimensional space are then calculated such that the Kullback-Leibler divergence between the similarity models in high and low resolution is minimized. This forces the new representation to have approximately the same pair-wise similarities as the high-dimensional representation. t-SNE has gained a lot of popularity to understand representations, usually learned by deep learning methods. Results for word and document embeddings (NLP) as well as for image classification will be presented.
Ressources :

Shannon Seminar Room, Place du Levant 3, Maxwell Building, 1st floor -- Thursday, 12 May 2016 at 14:15 (45 min.)