- Date: August, 26th
- Room: Escale en Arz, in Arz island.
- Title: Seminaire au vert! (Annual team seminar)
- Program
- Date: Wednesday, October 23rd – 10:00
- Room: A103, ENSIBS Vannes
- Speaker: Tom Avellaneda (IRISA Obelix) – Internship oral presentation
- Title: Efficient Remote Sensing Change Retrieval without Learning using Pattern Spectra
- Abstract: In the context of remote sensing image analysis and earth observation, this study focuses on the problem of spatio-temporal patch retrieval. Given a small pattern with spatial and temporal characteristics, we want to retrieve similar patterns from a very large satellite image time series. To solve this problem we leverage pattern spectra, attribute histograms which can serve as features and can be computed very efficiently using morphological hierarchies. The extension of pattern spectra to spatio-temporal data has not yet been fully explored, and their potential use for spatio-temporal pattern retrieval has not been tested. We propose a large variety of possible patch retrieval models using pattern spectra, which can be computed as a grid very efficiently, and test the performance of these approaches on the SECOND change detection dataset.
- Date: WEDNESDAY, November 13th – afternoon
- Room : in Rouen
- Speaker: Guillaume MAHEY (IRISA Obelix) — Ph.D. defense
- Title : Unbalanced and Linear Optimal Transport for Reliable Estimation of the Waserstein Distance
- Abstract: Despite being very elegant in theory, optimal transport (OT) might suffer from several drawbacks in practice. Notably, the computational burden, the risk of over fitting and its sensibility to artifacts motivated the introduction of variants for the OT loss in the ML community. In this thesis, we propose such variants in order to, first, reduce computational and statistical burden and second the sensibility to artifacts of sampling of the OT loss. To do that, we built on the top of proxies distributions introduced both by the Linear and the Unbalanced OT variants.spaces for representation learning by proposing a new classification method in hyperbolic spaces. Our method which is hierarchically-informed improves the performance in hierarchical classification.
as Guillaume’s defense will be held in Rouen, he will repeat his presentation in Vannes on Friday, 13th december at 11:00 am / Room A102 (ENSIBS building)
- Date: Thursday, November 28th – 11:30
- Room: A102, ENSIBS Vannes
- Speaker: Titouan Vayer (researcher Inria–OCKHAM, ENS Lyon)
- Title: Learning graphs with precision matrices: statistical estimators, compressive approches and unrolled neural networks
- Abstract: Learning a graph of relationships between variables from signals is a fundamental problem, particularly for analyzing geographic data or EEG brain data. Such a graph provides insights into the statistical dependencies between key variables. In this presentation, I will first review standard statistical approaches and convex algorithms or solving this problem, then introduce two distinct works: one focused on inverse problems, and the other on data-driven neural network approaches. The first work demonstrates how to estimate such a graph from a sketch of the data, which is a small summary built from random projections. I will show that, from an information-theoretic perspective, it is possible to recover the graph from a limited number of measurements, under sparsity assumptions. In the second part, I will present a novel neural network architecture, called SpodNet (Schur’s Positive-Definite Network), which tackles the graph estimation problem in a data-driven manner.
- Date: FRIDAY, December 13th – afternoon
- Room : Amphi ENSIBS, Vannes
- Speaker: Paul BERG (IRISA Obelix) — Ph.D. defense
- Title : Contributions to Representation Learning in Computer Vision and Remote Sensing
- Abstract : Deep Learning has become an ubiquitous tool for the resolution of image analysis tasks, notably in the remote sensing domain. As such, the need for annotated data have largely increased. But, annotating data can be costly and time consuming. Therefore, a whole field of the literature is dedicated to image representation learning by decreasing the dependence to annotations using so called self-supervised methods. The learnt representations are then usable in downstream tasks because of their discriminative nature with respect to the labels. In this context, we evaluate in this thesis how these methods can be exploited in the remote sensing domain by investigating tasks such as multi-modal scene classification for which we propose a self-supervised framework. We leverage the optimal transport problem to model several problems et propose new methodological contributions to contrastive learning. Finally, we propose to go beyond Euclidean spaces for representation learning by proposing a new classification method in hyperbolic spaces. Our method which is hierarchically-informed improves the performance in hierarchical classification.
- Date: Thursday, February 27, 2025 at 11:30 a.m.
- Room: A102, ENSIBS Vannes
- Speaker: Antoine Bralet
- Title: Deep learning for multimodal detection of sudden and slow moving slope instabilities on bitemporal remote sensing images
- Abstract: The aim of this work is to develop artificial intelligence algorithms for the detection of sudden and slow slope instabilities from synthetic aperture radar (SAR) and optical satellite images. The first contributions led to the creation of a new neural network architecture, SARDINet, enabling SAR images to be translated into optical images, so that comparable images can be obtained independently of weather conditions. Several architectural modifications improved the relevance and contrast of the translated images. These translations were then conditioned using new cost functions to guide the translator in delivering images that can be used a posteriori for classification or change detection tasks. These translations are then applied to the detection of sudden landslides in Haiti in 2021. Finally, the detection of slow instabilities was the subject of the creation of a new dataset, ISSLIDE. The dataset is based on a new manual routine of annotation on SAR interferograms and exploited by standard segmentation neural networks. It is also used to develop a new segmentation strategy named ECSPLAIN. The latter relies on the explainability of a classifier network during the training phase to generate segmentation maps and force the classifier to be explicable. Thus, this work provides new artificial intelligence methods applied to satellite images, while meeting the need for understanding in the geosciences.
- Date: Thursday, March 6, 2025 at 11:30 a.m.
- Room: A102, ENSIBS Vannes
- Speaker: Ambroise Odonnat
- Title: MaNo: Exploiting Matrix Norm for Unsupervised Accuracy Estimation under Distribution Shifts
- Abstract: Leveraging the model’s outputs, specifically the logits, is a common approach to estimating the test accuracy of a pre-trained neural network on out-of-distribution (OOD) samples without requiring access to the corresponding ground-truth labels. Despite their ease of implementation and computational efficiency, current logit-based methods are vulnerable to overconfidence issues, leading to prediction bias, especially under the natural shift. In this work, we first study the relationship between logits and generalization performance from the view of low-density separation assumption. Our findings motivate our proposed method MANO that (1) applies a data-dependent normalization on the logits to reduce prediction bias, and (2) takes the Lp norm of the matrix of normalized logits as the estimation score. Our theoretical analysis highlights the connection between the provided score and the model’s uncertainty. We conduct an extensive empirical study on common unsupervised accuracy estimation benchmarks and demonstrate that MANO achieves state-of-the-art performance across various architectures in the presence of synthetic, natural, or subpopulation shifts.
