Seminaires

Prochains séminaires

  • Seminaire IA : lundi 13 mars 2023, 12h30 en hybride (amphi Pelvoux)

Title: Bayesian inference and the inverse problem

Abstract: The tools for Bayesian inference are becoming more and more efficient thanks to the increase of available data. These last years have seen the growth of methods based on neural networks, allowing to obtain very good results by training algorithms to perform specific tasks. Although impressive performances, these methods are often criticized due to a lack of interpretability as well as some difficulties in estimating the confidence attributed to a prediction. This presentation will reintroduce the basic concepts of Bayesian inference, and will discuss the choices related to the observation model, the interest of a priori models as well as the classical strategies. Examples based on recent research work will be used as support.

slides

  • Seminaire IA : vendredi 15 avril 2022, 14h en hybride (amphi BX30 Pelvoux)

Titre: Tutorial sur les mécanismes d’attention et les modèles transformers.

Résumé:Les mécanismes d’attention et les transformers représentent une importante et récente évolution de l’apprentissage profond qui a beaucoup d’impact dans de nombreux domaines d’application (NLP, images, signal, données temporelles, modèle génératif). Lors de ce tuto, nous présenterons les principes de base de l’attention et les principaux modèles neuronaux qui en sont dérivés.

slides

Séminaires passés

  • Seminaire IA : mercredi 1 décembre 2021, 13h30 en présentiel (IBGBI)
    • Geoffroy Peeters, Full-professor at Télécom-Paris, Institut Polytechnique de Paris, IDS (Image-Data-Signal) department, S2A (Signal-Statistique-Apprentissage) team, ADASP (Audio Data Analysis & Signal Processing) group

Title: Deep learning for music audio signal processing

Abstract:As in many fields, deep neural networks have allowed important advances in the processing of musical audio signals. We first present the specificities of these signals and some elements of audio signal processing (as used in the traditional machine-learning approach. We then show how deep neural networks (in particular convolutional neural networks) can be used to perform feature learning. We first recall the fundamental differences between 2D images and time/frequency representations. We then discuss the choice of input (spectrogram, CQT, or raw-waveform), the choice of convolutional filter shape, autoregressive neural models, and the different ways of injecting a priori knowledge (harmonicity, source/filter) into these networks. Finally, we illustrate the different learning paradigms used in the music audio domain: classification, encoder-decoder (source separation, constraints on latent space), metric learning (triplet loss), and semi-supervised learning.

  • Séminaire axe IA : vendredi 7 mai à 14h
    • Aymane Souani : Modèle d’apprentissage profond pour calibration de capteurs de polluants. Application à l’identification des polluants industriels atmosphériques.
    • Nadia Abchiche-Mimouni
  • vendredi 29 janvier 2021 à 14h en visio , prochaine réunion de l'axe IA en distanciel
    • Hedi Tabia, (PR, IRA2) The impact of kernel functions on convolutional neural networks

abstract: CNNs have been used for a multitude of visual tasks. They showed to perform very competitive results while linear operations are used at different layers of the network. Linear functions are efficient, particularly, when the original data is linearly separable, which should have, in general, a high dimensional representation. In such a case, the decision boundary is a linear combination of the original features. However, it is worth noting that not all high dimensional problems are linearly separable. This talk presents the impact of the usage of kernels as an alternative of linear functions in CNNs.

  • Yassine Kebbati,(doctorant, SIAM) Adaptive control using online learning and meta-heuristics
  • mercredi 7 octobre 2020, 11h en visio

Programme (1h) :

  • R. Souriau (doctorant IBISC, équipe SIAM), Titre: Continuous Restricted Boltzmann Machine for learning signal representation: Application to control on Deep-Brain Stimulation
  • Nataliya Sokolovska (MCF HDR Faculté de Médecine, Univ. Sorbonne), Titre: Learning scores (interpretable models) for clinical applications. ref
  • Séminaire de Jonathan KOBOLD, Jeudi 5 mars 2020 à 14h, Bx30 Pelvoux: «The Receptive Field in CNNs» Info
  • jeudi 30 janvier 2020, 11h IBGBI., Programme (1h) :

Learning from a single modality (image, text or sound) is now a widely addressed problem with many well performing techniques such as CNNs. In this talk we will focus on the use of more than one modality, and, in particular, we will mainly discuss the principles of multimodal deep learning. The goal is to provide some insight into these new new approaches.

  • : Victoria Bourgeais : Méthodes semi-supervisées pour l’apprentissage profond

à partir de jeux de données transcriptomiques de petite taille.

  • Ludovic Ishiomin : Retour sur l’utilisation de nos serveurs GPU
  • 15 Novembre 2019: Blaize Hanczar, Tutorial sur les GAN, Sarra Houidi, NILM using feature selection and neural networks details
  • 2 Octobre 2019, Réunion de lancement de l'axe transversal IA au laboratoire IBISC