Speech has always been a dominant mode of social connection and communication. However, speech processing and modeling have been challenging due to its variability. Classic speech technologies rely on cascade modeling, i.e. transcribing speech to text with an Automatic Speech Recognition (ASR) system, processing transcribed text using Natural Langu...
International audience
Dans cette thèse, nous examinons le problème de la surcharge des canaux de communication dans le contexte des systèmes de transport intelligents coopératifs (C-ITS). Nous visons à améliorer le mécanisme de communication entre les véhicules et nous nous concentrons sur la partie sécurité de la communication, qui est la plus coûteuse en termes de res...
This doctoral thesis is dedicated to three interrelated Non-Intrusive Load Monitoring (NILM) topics.In the central part, the Unsupervised Multi-Sequence Appliance Identification (UMSAI), an algorithm for detecting electronic devices, is developed. This algorithm identifies simple and complex devices from the total power signal. The approach employs...
The goal of this thesis is to develop machine learning approaches to analyze collections of images without annotations. Specifically, given a collection of images as input, the machine should discover analysis concepts related to properties of the world, such as the object class or the underlying 3D geometry, without being taught this concept via m...
The goal of this thesis is to develop machine learning approaches to analyze collections of images without annotations. Advances in this area hold particular promises for high-impact 3D-related applications (e.g., reconstructing a real-world scene with 3D actionable components for animation movies or video games) where annotating examples to teach ...
The contributions of this thesis are two-fold. First, we deal with deep learning approaches for fully unsupervised motion segmentation from an optical flow field. We leverage a loss function based on the EM algorithm and involving parametric motion models. We then gradually extend this framework to longer sequences of input flows. With a triplet of...
Image generation has made substantial progress in recent years to the point where it becomes difficult to distinguish fake images from real ones. This success is mainly due to the concept of adversarial training: two models train to satisfy conflicting objectives. One called the "discriminator" tries to distinguish the dataset images from images ge...
This thesis explores new deep-learning approaches for modeling and analyzing real-world 3D data. 3D data processing is helpful for numerous high-impact applications such as autonomous driving, territory management, industry facilities monitoring, forest inventory, and biomass measurement. However, annotating and analyzing 3D data can be demanding. ...
Machine learning answers to the problem of handling dedicated tasks with a wide variety of data. Such algorithms can be either simple or difficult to handle depending of the data. Low dimensional data (2-dimension or 3-dimension) with an intuitive representation (average of baguette price by years) are easier to interpret/explain for a human than d...
