Quantized Compressive Learning: from broad intuitions to precise guarantees
| Quand ? |
Le 15/11/2021, de 16:00 à 17:00 |
|---|---|
| Participants |
Vincent Schellekens |
| Ajouter un événement au calendrier |
vCal
iCal
|
Title : Quantized Compressive Learning: from broad intuitions to precise guarantees
Asbtract : The compressive learning (or sketched learning) framework has been proposed to reduce the computational burden of machine learning from massive data. In this framework, one does not learn from the dataset directly, but first compresses it as a "sketch", an empirical average of transformations of the learning examples through a feature map (typically random Fourier features, i.e., random projections followed by complex exponentiation).
After a rapid overview of this framework, I will present our work on quantized compressive learning (QCL), where the feature map used during sketching produces binary outputs, which should allow for further computational savings. This presentation will be structured into two levels of abstraction. In the first broad level, I will answer the core questions: what is our motivation is for QCL, what are the concrete challenges it poses, how we tackled them, what is the intuition behind our strategy, what are the main empirical results. In the second level, as time allows, we will take a closer look at QCL (and its generalizations) as we gradually build our way up towards sound theoretical guarantees.
More information : Vincent Schellekens (https://schellekensv.github.io/) le lundi 15 Novembre de 16h à 17h . Vincent est un nouveau post-doctorant dans l'équipe DANTE et a travaillé notamment sur les méthodes d'apprentissage compressif.