Main topics
Statistical Machine Learning, Deep Learning, Neural Networks, Artificial Intelligence
Current projects
The project aims at creating a map of the brain areas according to its processing level of input stimuli by comparing cerebral representations with learned representations in a deep neural net.
Comparing computational and cerebral representations of voice and sounds
With Bruno Giordano
Funding: Projet ANR AAPG 2021 BrainSoundSem
A mathematical framework for morpho-transcriptomics
Funding: Malek Senoussi’s Ph.D. funding from Paul Villoutreix’s Centuri chaire
Explainability in Neural Networks for sequence data
With Stephane Ayache (LIS) and Hamed Benazha (LIS)
Funding : Projet ANR AAPG 2020 TAUDOS
Past projects
Sequence modeling, labeling and classification
Design of statistical models for sequence labeling and classification based on Hidden Markov Models, Conditional Random Fields, and variants like Contextual Hidden Markov Models, Hidden Conditional Random Fields, neural extensions of these models etc.
Extreme Classification
Extreme classification means classification in a very large number of classes (up top millions), either in monolabel or multilabel classification settings. This projet consisted in part in designing new methods for this challenging setting and in organizing international large scale classification challenges (LSHTC series and BioAsQ series).
Budgeted Learning
Budget learning covers various topics. This project aimed at designing classifiers that are economic in the number of features that are required to produce a decision on an input data.
Adversarial Learning For Animation
With Qi Wang.
Funding : Chinese doctoral program CSC-ECM
We designed synthesis systems for generating motion capture data that matchs some high level features ( e.g. the emotion with which an motion is performed) using adversarial learning and recurrent neural nets.
Disentangling factors of variations
Disentangling factors of variations aims at learning factors of variations from data. Once learned it allows changing the stye of an input data (e.g. viewpoint for image data) to the style of another image. We have applied such techniques, relying on adversarial learning on images and on motion capture sequences.
