Analysis of the consistency of audiovisual concepts in an unsupervised manner, specifically understanding whether there is a relationship between a sound and a sequence of images without human intervention.This thesis has been divided into 2 parts:The first part consisted in analyzing audio and video through spiking networks for its temporal properties with hebbian learning rules (STDP and its variants), then applying self-supervised learning methods from deep learning in the SNNs… Show more Analysis of the consistency of audiovisual concepts in an unsupervised manner, specifically understanding whether there is a relationship between a sound and a sequence of images without human intervention.This thesis has been divided into 2 parts:The first part consisted in analyzing audio and video through spiking networks for its temporal properties with hebbian learning rules (STDP and its variants), then applying self-supervised learning methods from deep learning in the SNNs domain.The second part focuses on the explicability of neural networks (XAI) to help the convergence of models on the first task (finding a good initialization, a good architecture, understanding what is important for a given task). Show less

Development of a vehicle recognition pipeline (by make, model and year) in CCTV streams.Study of the impact of dataset errors on model performance

Temporal analysis of audio files to find repeating sounds using Spiking Neural Networks (SNNs).

Development of a Spiking Neural Networks framework using unsupervised hebbian learning rules like STDP (Spike-Timing Dependent Plasticity).Trained five/six people on deep learning methods using PyTorch, and on the basics of SNNs.Supervision of an engineer on the reverse engineering of an old framework.

Development of an automatic speech recognition system with Kaldi adapted to children with autism within the framework of a national project : INSIDE (Intelligent Networked Robot Systems for Symbiotic Interaction with Children with impaired Development).

Analysis of repeated gestures from the Kinect sensor using similarity matrices.

Design and Implementation of a Physical Access Control System