📢 Announcement: LRE Students Seminar 🎓

Attention all EPITA students and faculty members! We are thrilled to invite you to the upcoming EPITA Research Laboratory (LRE) Seminar, where the LRE Research Students will expose their work.

📅 Jan 24, 10th ⏰ 10.00 - 12.00 and 16.00 - 18.00 🏠 KB 402

Schedule

• ❶ 10.00 - 12.00 AA,Image,Secu 📹 Lien Teams
• ❷ 16.00 - 18.00 IA 📹 Lien Teams

Abstracts

10:00 ↦ 10:25: Combining reactive synthesis and motion planning to control complex systems, Rostan Tabet

?????

10:30 ↦ 10:55: SAT-Based learning for CTL formulas, Simon Scatton

?????

11:00 ↦ 11:25: Implémenter des capacités de perception basées vision de l’environnement proche en frugalité embarquée, Célien Aubry

Percevoir son environnement est essentiel pour un robot mobile car il doit se déplacer dans un monde qui lui est inconnu ou changeant. Il s’agit dans ce projet de recherche de se focaliser sur les solutions à forte frugalité embarquée et implémenter les plus prometteuses tout en restant compatible avec Ros2.

Slides

11:30 ↦ 11:55: ????, Younes Benreguieg

?????

16:00 ↦ 16:25: Explicabilité des réseaux de neurones sur graphe au niveau des embeddings, Elouan Vincent

While Graph Neural Networks (GNNs) excel in learning node representations and exhibit remarkable performance on graph-related tasks, they grapple with the challenge of opaqueness, commonly referred as the “black-box” problem. Motivated by the need for transparency and interpretability in these models, this research is dedicated to enhancing our understanding of their inner workings. In previous work, we have formulated a method that integrates Hamiache-Navarro techniques and pattern mining, strategically tailored to construct representative graph of an activation rule within GNNs. In this extensive study, we thoroughly evaluate this approach, including a comprehensive exploration of its application on various GNN tasks, such as graph and node classification, and an in-depth analysis of the impact of different hyperparameters.

Slides

16:30 ↦ 17:55: Fine tuning of Automatic Speech Recognition Models for Self Supervised Speaker Verification, Victor Miara

Automatic Speech Recognition (ASR) models, characterized by their expansive size and training by corporations with substantial computational resources, inherently grasp extensive auditory information during training. This comprehension spans two major levels: speaker’s information and phonetic information.Though ASR models primarily utilize phonetic information, this study highlights their potential in speaker verification tasks when fine-tuned.The current state of the art in speaker verification on the SUPERB benchmark is Microsoft’s Wav-LM Large model. We are therefore seeking to utilize a pre-trained ASR model like Wav-LM for speaker verification tasks using a self-supervised approach. Techniques such as employing contrastive loss or utilizing pseudo labels generated by another self-supervised speaker verification model are considered, and could offer a robust, enhanced, and self-reliant system for speaker verification.

Slides

17:00 ↦ 17:25: Détection d’usurpation d’identité dans l’audio, Noe Audemard

Speaker verifications systems are vulnerable to spoofing attacks in various forms. Namely, voice conversion, voice synthesis and replay attacks. Many countermeasures to such attacks have been developed but remain vulnerable to unseen attacks or conditions different from the training data. In this work, we will try to propose a method to detect spoofing, and classify the attack according to the method used to generate the audio file.

Slides