Seminare from Mano Brabant, PhD student at LIUM

 

Date: 6/01/2023
Time: 11h00
Localization: IC2, Boardroom
Speaker: Mano Brabant
 
 

Lightning exploration in interpretability of speech systems.

XAI (eXplainable Artificial Inteligence) is a research field that focuses on the understanding and explanation of Machine Learning systems. Recent Deep Learning systems are more and more complex and less and less interpretable, which can be a problem because this notion of interpretability is mandatory in certain fields where decisions taken have heavy consequences, such as in the legal or medical fields.

This notion is also very useful for other purposes such as bias reduction or to help researchers to improve their models more easily. Numerous advances have been made, especially in the field of text and images. We study these different approaches for applications in the speech domain. This seminar presents the XAI domain and then reviews the first exploratory work we have been able to do, which consists in different visualizations of the embeddings of a WavLM system. The future research works considered will then be presented.

 

“Unlimited power” or “Aucune limite” to Jean Zay’s power.

This synthesis is the follow-up of a training on Optimized Deep Learning for the Jean Zay national supercomputer. It contains a quick presentation of the supercomputer as well as different techniques to optimize learning time, memory consumption and GPU efficiency.

 

Coup de foudre for Pytorch Lightning.

This presentation provides an overview of the PyTorch Lightning tool, an overlay of PyTorch that enables faster development of Machine Learning systems and easier reproducibility of research work.