Amaury Bouchra Pilet successfully defended his Ph.D. entitled “Contributions to distributed multi-task machine learning” on Wednesday, November 10, 2021, at IRISA/Inria in Rennes.
- Marc TOMMASI, Professor, Université de Lille
- Giovanni NEGLIA, Research Scientist, Inria, Sophia Antipolis
- Sara BOUCHENAK, Professor, INSA Lyon,
- Aurelien BELLET, Research Scientist, Inria, Lille
- Alexandre TERMIER, Professor, Université de Rennes 1
- Davide FREY, Research Scientist, Inria, Rennes
- François TAIANI, Professor, Université de Rennes 1
Title: Contributions to distributed multi-task machine learning
Machine learning is one of the most important and active fields in present computer science. Currently, most machine learning systems are still using a mainly centralized design. Even when the final application is to be delivered in several systems, potentially millions (and even billions) of personal devices, the learning process is still centralized in a large data center. This can be an issue if the training data is sensitive, like private conversations, browsing histories, or health-related data.
In this thesis, we tackle the problem of distributed machine learning in its multi-task
form: a situation where different users of a common machine learning system have similar but different tasks to learn, which corresponds to major modern applications of machine learning, such as handwriting recognition or speech recognition.
We start by proposing a design of an effective distributed multi-task machine learning system for neural networks. We then propose a method to automatically optimize the learning process based on which tasks are more similar than others. Finally, we study how our propositions fit the individual interests of users.