Personalized collaboration for learning
Event details
| Date | 26.08.2021 |
| Hour | 08:00 › 10:00 |
| Speaker | El Mahdi Chayti |
| Category | Conferences - Seminars |
EDIC candidacy exam
exam president: Prof. Rachid Guerraoui
thesis advisor: Prof. Martin Jaggi
co-examiner: Prof. Nicolas Flammarion
Abstract
In many scenarios we have agents that collaborate to learn useful representations, such scenarios are for example Federated learning or decentralized learning in which each agent is a node and the goal is to train a global model which is, in general, a weighted average of all the available agents. It is a very well-known fact that the global model does not generalize well, in many interesting cases, at the local level, the main reason for this is the dissimilarity between the agents (agents behave differently, biased towards their own preferences). This fact justified the need for personalization i.e collaborative approaches that generate models that behave well at the local level. In particular, it has been noted the similarity between this framework and the meta-learning / few-shot learning paradigm. Many ideas were suggested in all of these learning domains, but while there is no shortage of ideas, there is clearly a need for theoretical guarantees. In this work, we will investigate approaches to solve the personalization problem in particular and answer the more general question: how agents that do not necessarily behave in the same way can have a beneficial collaboration with each other?
Background papers
On the Theory of Transfer Learning: The Importance of Task Diversity
https://arxiv.org/pdf/2006.11650.pdf
Which Tasks Should Be Learned Together in Multi-task Learning?
https://arxiv.org/pdf/1905.07553.pdf
Survey of Personalization Techniques for Federated Learning
https://arxiv.org/pdf/2003.08673.pdf
exam president: Prof. Rachid Guerraoui
thesis advisor: Prof. Martin Jaggi
co-examiner: Prof. Nicolas Flammarion
Abstract
In many scenarios we have agents that collaborate to learn useful representations, such scenarios are for example Federated learning or decentralized learning in which each agent is a node and the goal is to train a global model which is, in general, a weighted average of all the available agents. It is a very well-known fact that the global model does not generalize well, in many interesting cases, at the local level, the main reason for this is the dissimilarity between the agents (agents behave differently, biased towards their own preferences). This fact justified the need for personalization i.e collaborative approaches that generate models that behave well at the local level. In particular, it has been noted the similarity between this framework and the meta-learning / few-shot learning paradigm. Many ideas were suggested in all of these learning domains, but while there is no shortage of ideas, there is clearly a need for theoretical guarantees. In this work, we will investigate approaches to solve the personalization problem in particular and answer the more general question: how agents that do not necessarily behave in the same way can have a beneficial collaboration with each other?
Background papers
On the Theory of Transfer Learning: The Importance of Task Diversity
https://arxiv.org/pdf/2006.11650.pdf
Which Tasks Should Be Learned Together in Multi-task Learning?
https://arxiv.org/pdf/1905.07553.pdf
Survey of Personalization Techniques for Federated Learning
https://arxiv.org/pdf/2003.08673.pdf
Practical information
- General public
- Free
Contact
- edic@epfl.ch