Network Classifier
Event details
| Date | 27.08.2019 |
| Hour | 14:00 › 16:00 |
| Speaker | Elsa Rizk |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Rüdiger Urbanke
Thesis advisor: Prof. Ali Sayed
Co-examiner: Prof. Volkan Cevher
Abstract
The surge of data defined over a network calls for tools to be adapted to handle them. We have already explored designing machine learning algorithms for data defined in Euclidean space, but now we wish to extend it to data defined over graphs. Some works have already explored this extension, such as graph neural networks and representation learning. However, we focus on designing a network of classifiers that collaborate towards either a common goal, or towards multiple related goals. In this research proposal, we define the problem to be tackled: network classifiers. We present and discuss three publications that will help steer the research plan. The first paper introduces federated learning which is closely linked to network classifiers; the latter encapsulate the problems solved by federated learning. Next, the second paper looks at multitask networks, and develops a distributed optimization algorithm that enforces smoothness on the parameter vectors. It functions as a building block for formulating an optimization problem of the network classifiers. Finally, the third paper aims at privatizing distributed online learning algorithms while relying on the concept of differential privacy. This paper shall also function as a building block when the privacy level is important to the network classifiers. Thus, in our work, we hope to formulate the optimization problem of the network classifiers with all its variations, solve it in a private way, and study the trade-off between privacy and utility.
Background papers
Communication-Efficient Learning of Deep Networks from Decentralized Data, by Brendan McMahan, H., et al.
Distributed Inference over Multitask Graphs under Smoothness, by Nassif, R., et al.
Differentially Private Distributed Online Learning, by Li, C., et al.
Exam president: Prof. Rüdiger Urbanke
Thesis advisor: Prof. Ali Sayed
Co-examiner: Prof. Volkan Cevher
Abstract
The surge of data defined over a network calls for tools to be adapted to handle them. We have already explored designing machine learning algorithms for data defined in Euclidean space, but now we wish to extend it to data defined over graphs. Some works have already explored this extension, such as graph neural networks and representation learning. However, we focus on designing a network of classifiers that collaborate towards either a common goal, or towards multiple related goals. In this research proposal, we define the problem to be tackled: network classifiers. We present and discuss three publications that will help steer the research plan. The first paper introduces federated learning which is closely linked to network classifiers; the latter encapsulate the problems solved by federated learning. Next, the second paper looks at multitask networks, and develops a distributed optimization algorithm that enforces smoothness on the parameter vectors. It functions as a building block for formulating an optimization problem of the network classifiers. Finally, the third paper aims at privatizing distributed online learning algorithms while relying on the concept of differential privacy. This paper shall also function as a building block when the privacy level is important to the network classifiers. Thus, in our work, we hope to formulate the optimization problem of the network classifiers with all its variations, solve it in a private way, and study the trade-off between privacy and utility.
Background papers
Communication-Efficient Learning of Deep Networks from Decentralized Data, by Brendan McMahan, H., et al.
Distributed Inference over Multitask Graphs under Smoothness, by Nassif, R., et al.
Differentially Private Distributed Online Learning, by Li, C., et al.
Practical information
- General public
- Free