course offered during spring 2023 at Aalto University and to adult learners via Finnish Network University
If you have not registered for this course via your university course enrolment system (such as sisu.aalto.fi) or fitech.io (Finnish adult learners), you can also register via this google form: Registration for Externals
Lectures *** Lecture Notes *** Assignments *** FL Project
Many application domains of machine learning (ML), such as numerical weather prediction, generate decentralized collections of local datasets. A naive application of basic ML methods [1] would require collecting these local datasets at a central point. However, this approach might be unfavourable for several reasons, including inefficient use of computational infrastructure or the lack of privacy.
Federated learning (FL) aims at training ML models in a decentralized and collaborative fashion. FL methods require only the exchange of model parameter updates instead of raw data. These methods are appealing computationally and from a privacy protection perspective. Indeed, FL methods leverage distributed computational resources and minimize the leakage of private information irrelevant to the learning task.
This course uses an optimization perspective to study some widely used FL models and algorithms. In particular, we will discuss total variation minimization as a unifying design principle for FL methods. The geometry of total variation determines these methods' computational and statistical properties.
[1] A. Jung, "Machine Learning. The Basics," Springer, Singapore, 2022. available via Aalto library here. preprint.