FCAI SIG

Federated Learning

Coordination: Prof. Alex Jung (Aalto University)

Federated learning (FL) is a collection of machine-learning methods that train local personalized models collaboratively without sharing the data. Such techniques facilitate the preservation of data privacy, reduction of communication and computational cost for edge devices, as well as leverage parallel optimization algorithms. These properties make federated learning particularly relevant for healthcare, telecommunications, and edge AI.

This SIG will facilitate collaboration between research groups, developing practical algorithms and theoretical foundations for FL. We will provide a platform for discussing the impact of FL as the enabling technology for future AI applications.

To this end, we will organize a bimonthly seminar at Aalto University to discuss the latest developments in FL and launch a mailing list to notify people about upcoming events. We also expect to organize further events in collaboration with other FCAI SIGs, especially Edge AI and AI for Health, and arrange a thematic session during the AI Days.

Research Programs

Networked FL

 Many application domains generate collections of local datasets related by an intrinsic network structure ("big data over networks"). A timely application domain that generates such big data over networks is the management of pandemics.

Individuals generate local datasets of biophysical parameters via their smartphones and wearables. The statistical properties of local datasets are related via different network structures that reflect physical ("contact networks"), social, or biological proximity. FL methods exploit the network structure to pool local datasets into training sets for personalized models adaptively. This program studies conditions on the network structure, models, and computational infrastructure that ensure FL methods succeed.

Relevant publications:

A. Jung, “On the Duality Between Network Flows and Network Lasso,” in IEEE Signal Processing Letters, vol. 27, pp. 940-944, 2020, doi: 10.1109/LSP.2020.2998400.

A. Jung, “Networked Exponential Families for Big Data Over Networks,” in IEEE Access, vol. 8, pp. 202897-202909, 2020, doi: 10.1109/ACCESS.2020.3033817.

A. Jung, A. O. Hero, III, A. C. Mara, S. Jahromi, A. Heimowitz and Y. C. Eldar, “Semi-Supervised Learning in Network-Structured Data via Total Variation Minimization,” in IEEE Transactions on Signal Processing, vol. 67, no. 24, pp. 6256-6269, Dec., 2019, doi: 10.1109/TSP.2019.2953593.

A. Jung and N. Tran, “Localized Linear Regression in Networked Data,” in IEEE Signal Processing Letters, vol. 26, no. 7, pp. 1090-1094, July 2019, doi: 10.1109/LSP.2019.2918933.


Explainable FL

 A key challenge for the widespread use of machine learning methods is the explainability of their predictions. We have recently developed a novel approach to constructing personalized explanations for the predictions delivered by machine learning methods. We measure the effect of an explanation by reducing the conditional entropy of the prediction given the summary that a particular user associates with data points. The user summary characterizes the background knowledge of the "explainee" to compute tailored explanations for her. Our method only requires some training samples that consist of data points and their corresponding predictions and user summaries to compute the explanations. Thus, our method is model-agnostic and can be used to compute explanations for different machine-learning methods.

Relevant publications:

Zhang, L., Karakasidis, G., Odnoblyudova, A. et al. Explainable empirical risk minimization. Neural Comput & Applic 36, 3983–3996 (2024). https://doi.org/10.1007/s00521-023-09269-3

A. Jung and P. H. J. Nardelli, “An Information-Theoretic Approach to Personalized Explainable Machine Learning,” in IEEE Signal Processing Letters, vol. 27, pp. 825-829, 2020, doi: 10.1109/LSP.2020.2993176.


Resource efficient FL for cyber-physical systems

 FL methods require relatively large amounts of data exchanges, which impose challenges in using the communication network resources, particularly when wireless connectivity is considered. In previous research, we have worked with the benefits of different types of event-based sampling and communication to decrease the need for explicit communication over channels by using implicit communication when no actual transmission is detected. This method was extended to build a semantic-functional approach to resource-efficient communication networks that support cyber-physical systems (CPSs) operation. Our current research combines this new approach in a case where many elements of the CPS jointly perform FL over networks. By sampling and transmitting data that are meaningfully mapped into events that are semantically informative in terms of the FL role in the CPS, we decrease the communication (and energy) resources used while keeping the CPS functioning within its desired operational range. 

Relevant publications:

Silva, Pedro E. Gória, et al. "Enabling Semantic-Functional Communications for Multiuser Event Transmissions via Wireless Power Transfer." Sensors 23.5 (2023): 2707.

Silva, Pedro E. Gória, et al. "A novel semantic-functional approach for multiuser event-trigger communication." arXiv preprint arXiv:2204.03223 (2022).

de Castro Tomé, Mauricio, et al. "Event-driven data acquisition for electricity metering: a tutorial." IEEE Sensors Journal (2022).

Narayanan, Arun, et al. "Key advances in pervasive edge computing for industrial internet of things in 5G and beyond." IEEE Access 8 (2020): 206734-206754.

Nardelli, Pedro HJ. Cyber-physical Systems: Theory, Methodology, and Applications. John Wiley & Sons, 2022.

People

The following researchers already take part in the SIG. If you would like to join the SIG, please contact the coordinator.

  • Alexander Jung (Aalto University) - coordinator

  • Olga Kuznetsova (Aalto University)

  • Pedro Nardelli (LUT University)

  • Fabricio Oliveira (Aalto University)

  • Ella Peltonen (University of Oulu)

  • Simo Särkkä (Aalto University)

  • Jussi Kangasharju (University of Helsinki)

  • Kaie Kubjas (Aalto University)

  • Antti Honkela (University of Helsinki)

  • Kristiina Tähtinen (Tampere University)

  • Marlon Tobaben (University of Helsinki)

  • Udayanto Atmojo (Aalto University)