Project Coordinator (EU) :UNIVERSITAT POLITECNICA DE CATALUNYA
Country of the EU Coordinator :Spain
Organisation Type :Academia
Project participants :
US: Texas Tech University (TTU)
Lu Wei is an assistant professor at the Department of Computer Science at Texas Tech University (TTU). He obtained a Ph.D. degree from Aalto University, Finland, in 2013. He held postdoctoral positions at University of Helsinki, Finland, from 2013 to 2015 and at Harvard University, USA, from 2015 to 2016.
US: Mississippi State University (MSU):
Chun-Hung Liu is an assistant professor at the Department of Electrical and Computer Engineering of Mississippi State University (MSU). He obtained a Ph.D.degree from University of Texas at Austin, USA, in 2011.
Felix Freitag is an associate Professor at the Department of Computer Architecture of UPC. He obtained a PhD degree from the UPC in 1998. He coordinated the FP7 CLOMMUNITY project (2013-2015). His recent research focuses on community clouds and federated learning. Felix supervised several PhD theses. The full list of his publications is available at https://futur.upc.edu/FelixUrsErichFreitag.
State of US partner :Mississippi
Starting date :
Adaptive decentralized federated learning in wireless mesh network (FLESHNET)
The FLESHNET project will apply an experimentally driven research in which the design of adaptive decentralized federated learning will obtain feedback from the prototype experimentation, experimentally evaluated in a testbed deployed within the GuifiSants wireless city mesh network.
This testbed consists of a set of low-capacity devices. Specifically, two main types of hardware are used as nodes of the testbed, namely Minix mini-PCs and PC Engines APU2. Using more than one type of device adds heterogeneity in hardware, which can be targeted by certain types of experiments. Both devices have Debian 10 and Docker support installed.
The federated learning components will be deployed on these distributed testbed nodes and the performance will be measured. The approach is to use Docker images which contain the specific modification of the federated learning design to be experimentally evaluated. As such, the Docker images will need to be deployed in the chosen testbed nodes.
Typically, in user environments these computing devices run as home servers to manage several user-oriented services. Therefore,these devices are not dedicated exclusively to a machine learning application. A real scenario for this situation is the Guifi.net community network, where some users provide applications on such low-capacity nodes to other users. From this scenario, an important research topic can be motivated: it is important to understand how federated learning consumes computing resources, since in nondedicated home devices, i.e., those running multiple services, the resource consumption of federated learning may need to be transparent to the user and not affect other applications running simultaneously in the device.
Using experimentation with real devices connected to this wireless mesh network, our research will target to understand the resource usage of federated learning in alternative designs and models. We aim to understand better how for the different designs the phases of federated learning of a learning round (i.e., the model exchange and local model training) affect the CPU and memory consumption of the low capacity device and the traffic in the network.
Implementation plan :
The FLESHNET project will apply an experimentally driven research in which the development of federated learning designs will obtain feedback from the prototype experimentation, experimentally evaluated in the our testbed. The FLESHNET project is organized in three work packages (WPs) with two focus areas: WP1 (Design) and WP2 (Integration) is dedicated to design and integrate federated learning components in a deployable prototype. WP3 (Experimentation) conducts the experimentation in the wireless mesh network testbed. FLESHNET will have two iterations with short design and experiment cycles of three months. Initial experimental results at M3 will feed the research of the second cycle and to obtain final outcomes at M6.
During M1, the activities in WP1 will prepare the initial building blocks. From M2 to M4 the components will be integrated in a deployable prototype by WP2 and experimentation will be carried out by WP3. The first cycle finishes in M4 where WP1 carries out the analysis and feeds the results to WP2, to develop in M5 the improved design, and for being experimented in M6 by WP3 closing the loop.
The consolidation of the findings of WP1 in the second cycle will provide the results for objective 1, namely experimentally-driven design and integration of adaptive decentralized federated learning components. The consolidation of the findings of WP3 in the second cycle will provide the results for objective 2, namely experimental validation of federated learning in wireless mesh networks.
The experiments validate the designs in a relevant environment (i.e., an operational wireless mesh network with real nodes). Such an environment can classify as TRL 5.
Expected Impacts :
Impact 1: Enhanced EU – US cooperation in Next Generation Internet, including policy cooperation.
FLESHNET establishes a new collaboration of three partners, namely UPC (EU), TTU (US) and MSU (US). The collaboration addresses a currently very relevant research topic. All three partners benefit from the complementary knowledge within the consortium, which is needed to undertake the experiment proposal. We hope that this team of the three partners will be a starting point for growth and opens up new collaborations.
Impact 2: Reinforced collaboration and increased synergies between the Next Generation Internet and the Tomorrow's Internet programmes.
FLESHNET partners will monitor the evolution of the EU's NGI initiative and the USA's Internet related programmes for opportunities to increase synergies.
Impact 3: Developing interoperable solutions and joint demonstrators, contributions to standards.
The deployed prototype used for experimental evaluation in the testbed can be presented in the form of a demonstrator. Resource consumption of machine learning is a growing concern and there are standardization efforts on-going, e.g., ITU.FG-AI4EE aims to establish indicators. Experimental results from the project looking at the resource consumption of different designs may help to make contributions in this direction.
Impact 4: An EU - US ecosystem of top researchers, hi-tech start-ups / SMEs and Internet-related communities collaborating on the evolution of the Internet.
Since our results are obtained from experiments with a real environment rather than from simulations, there is a possibility to also attract practitioners besides the research community.
Expected Results :
FLESHNET proposes to experiment adaptive decentralized federated learning in wireless mesh networks. The results of this experiment will advance the state-of-the-art in the following aspects.
- Understand the performance of federated learning in a real edge environment: Federated learning is still a very recent research area where almost all research works are evaluated in simulations. Differently, FLESHNET will provide results from field experiments. Specifically, we will obtain performance results from a real edge environment given by our testbed. We follow the approach of previous works but go beyond in terms of the testbed environment.
- Adaptation of federated learning: In FLESHNET we aim to assess a cross-layer design for federated learning which integrates context-aware knowledge such as that of the network in the decisions of federated learning workers, allowing them to adapt to the dynamics and changes of edge networks, leading to more resource efficient computing. Such results build upon ideas of recent work such as to better understand the strategies for making federated learning more efficient in real systems.
- New paradigm for federated learning: The current federated learning paradigm is conceived as a Master-Slave approach, where the control and “intelligence” is centralized in the Master node, and worker nodes (clients) are passive computing machines. Differently, in FLESHNET, we argue for a more adaptive and decentralized organization of federated learning, where the intelligence, decision and control is more distributed among the components. Most current works explore this direction by a theoretical approach. The results from the experimentation in FLESHNET will provide evidence to support this few federated learning model.