September 24, 2020 16:17
At this year's ECML-PKDD, the publication "Resource-Constrained On-Device Learning By Dynamic Averaging" received a Best Paper Award at the "Workshop on Parallel, Distributed and Federated Learning". The cooperation between the TU Dortmund University, ML2R, CRC 876, the University of Bonn, Fraunhofer IAIS, and Monash University was initiated by Katharina Morik's research stay in Melbourne.
The paper demonstrated that distributed learning of probabilistic graphical models can be realized completely with integer arithmetic. This results in reduced bandwidth requirements and energy consumption, thus enabling the use of distributed models on resource-constrained hardware. Furthermore, the possible error of the approximation was theoretically analyzed and bounded.
