Federated Learning for Online Collaborative Knowledge and Decision-making (FLOCKD) has been accepted for funding by the Danish Independent Research Fund (DFF).

The FLOCKD project will investigate the distribution of Deep Neural Networks (DNNs) in smart camera networks, allowing individual cameras in a networked setting to classify and prediction trajectories and actions of observed objects. While DNNs are indeed very successful in identification and prediction tasks, they are resource expensive to train and maintain. To overcome this, federated learning has been proposed, combining the learned models of different devices. However, due to the different perceptions of cameras, a single common DNN might not be viable and individual, specialised DNNs are required. While utilising such individual specialised networks, we will also develop approaches allowing cameras to request feedback from each other by sharing their specialised networks upon request. We hypothesise this will lead to better network-wide inference.

The FLOCKD project I will work with Alexandros Iosifidis (Aarhus University) and we will collaborate with Prof Mohan Kankanhalli (National University of Singapore), Prof Bradley McDanel (Franklin and Marshall College), and Prof Andrea Prati (University of Parma).