Main goals

• Neural model reduction. The aim is to sparsify a deep neural network by cutting of connections of neurons among the network that don't contribute to the neural network training, during the training phase. In particular, we employ specific rule-oriented concepts, developed in the realm of network science (We base our motivation on observations in real neural networks whose actual topology is scale-free (or smallworld)),in order to sparsify the neural network with the aim of keeping the most significant linkages, which are responsible for better information distribution in the network. Thus, we reduce, drastically both the number of trainable variables and the size of the model, which leads to training acceleration and therefore our algorithms are applicable to resource-constrained (TinyML) devices.


• Federated learning over ad hoc networks. The aim is to drastically reduce communication in purely distributed ad hoc network, by allowing only a percentage of nodes (not every node), belonging to the network, to cooperatively train a global model, in Federated Learning techniques. So, we focus on finding smart node participation protocols (like following data similarity criteria) in order to exclude specific nodes of the training procedure and therefore reduce the data, needed to be transmitted among the nodes in the network.


• Multi-layer distributed backbones to facilitate federated learning. In the existing literature, most published studies have concentrated on single-layer networks; however, these works often fall short in representing large-scale real-world networks. This project addresses the challenge of constructing compact and robust multi-layer backbones, aiming to facilitate efficient information dissemination across network nodes. The proposed approach not only supports the federation process but also enhances the scalability of the developed algorithms.

Contact

>> Fragkou Evangelia, email: efragkou@uth.gr