NEWS
 BRIDGE GRANT FROM yrCSS. Riccardo Muolo, PhD student @naXys, and Joey O’Brien, PhD student @MACSI, University of Limerick (Ireland), have been selected for the Bridge Grant, funded by the Complex Systems Society (yrCSS) and designed to encourage scientific collaboration between young scientists. The grant will allow Joey to visit naXys in the next summer and to work with Riccardo on synchronization phenomena in real-world networks. The project is supervised by Prof. Timoteo Carletti (naXys), Prof. James Gleeson (MACSI) and Dr. Malbor Asllani (University College Dublin).
 CONGRATULATIONS to our naXys member Nicolas Franco for his nomination as “Namurois de l’année”. Every year, Namur city honors 12 personalities who played a major role in the society. Nicolas has been elected in the category “Science” for his modelling work of the expansion of COVID-19 in Belgium. Based on Sciensano statistical data, Nicolas and his collaborators elaborate scenarii for the evolution of the pandemic. The models project the situations we might have to undergo in Belgium according to the decisions made by the Authorities. As specified by the Jury, « Outil crucial dans la gestion de la crise du covid-19, ces modèles aident aussi à faire comprendre l’utilité des énormes efforts nécessaires demandés à tous pour combattre ce virus, ralentir sa propagation et ainsi sauver des vies ». More information
 PUBLICATION IN NATURE COMMUNICATIONS. A recent work on machine learning in spectral domain involving naXys researcher Timoteo Carletti has been published in the prestigious journal Nature Communications. More information
“Artificial intelligence is playing an increasingly role with applications interesting different fields of research, from the study of language to particle physics, from the prediction of the dynamics of financial markets to the classification of images. Neural networks in particular perform these tasks by training artificial brains made up of simple neurons, the computation units, connected together by links provided with appropriate weights, the synapses. The intelligence of a network lies precisely in this enormous collection of optimised weights which define its overall structure.
Even to manage relatively simple tasks, it is necessary to use networks formed by a considerable set of neurons and the number of weights to be assigned with the learning process therefore grows very rapidly. This is ultimately an intrinsic limitation which must be duly taken into account. In this work, we have developed an innovative approach to train neural networks: we have reformulated the problem in an abstract space where we can access global quantities that allow us to train simultaneously and all at once the weights of the artificial brain. In this way, the number of parameters to be learned by the training procedure is significantly reduced, while the performance of the network in carrying out the tasks assigned to it remains comparable to that obtained with conventional methods. To test the effectiveness of the method, we used two image databases. The first is a database of handwritten numbers (NMIST) that the neural network must be able to recognise. The second (Zalando) is a set of clothing images that must be properly classified. In both cases, the method developed produced results in line with those obtained with conventional approaches which are more expensive in terms of computer resources.”
Reference: Giambagli, L., Buffoni, L., Carletti, T. et al. Machine learning in spectral domain. Nat Commun 12, 1330 (2021).
|
