Computational cost reduction in adaptive diffusion networks and in the distributed training of neural networks: theory and applications

• Grant number: 25/05859-7
• Support Opportunities: Regular Research Grants
• Start date: September 01, 2025
• End date: August 31, 2027
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Daniel Gilio Tiglea
• Grantee: Daniel Gilio Tiglea
• Host Institution: Escola Politécnica (EP). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• Associated researchers: Luis Antonio Azpicueta Ruiz ; Magno Teófilo Madeira da Silva ; Renato Candido

Abstract

Distributed signal processing techniques have attracted the attention of the scientific community due to their advantages in comparison with centralized approaches. In this area, sampling and censoring techniques have been topics of particular interest, since the computational and/or energy costs associated with measuring, processing and/or transmitting data throughout the network can be prohibitive in certain cases. This project aims to conduct an in-depth study of sampling and censoring techniques in diffusion networks, proposing improvements to algorithms that have recently emerged in the literature and seeking to obtain theoretical results about the effect of these techniques on network performance. Furthermore, we intend to investigate the behavior of solutions for distributed signal processing in practical applications, such as acoustic impulse response estimation and the detection of artifacts in electroencephalogram signals. Finally, we intend to study the distributed training of neural networks. This approach has emerged in the literature as a possible solution for machine learning in situations in which it is desirable to mitigate the risks associated with data privacy. This would make it possible to obtain a model from all of the available data, without having to concentrate sensitive data on a single machine -- as is often the case in medical applications. In this case, the focus of the research will also be on reducing the computational cost associated with the learning task, without causing significant undesirable effects on the performance of the model obtained. (AU)