From adaptive signal processing to distributed learning: computational cost, privacy, and applications

• Grant number: 26/01471-7
• Support Opportunities: Regular Research Grants
• Start date: June 01, 2026
• End date: May 31, 2029
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Magno Teófilo Madeira da Silva
• Grantee: Magno Teófilo Madeira da Silva
• Host Institution: Escola Politécnica (EP). Universidade de São Paulo (USP). São Paulo , SP, Brazil
• City of the host institution: São Paulo
• Associated researchers: Daniel Gilio Tiglea ; Renato Candido

Abstract

In recent years, adaptive diffusion networks have become established as one of the main tools for distributed signal processing. In this context, sampling and censoring techniques have played an important role, since the costs associated with data acquisition, processing, and transmission across the entire network can become prohibitive in large-scale applications. Additionally, in sensitive domains such as those involving medical data, privacy preservation has ceased to be a desirable requirement and has become a fundamental necessity. As a result, privacy-preserving techniques have been incorporated into adaptive networks. However, as these approaches introduce noise into the shared parameters, the challenge of ensuring adequate levels of privacy without compromising algorithmic performance remains open.In parallel, the field of machine learning has experienced rapid growth, with applications becoming ubiquitous. The massive volume of data has imposed limitations on centralized processing, driving the development of distributed learning for neural networks. This project proposes to address challenges in these areas by integrating signal processing and machine learning.Within adaptive diffusion networks, the project aims to investigate optimal sampling strategies, kernel-based sampling and censoring techniques, approaches based on information-theoretic criteria, as well as extensions involving the distributed Kalman filter. In machine learning, the goal is to incorporate recent results from adaptive diffusion networks into the distributed training of neural networks, addressing in a unified manner topics such as sampling, censoring, and privacy. Furthermore, the project seeks to develop robust machine learning solutions for medical applications, with a focus on cardiac arrhythmias and epilepsy, as well as image super-resolution applied to electronic games. Finally, the project intends to study methods that enhance model interpretability without neglecting computational cost. In this way, the proposal is situated in an interdisciplinary context, combining theoretical rigor, methodological innovation, and practical relevance. The proposer's extensive experience in adaptive signal processing lends solidity to the proposal and enables a comprehensive and innovative approach to the machine learning problems considered. In addition, the project aims to intensify strategic actions focused on human resource training and on expanding the proposer's presence in the national and international scientific communities. (AU)