Deep Learning for Nonlinear Blind Source Separation

• Grant number: 25/12174-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: September 01, 2025
• End date: August 31, 2026
• Field of knowledge: Engineering - Electrical Engineering - Telecommunications
• Principal Investigator: Levy Boccato
• Grantee: Guilherme Filizatti Ramalho
• Host Institution: Faculdade de Engenharia Elétrica e de Computação (FEEC). Universidade Estadual de Campinas (UNICAMP). Campinas , SP, Brazil
• Company: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação (FEEC)
• Associated research grant: 20/09838-0 - BI0S - Brazilian Institute of Data Science, AP.PCPE

Abstract

This scientific initiation project focuses on the problem of blind source separation in the context of post nonlinear (PNL) mixtures and aims at studying a separation framework based on autoencoders (AEs) and adversarial learning. The research will be based on the Anica model, originally proposed by Brakel and Bengio (2017), in which an encoder-decoder structure was combined with a discriminator network to simultaneously produce an informative latent representation, in the sense that the mixtures can be reconstructed from the latent variables, and achieve statistical independence among these variables, thereby recovering estimates of the sources.Our goal is to explore and expand the potential of Anica by incorporating known hypotheses about the mixing system, such as imposing monotonicity on the nonlinear functions of the decoder, using monotonic network architectures. Furthermore, we intend to investigate efficient estimators of information measures, such as the Kullback-Leibler divergence, to replace the discriminator network (e.g., via the MINE method), seeking to improve training stability and/or convergence, and reduce the need for large volumes of data.The developed separation frameworks will be applied in various scenarios of the problem, considering different types of nonlinearity, mixing matrices, and sources, so as to evaluate the performance of each technique and establish a comparative overview among them. (AU)