Banca de QUALIFICAÇÃO: LUCAS HENRIQUE GOIS DE CAMPOS

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
STUDENT : LUCAS HENRIQUE GOIS DE CAMPOS
Date: 16/04/2026
TIME: 15:15
LOCAL: https://conferenciaweb.rnp.br/sala/aline-14
TITLE:

Advances in Signal Processing for Latent Component Estimation: Kernel Adaptive Filtering and Independent Component Analysis

PAGES: 62
BIG AREA: Engenharias
AREA: Engenharia Elétrica
SUMMARY:

This thesis presents two distinct contributions to the field of advanced signal processing and machine learning, addressing challenges in nonlinear filtering and neuroimaging data analysis. The first part investigates Kernel Adaptive Filtering (KAF), a powerful equalization method for nonlinear scenarios and impulsive noise environments where classic linear methods fail. By mapping input data into a high-dimensional Reproducing Kernel Hilbert Space (RKHS), KAF algorithms update a filter network of Radial Basis Functions (RBF) using Information Theoretic Learning (ITL) criteria, such as Kernel Maximum Correntropy (KMC). While the Gaussian kernel is commonly used in the literature, the Epanechnikov kernel is considered optimal for probability density estimation under certain conditions. Building on prior work that showed promising results but also presented instability with this kernel, we propose and analyze the application of sparsification methods to address the linear growth of computational complexity inherent to KAF, creating efficient input dictionaries while stabilizing performance.

The second part focuses on the blind source separation (BSS) of multi-subject functional Magnetic Resonance Imaging (fMRI) data. We evaluate constrained Independent Component Analysis (ICA) and Independent Vector Analysis (IVA) algorithms designed to capture complex patterns of functional brain organization. Specifically, we compare the performance of threshold-free (tf-cIVA) and adaptive-reverse (ar-cIVA/ar-cEBM) constraint formulations. We also proposed a novel domain-informed algorithm, ditf-cIVA that incorporates prior knowledge of functional domains into the regularization term. The new ditf-cIVA is shown to preserve correlations within domains while enforcing separation between distinct functional systems. Our studies on a cohort of 116 subjects demonstrates that this domain-informed approach yields superior spatial consistency and sensitivity to group differences in schizophrenia, highlighting the importance of tailored constraints in neuroimaging analysis.

COMMITTEE MEMBERS:
Presidente - Interno ao Programa - 1544392 - ALINE DE OLIVEIRA NEVES PANAZIO
Membro Titular - Examinador(a) Externo à Instituição - CHARLES CASIMIRO CAVALCANTE - UFC
Membro Titular - Examinador(a) Externo à Instituição - LEONARDO TOMAZELI DUARTE - UNICAMP
Membro Suplente - Examinador(a) Interno ao Programa - 2356637 - KENJI NOSE FILHO