Banca de DEFESA: ANGELO GUSTAVO SOUZA MARINHO MORAIS DE SALES
Uma banca de DEFESA de MESTRADO foi cadastrada pelo programa.STUDENT : ANGELO GUSTAVO SOUZA MARINHO MORAIS DE SALES
DATE: 14/08/2026
TIME: 14:00
LOCAL: https://meet.google.com/aue-xihb-zvh
TITLE:
KEY WORDS:
PAGES: 61
BIG AREA: Ciências Biológicas
AREA: Biologia Geral
SUMMARY:
The decoding of animal communication frequently groups vocalizations into discrete categories based solely on static acoustic structure, limiting the understanding of the rich information contained in sequential order and intra-category variations. This thesis investigated the sequential structure of common marmoset (Callithrix jacchus) vocalizations, proposing novel computational approaches based on deep learning and state modeling to overcome the limitations of traditional categorization. The first study addressed the information loss inherent in standard acoustic tokenization. By developing a reassignment transformer, it was demonstrated that reinjecting acoustic embeddings (AVES) into the model improves predictive power, lowering the crossentropy of next-call predictions by 8.6%. A centroid ablation further showed that most of this gain comes from the acoustic variation within a category rather than from its average: replacing each call with its category-mean embedding discarded the majority of the improvement. Building on the premise that a call is defined by what follows it, the second study introduced a complementary approach focused on the unsupervised discovery of vocal categories from their sequences. Employing predictive-state reclustering and variable-length Markov chains (VLMC), calls were categorized by their behavioral consequences in the sequence rather than merely how they sound. This approach reorganized the marmoset repertoire, revealing, for example, that the acoustically variable phee call actually consists of four distinct sub-types, each occupying a specific statistical position in the vocal stream. The method successfully generalized to a taxonomically distant species (Bengalese finch). It is concluded that while acoustic nuances are highly predictive, the ultimate boundaries of a vocal category must be grounded in its behavioral role. The integration of deep attention mechanisms with ecologically grounded state reclustering provides a new, behaviorally valid framework for decoding the structural complexity of animal communication
COMMITTEE MEMBERS:
Externo à Instituição - THIAGO TARRAF VARELLA
Presidente - 3086031 - DANIEL YASUMASA TAKAHASHI
Externo ao Programa - 2069422 - DIEGO ANDRES LAPLAGNE - nullInterno - 3083298 - RENAN CIPRIANO MOIOLI