Biologically inspired approaches to building spatial maps for spatial navigation and learning

• Ano de defesa: 2023
• Autor(a) principal: MENEZES, Matheus Chaves
• Orientador(a): OLIVEIRA, Alexandre César Muniz de
• Banca de defesa: OLIVEIRA, Alexandre César Muniz de , FREITAS, Edison Pignaton de , BARONE, Dante Augusto Couto , COUTINHO, Luciano Reis , ALMEIDA NETO, Areolino de
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO DOUTORADO EM CIÊNCIA DA COMPUTAÇÃO
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Navegação Espacial; Mapas Cognitivos; Aprendizado Latente; Representações Sucessoras; RatSLAM
• Palavras-chave em Inglês: Spatial Navigation; Cognitive Maps; Latent Learning; Successor Representation; RatSLAM
• Área do conhecimento CNPq: Arquitetura de Sistemas de Computação
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/5358

Navigating unfamiliar spaces and searching for resources, such as food and water, is fundamental for survival in many animals, including humans. For nearly a century, behavioral and cognitive neuroscience research has supported the existence of cognitive maps, which animals employ to navigate spatially. Cognitive maps enable animals to perform complex tasks, including acquiring a global map from distinct contexts once connections are established. Furthermore, studies have revealed that building a cognitive map before engaging in reward-based tasks can enhance learning speed, as evidenced by latent learning experiences. However, the specific factors contributing to the observed differences in learning speed, as influenced by experimental design and exploration strategies in latent learning, remain an open question. This doctoral thesis proposes novel computational approaches inspired by biological principles for building spatial maps to facilitate spatial navigation and learning. The Simultaneous Localization and Mapping (SLAM) algorithm, inspired by the navigation process in rodent brains, known as RatSLAM, has been extended by developing a novel structure merge approach to address the challenge of multisession mapping. RatSLAM is also integrated as a state representation learning algorithm within the CoBeL-RL framework, a reinforcement learning framework built on recent neuroscience findings, enabling agents to learn spatial tasks in unknown environments. By utilizing this framework, latent learning experiments are investigated to gain insights into the impact of different experimental designs and exploration strategies on learning speed. The results demonstrate RatSLAM’s successful performance in multisession mapping using real-world datasets and the ability of virtual agents to learn spatial tasks in unfamiliar environments. Additionally, it is shown that agents acquire distinct Successor Representations based on the specific experimental designs, providing a potential explanation for variations in learning speed for latent learning experiments. Overall, this thesis contributes to robotics and computational neuroscience by deepening our understanding of the cognitive processes involved in spatial navigation and providing practical insights for developing more effective robotic systems and computational models inspired by biological principles.