“Geometric Investigations of Representations, Learning, and Generalization in Minds, Brains, and Machines”
During this CONECT seminar, Leyla Roksan Caglar will present her work.
Understanding how abstract information is represented in the brain - and how such representations support efficient storage, retrieval, and generalization - has been a central challenge in cognitive science, neuroscience, and artificial intelligence. In a series of studies, I will illustrate how my research addresses these questions by 1) applying geometric and topological approaches to probe the structure of neural representations, and 2) by using comparative approaches across minds, brains, and machines. First, I will use a predictive encoding model and geometric analyses to show that object-directed action representations (e.g. of tools) contain motor information in a compositional manner that facilitates flexible retrieval. Then, I will use mathematical models of similarity to investigate the structure, shape, and metricity of representational manifolds and their congruency across behavioral and neural data. The results illustrate the importance between representational form and function, as well as the learning process’ representational constraints. Building on this, I will discuss some ongoing work exploring the intimate link between information-compression, generalization processes, and the shape of neural representations, illustrating how representations are dynamically adapted to task performance. Using a comparative lens across biological and artificial neural information processing systems and bridging information theoretical approaches with topological data analysis, this work aspires to uncover fundamental principles of generalization shared across humans, animals, and artificial neural networks.