A wide range of computational models have been proposed to explain how the hippocampus supports spatial and non-spatial reasoning. A recent model demonstrated that prediction of observations alone creates representations which contain spatial information and resemble the activity of hippocampal cells. This work explores three key questions. Does the prediction of visual elements in natural scenes induce a true, usable world model, challenging the notion that specialized neural architectures are necessary for spatial cognition? Do the ”place cells” identified in this model function in accordance with our current understanding of their biological counterparts? And, are functional cell types such as place cells truly ”functional”, or are they heretofore mislabeled correlates of sensory information? This investigation reveals that prediction of visual elements in this scheme induces not a cognitive map, but instead local and non-functional features which are easily misidentified as true place cells. The study further proposes that genuine hippocampal features may serve more complex functions than the image-processing artifacts that superficially resemble them.

PDF

Masters Thesis, Department of Computer Science, The University of Texas at Austin.

@mastersthesis{vonebers:msthesis24, title={From Sensory Input to Cognitive Maps: Exploring the Significance of Spatial Representations in Artificial Hippocampal Models}, author={Margaret Cordelia von Ebers}, month={ }, school={Department of Computer Science, The University of Texas at Austin}, address={Austin, TX}, url="http://www.cs.utexas.edu/users/ai-lab?vonebers:msthesis24", year={2024} }