An accessible hippocampal dataset for benchmarking models of cognitive mapping

Recent empirical studies have revealed spatial tuning in traditionally non-spatial brain regions, as well as coding differences across species in traditionally spatial regions. These findings have in part motivated new models of cognitive mapping that make quantitative predictions at multiple levels of explanation. Taken together, these advances highlight a growing need to quantitatively compare spatial codes across brain regions, species, levels of explanation, and against computational models. To address this need, here we leverage a Representational Similarity Analysis framework to quantitatively compare the geometric determinants of the rodent hippocampal code to model predictions. We recorded large hippocampal populations via head-mounted miniscope imaging as mice repeatedly explored ten uniquely-shaped environments generated by opening and closing parts of a 3 x 3 grid. Traditional analyses replicated and extended several classic findings while also revealing previously unreported dynamics, such as stabilization of representational drift with experience and more heterogenous single-cell responses in more complex shapes. Critically, the grid-based design of this shapespace afforded the construction of a rich and reliable (noise ceiling: Kendall’s Tau ~0.7) representational matrix summarizing the similarity of the hippocampal code within and across environments – ideal for quantitative comparison with other assays via the RSA toolbox. We then compared this empirical matrix to those generated from several popular models of spatial coding, including a stable model, a dynamic boundary-tethered model, boundary vector cell (BVC) models, and boundary-based successor representation models. Of the models tested, only some parameterizations of BVC and boundary-based successor models approached the theoretical maximum fit. These results thus demonstrate that this dataset can serve as an empirical benchmark for tuning, evaluating, and comparing models of cognitive mapping. Moreover, these results motivate future work characterizing and comparing the geometric determinants of spatial representations across brain regions, species, and levels of explanation with this paradigm.