P001 A Recurrent Neural Network Model of Cognitive Map Development
Marco P. Abrate*1, Tom J. Wills1, Caswell Barry1
1Department of Cell and Developmental Biology, University College London, London, UK
*Email: marco.abrate@ucl.ac.uk Introduction Animals use an allocentric cognitive map of self-location, constructed from sequential egocentric observations, to navigate flexibly[1-4]. In the hippocampal formation, spatially modulated neurons support navigation, such as place cells, Head Direction (HD) cells, grid cells, and boundary cells[5-8]. The early development of these neurons is well characterised[9] but the mechanisms driving maturation and the relative timing of their emergence are unclear. We hypothesize that changes in locomotion shape the development of spatial representations. Combining behavioural analysis with a Recurrent Neural Network (RNN), we prove that movement statistics determine the development of spatial tuning, mirroring biological timelines. Methods Rats from post-natal day 12 (P12) to P25[10-12] were grouped according to their movement statistics. Rodent trajectories were simulated, using the RatInABox toolbox[13], in a square arena matching these locomotion stages. An RNN was trained to predict upcoming visual stimuli based on previous visual and vestibular inputs - mimicking the predictive coding function of biological systems[14]. The hidden units' activity was analysed against the position and the facing direction of the agent. Finally, these units were classified as place units based on their spatial information content or as HD units based on their Rayleigh vector length and KL divergence vs a uniform distribution - standard metrics for hippocampal neural recordings. Results Behavioural analysis revealed three distinct stages of locomotion during development with median ages P14, P15, and P21, respectively (Fig. 1a). The RNN trained on adult-like locomotion (Fig. 1b), solving the predictive task with biologically plausible inputs, showed spatially tuned units resembling hippocampal place and head direction cells (Fig. 1c). Crucially, when trained separately on simulated locomotion styles corresponding to the identified developmental stages, the model recapitulated the progressive emergence of spatial tuning observed experimentally. Specifically, spatial measures and consequently the number of units classified as place and head direction neurons steadily increased with improved locomotion (Fig. 1d). Discussion Our model establishes locomotion-dependent sensory sampling as a sufficient mechanism for cognitive map formation, extending predictive coding theories[3,4,15]. The RNN's ability to replicate spatial cell maturation patterns suggests that sensory-motor experience significantly shapes hippocampal spatial tuning. Furthermore, our results inform how manipulations of locomotion or sensory inputs could influence the development of spatial representations, which can then be tested in real-world experiments. Future work will directly compare the RNN's units with hippocampal neurons through Representational Similarity Analysis, search what drives grid patterns formation in our model, and investigate the changes in the geometry of the latent space. Figure 1. (a) 3-d UMAP representation of rats’ movement statistics coloured into locomotion stages. (b) Example of an agent’s trajectory (left) and snapshot of current visual input (right). (c) Architecture of the RNN. The latent space’s units are analysed for spatial responses. (d) Trend in the number of the RNN’s units classified as Place units (left), HD units (right), or Place and HD units (both). Acknowledgements NA References
