Dimensionality of population-level latent mechanisms encoding spatial representations

How does the brain efficiently encode space, and can this be achieved with low-dimensional neural codes? In this work, we show that the answer depends critically on whether spatial information is encoded continuously or discretely. We trained recurrent neural networks (RNNs) on two path integration tasks: one with continuous spatial outputs and another with discrete, place-cell-like responses. RNNs encoding continuous spatial coordinates developed low-dimensional latent dynamics, whereas in the latter case, the spatial resolution demanded by the place fields, rather than their number, determined the dimensionality of the neural code. Overall, by shifting focus from individual neuron tuning to population-level representations, our work identifies a fundamental constraint on the computational resources required for different neural coding strategies in spatial navigation.