Navigating without landmarks relies on path integration (PI), a computation thought to be performed by grid cells in the entorhinal cortex (EC). Classical models either assume fixed connectivity or fail to account for PI. To address this gap, we introduce a self-organizing continuous attractor neural network (CANN) model where PI and predictive coding emerge naturally through learned dynamics. Guided by local Hebbian plasticity, spike-frequency adaptation, and global inhibition, the network develops a continuous spatial code from structured input. Notably, this feedforward learning spontaneously generates prospective coding, anticipating future inputs. This emergent predictive shift strongly aligns with recent experimental findings in the superficial layers of the EC [1]. Furthermore, stacking multiple feedforward layers amplifies this predictive effect, forming a functional predictive hierarchy. When recurrent synapses are introduced, they self-organize under the same biological rules to support continuous PI, allowing the network to update its activity without external input. By bridging established frameworks [2,3], this model reveals how vital navigation and prediction capabilities can spontaneously emerge from basic principles of self-organization.
[1] Ouchi, A., & Fujisawa, S. (2024). Science, 385, 776-784.
[2] Kropff, E., & Treves, A. (2008). Hippocampus, 18, 1256-1269.
[3] Mi, Y., et al. (2014). NeurIPS, 27.
