A self-organizing developmental framework for the head direction ring attractor

Abstract

The head direction (HD) system provides a continuous representation of the animal's facing direction and is a core component of the spatial navigation network. Although mature HD circuits have been studied extensively, their developmental origin remains unclear, particularly because young animals already show attractor-like internal coherence before stable visual anchoring is established. Here we propose a developmental computational framework of the lateral mammillary nucleus and dorsal tegmental nucleus (LMN-DTN) circuit, in which a functional HD ring attractor emerges through staged self-organization without preconfigured topology or early external supervision. In the model, a randomly connected LMN network first generates stable travelling waves through relative refractoriness. These waves provide a temporal scaffold for subsequent learning. Local STDP-like and structural plasticity then consolidate wave-based statistics into locally enhanced recurrent connectivity, followed by direction-selective shift-feedback learning in the DTN→LMN pathway. At a later stage, recruited long-range inhibition transforms the travelling-wave manifold into a stable single activity profile (bump) attractor. Visual input acts mainly as a late calibration signal for global integration gain: it improves short-timescale path integration, but does not remove long-term drift. The framework provides a testable account of the developmental dissociation between early internal organization and later sensory anchoring. More broadly, it suggests that key components of classical ring attractor models, including local recurrent structure, direction-selective shift feedback, and late stabilizing mechanisms, can arise through a staged developmental process rather than relying on a preconfigured network architecture.