Representational Drift from Local Synaptic Learning in Recurrent Circuits

Neural representations in cortex are known to drift gradually over time, even under stable behavioral conditions. The origin of such drift—whether it reflects noise, instability, or a functional property of learning—remains actively debated. In this work, we show that representational drift can emerge naturally from local synaptic plasticity in recurrent circuits driven by structured, temporally correlated inputs. Using a minimal learning model with biologically motivated Hebbian-type updates, we find that feature selectivity evolves along a low-dimensional manifold defined by input statistics, giving rise to slow reorganization of internal representations despite stable coding performance. This framework links drift to the geometry of learned representations and provides a dynamical perspective on long-term changes in neural population activity. Our results suggest that drift may arise as a generic property of plastic neural circuits operating under normalization and recurrence, rather than as a consequence of external perturbations or noise.