Sequential critical periods support efficient local representation learning in a model of visual processing

Abstract

The emergence of abstract object representations in the mammalian ventral visual stream remains a central challenge for biologically plausible learning theories. While deep artificial networks trained via backpropagation achieve high performance, the algorithm lacks biological realism and conflicts with the staggered timeline of critical periods observed in cortical development. Here, we present a hierarchical model of the visual stream that learns invariant representations using only local synaptic plasticity rules modulated by lateral predictive signals. We demonstrate that imposing staggered critical periods -- where plasticity windows open and close sequentially from V1 to inferotemporal cortex -- significantly enhances representation quality for local learning rules, whereas it degrades performance in backpropagation-based networks. Furthermore, this sequential regime improves the learning efficiency in terms of number of synaptic updates, suggesting a metabolic advantage consistent with evolutionary constraints. We validate the functional utility of these acquired representations through reinforcement learning agents that successfully solve navigation and visual discrimination tasks without further fine-tuning of the visual encoder. These findings suggest that staggered critical periods are not merely a developmental constraint but a functional mechanism that enables efficient, local, and metabolically economical learning in hierarchical neural systems.