Abstract
In the first months or years of life, newborns learn to perceive their environment at a remarkable pace. This capacity is often attributed to heightened neuronal plasticity that diminishes over time. While this decline in plasticity could be seen as constituting a mere biological constraint, we ask whether it might, in fact, serve an adaptive function. A natural consequence of time-limited plasticity is the preservation of sensory processing mechanisms formed early in life. Emerging evidence has supported the "adaptive initial degradation" hypothesis, which posits that early experience with degraded sensory inputs, such as blurred or color-reduced vision, helps instantiate important sensory mechanisms subserving robust perception later in life. Here, we employ deep neural networks as computational models to systematically probe how developmental improvements in visual fidelity interact with diminishing plasticity, modeled via decreasing learning rates. Across both acuity and color domains, we find that time-limited plasticity yields modest performance gains and helps stabilize early-formed representations. However, these improvements remain surprisingly modest. Further analysis of the networks' internal representations also reveals that earlier layers stabilize sooner than deeper layers, loosely paralleling findings from biological development. Overall, our results suggest that while declining plasticity can confer small benefits, it is not the primary force behind the more pronounced advantages of commencing visual experience with degraded inputs. These insights refine the "adaptive initial degradation" hypothesis and, more broadly, underscore how computational modeling can shed light on longstanding questions in developmental neuroscience. SUMMARY: Time-limited neuronal plasticity is often viewed as a mere biological constraint. However, might it provide adaptive advantages by preserving sensory mechanisms learned early in development? Simulations with initially degraded sensory inputs reveal modest performance gains when modeling time-limited plasticity as decreasing learning rates. Despite these gains, plasticity reductions contribute less than the benefits conferred by starting with degraded inputs, refining the "adaptive initial degradation" hypothesis. These results underscore the utility of computational models to illuminate questions in developmental neuroscience and point to interesting future research avenues.