Emergence of perceptual reorganisation from prior knowledge in human development and Convolutional Neural Networks

Abstract

Perception in the mature human visual system relies heavily on prior knowledge. Here we show for the first time that prior-knowledge-induced reshaping of visual perception emerges gradually, and late in childhood. To isolate the effects of prior knowledge on vision, we presented 4-to-12-year-olds and adults with two-tone images, which are degraded photos that are hard to recognise on first viewing. In adults, seeing the original photo causes a perceptual reorganisation leading to sudden, mandatory recognition of the two-tone version - a well-documented process relying on top-down signalling from higher-order brain areas to early visual cortex. We find that children younger than 7 to 9 years, however, do not experience this knowledge-guided shift, despite viewing the original photo immediately before each two-tone. To assess potential computations underlying this development we compared human performance to three state-of-the-art neural networks with varying architectures. We found that the best-performing architecture behaved much like 4- to 5-year-old humans, who display a feature-based rather than holistic processing strategy akin to neural networks. Our results reveal a striking age-related shift in the reconciliation of prior knowledge with sensory input, which may underpin the development of many perceptual abilities.