Endogenous and Exogenous Factors Shaping Visual Development in Early Infancy

Abstract

The human mind is capable of forming elaborate concepts and models of the world and manipulating them so that it can adapt to complex, rapidly changing environments. In contrast with that, machine learning methods and artificial neural networks (ANNs) only perform well within narrow task constraints and are poor generalists. Infants and preschool children are capable of learning things that have been shown to be challenging even for cutting edge computational methods, especially when it comes to scarce or unbalanced sets, distorted data and the ability to adapt to new situations. This difference is often attributed to the more complex neural architecture of the brain in comparison to ANNs. However, another potential factor contributing to these differences is the structure of the learning process itself. ANNs are often trained in a static way, with batches of examples randomly sampled from the whole training set throughout. In contrast, an infant?s experience is dynamic and non-stationary on many timescales. Consider the learning of visual classes. Some factors are exogenous. In the real world, during any short period, some visual classes are much more common than others, while over time the distribution of different classes fluctuates, depending on the context the infant finds themselves in. Other factors are endogenous. For example, over the first few months of life, the eyes undergo rapid development and visual experience changes dramatically. Initially, visual acuity is low and input almost monochromatic, but in a few months it become sharp and richly coloured. Furthermore, the infant is not a passive viewer. Following their curiosity, they can move their eyes and apply selective attention, to allow them to shape their own curriculum. These rich trajectories during learning will change the nature of the underlying representations and may play a valuable role creating robust and generalisable concepts. Our goal is to characterise these factors and study how they shape infant learning and change ANN performance. We investigated one exogenous factor - the statistics of categories observed by infants in their early years, and two endogenous factors - the impact of visual maturation on perceptual preferences, and the limitations of self-driven learning. Our results suggest that ML algorithms may benefit from better training procedures and better design of their curriculum.