Prenatal and postnatal development of laterally connected orientation maps

Abstract

Both environmental and genetic factors interact to produce the orientation maps found in the primary visual cortex of adult mammals. However, it is not clear how this interaction occurs during development, or whether both factors are crucial. Previous computational models have focused on either environmentally driven or genetically driven development alone. In contrast, we show that a two-stage model of development can account for a wider range of experimental data. The model explains how environmental and genetic information can be incorporated into the same neural hardware, using a common set of learning mechanisms. Our results suggest that while either environmental or genetically driven development is sufficient for maps and selectivity to form, prenatal activity speeds up early development and makes it more robust against environmental variation.

Introduction

Experiments in mammals such as cats, ferrets, and monkeys have shown that the orientation map patterns and receptive fields in primary visual cortex (V1) are produced by an interaction between environmental and genetic factors. Genetic influences are clear when measuring neural responses in newborns, before any visual experience. Even at or before natural eye opening, orientation-selective cells and orientation maps can be detected in newborn kittens and ferrets [5], [8], [9], [11], [12]. The overall shape of the orientation map changes very little during subsequent normal visual experience [9], [11]. Based on this evidence, one might conclude that orientation maps are largely genetically specified.

At the same time, it is clear that altering the visual environment can have dramatic effects on how orientation-selective neurons and maps develop. For instance, if kittens are raised in environments consisting of only one orientation of contour (e.g. vertical lines) during a critical period, an abnormally large number of their V1 neurons become responsive to vertical orientations [4]. Kittens raised in this way develop orientation maps with larger area devoted to the overrepresented orientation [14]. Cats whose visual environment was even more abnormal, e.g. with eyelids sutured shut during development, have few orientation-selective neurons at all in V1 [5], [11]. Even in normal adult animals, the distribution of orientation preferences is slightly biased towards horizontal and vertical [7], [10], mirroring the distribution of orientations in normal visual environments [17]. Such a bias is consistent with neurons learning orientation selectivity from the environment. Orientation selectivity also improves greatly after birth, and the maps become smoother and more well organized [9], [11]. Based on the postnatal experiments, one might conclude that orientation maps develop through visual experience alone.

Taken together, the evidence indicates that both genetic and environmental influences interact to produce the adult orientation map. However, important questions remain. How does this interaction actually occur? Could adult-like maps develop from environmental or genetic cues alone, or are both necessary? These questions are difficult to answer through biological experiments. Computational modeling, however, can lead to valuable insights, because it is easy to separate environmental and genetic influences in computational experiments. Existing models have been used to simulate how orientation maps can develop from visual input alone (e.g. natural images; [6]) or genetic factors alone, such as spontaneous neural activity (e.g. noise [13]; for review of existing models of each type, see [16].) However, models have not yet shown how V1 can have an initial map at birth that becomes smoother and more selective due to postnatal visual experience, while retaining the original map shape. The visually driven and internally driven models also differ in many ways besides the source of activity, and thus it has been difficult to determine whether the activity patterns alone account for any differences between the results.

In this paper we construct a single model to show how an initial map can develop from spontaneous neural activity, then be refined through visual experience to reflect the environment. The model develops orientation maps and selective neurons for a very wide range of training patterns. This result suggests that the processing implemented in orientation maps is very general, and does not need to be encoded specifically in the genome. The role of spontaneous activity may primarily be to speed up development so that even newborns have functional neural processing. This ability may make development more predictable in unusual environments, and allow higher levels to organize sooner.