Model study on developmental mechanism of innate functional circuits in the visual cortex

Abstract

In the ventral visual system, various functional circuits have been reported on scales ranging from single neuronal tuning to cortical map levels, even before the time of eye-opening. For example, neurons in the primary visual cortex show selective responses to visual features, such as orientation, ocular dominance, and the spatial frequency of visual stimuli, and these neurons form quasi-periodic cortical maps across the cortical surface. It has also been reported that such cortical maps are systematically organized to achieve efficient visual coding, such as a ‘hypercolumn’, which enables V1 sites to sample a complete set of visual features at any cortical location. These observations imply that the visual cortex already forms an efficient organization of basis functional circuits for further visual processes, even before the visual experience. However, it remains elusive as to how such systematic organizations arise innately. In this thesis, I suggest that a feedforward connection from the periphery could seed the initial blueprint of functional circuits that would be refined by cortical activity and visual experience later, rather than the scenario in which these functional circuits develop from a totally random organization. From this perspective, using a computer simulation I show that (1) the initial blueprint of systematically organized functional circuits in the early visual pathway could be seeded by the physical projection of the retino-cortical pathway, and (2) the initial tuning of higher-order cognitive functions could also be developed by the stacking of physical feedforward connections. First, I suggest that because the alignment and distance between ON and OFF retinal ganglion cells could seed the initial selectivity of orientation, ocular dominance, and spatial frequency in the primary visual cortex, the orthogonal organization of multiple cortical maps could be seeded by the interference pattern of the retinal mosaic. This retinal origin model can explain the development of several characteristics observed in the orientation map. For example, because orientation selectivity in V1 is seeded by the retinal structure, an asynchronous retinal wave could induce the co-activation of V1 neurons with similar orientation preferences, resulting in clustered horizontal connections linking these neurons. This model also provides an explanation of how different retino-cortical mapping ratios of diverse species induce distinct spatial organizations of the functional tuning, specifically the columnar and salt-and-pepper map, as the spatial organizations of cortical map are seeded by the retinal interference structure. Next, to expand the notion that the physical structure of the neural system could seed the initial blueprint of functional circuits from the early visual pathway to a higher visual area, I suggest that the stacking of such a physical feedforward projection could induce the spontaneous development of a face-detection function even without a training process. Using a deep neural network as a model of the ventral visual stream, I show that random feedforward connections could induce face-selectivity in the complete absence of training. In summary, these results suggest that the physical structure of the neural system, specifically the projection from the retina and the stacking of feedforward projection, could provide an initial blueprint of functional circuits before it would be further developed by cortical activity and visual experience. This research provides insight into the origin of innate cognitive functions in both biological and artificial neural networks.