A walk through the woods explains the space variant oblique effect

Abstract

Learning of receptive fields from natural image ensembles has usually assumed stationary statistics across the visual field and implicitly also independence from the active selection process due to eye movements as well as the imaging geometry. Such models cannot explain that visual performance depends on the position within the visual field. As an example, perceptual studies in humans have demonstrated reduced discrimination ability for obliquely oriented patterns as compared to horizontal and vertical ones in the central area of the visual field 1. This so called oblique effect varies with the position within the visual field in that with increasing eccentricity sensitivity to meridionally oriented stimuli increases 2. What is the origin of this space variant oblique effect? While it has been proposed to understand some aspects of the oblique effect through the second order statistics of natural images, as quantified by their power spectrum 3,4, we investigate the role of higher order dependencies and task behavior on properties of receptive fields across the visual field. Image sequences are obtained by simulating an agent navigating through wooded environments. Standard unsupervised learning algorithms are used to learn a sparse code of the images but contrary to previous approaches, the learning is separately carried out for different regions of the visual field and different gaze allocations during walking. The properties of the learned receptive fields are quantified through the parameters of best fitting Gabor functions. This analysis shows that the receptive fields have a preference for horizontal and vertical orientations at the center and more and more meridional directions in the periphery. They also show systematic changes with increasing eccentricity. Furthermore, we relate these results to the behavioral measurements through Fisher information, demonstrating qualitative agreement with the space variant properties of the oblique effect. Furthermore, it is shown that the distribution of preferred orientations significantly depends on where gaze is directed within the visual field during walking. Artificial scenes using textures of natural images only up to second order, as well as textures sampled from dead leaves models shown that the second order dependencies are not sufficient for observing the receptive fields anisotropies. The analysis was repeated on a dataset obtained from human subjects walking through a wooded environment also demonstrating comparable results. We conclude that the properties of model receptive fields not only depend on the statistics of visual scenes in the environment, but also on the statistics imposed on the stimulus by the imaging geometry and the statistics of the interaction with the environment during natural task execution. Taking all these determinants together, receptive fields can be learned that explain the space variant oblique effect. Acknowledgements: This research was supported by EC MEXT-project PLICON and the German Federal Ministry of Education and Research within the "Bernstein Focus: Neurotechnology" research grant.

Cite this document ^(BETA)