Simulating the natural input to the visual system

Author: Dominik Straub

Supervisor: Prof. Constantin Rothkopf, Ph.D.

Submission: Aug. 2019

Abstract:

If an observer wants to make optimal inferences about their environment, their internal model should match the environmental statistics. The efficient coding hypothesis states that the visual system achieves this by adapting neural representations to the statistics of its natural input. This has motivated many studies modeling the statistics of natural images, which have revealed biases that are also found in the early visual system. However, the input to the visual system is not only shaped by the statistics of the environment but also by the imaging process due to the eye’s optical properties.

Here, images were generated in a virtual forest environment and transformed according to a model of the eye’s geometrical optics. Differences in image statistics between regions of the visual field were examined using power spectra, edge filters and sparse coding. Supporting previous research, both second- and higher-order statistics revealed a bias towards cardinal orientations as well as a radial bias that becomes stronger with increasing eccentricity. There were also differences between the upper and lower visual field and between images taken from the eye height of a human versus a cat. Furthermore, data from an fMRI experiment, in which subjects viewed oriented gratings, were analyzed to provide a link between input statistics and neural representations. The fMRI responses also showed a radial bias, which was strong even at low eccentricities. Interestingly, there was no discernible bias in image statistics at the eccentricities considered in the fMRI data, which could be due to the rather uniform orientation distribution in the forest scene. Future work should address this by analyzing different scene types.