Publication
Neural Computation of Statistical Image Properties in Peripheral Vision
Christoph Zetzsche; Ruth Rosenholtz; Noshaba Cheema; Konrad Gadzicki; Lex Fridman; Kerstin Schill
In: Vision Science Society (Hrsg.). MODVIS. Computational and Mathematical Models in Vision (MODVIS-2017), located at Vision Sciences Society Annual Meeting, May 17-19, St. Pete Beach, FL, USA, Purdue University, 5/2017.
Abstract
In the peripheral field of view our visual system provides a much lower image quality than in the central region. This has often been attributed to a mere loss of spatial acuity, but recent investigations suggest that the system uses a more refined strategy. For lowering its data load it computes a statistical summary representation based on low-level image features. In a recent modeling approach the summary statistics refer to classical statistical measures, like auto- and cross- correlations, operating on wavelet-like filter outputs.
Here we investigate how such a statistical representation can be obtained in a neurobiologically plausible fashion. For this, we consider both the elementary neural operations and the architectural properties. For example, the neurobiological plausibility of multiplications which are an essential component of classic statistical operations, is unclear and often critically debated. Also, it remains to be determined how the characterization of a statistical distribution, classically achieved by moments or histograms, can be achieved by neurobiological hardware.
Furthermore, it is unclear which specific visual features are actually best suited for an efficient statistical summary representation. We address these problems by considering how basic neural nonlinearities, like cortical gain control, can contribute to the computation of statistical properties, how basic neural selectivities, e.g. for intrinsically two-dimensional signals are related to statistical features, and how results from the promising domain of deep learning can help to understand the role of architectural properties in a statistical representation. We show that the visual cortex can provide a reliable statistical characterization of the visual environment, and we discuss which role this representation can play for different visual tasks, e.g., for object recognition, gist estimation, and localization.