Understanding the Orientation Tuning of Surround Suppression in the Human Visual System

Abstract

Orientation-tuned surround suppression is important for figure-ground segmentation and object detection. This dissertation aims to understand the visual characteristics and neural mechanisms of orientation-tuned surround suppression. A shortcoming of previous models of surround suppression is the implicit assumption of a uniform distribution of orientations and the failure to consider possible anisotropies. Thus, the potential impact of the absolute orientation of a target on the orientation profile of surround suppression has been neglected. It is well known that cardinal orientations are overrepresented in the visual system in comparison to oblique orientations. In this dissertation, my aim was to investigate how surround suppression may be modulated by the overrepresentation of cardinal orientations in natural scenes by measuring the tuning of surround suppression across orientation space. In Project 1, I compared the orientation tuning of surround suppression around cardinal and oblique orientations using psychophysical experiments. While human observers were more sensitive to small deviations from cardinal than from oblique orientations (the oblique effect), I observed an inverted oblique effect for surround suppression, with broader tuning of suppression around cardinal than oblique targets. In Project 2, I evaluated multiple computational models with different orientation anisotropies and different divisive normalization procedures. The only model that showed an inverted oblique effect assumed more cardinal-tuned than oblique-tuned neurons, and implemented Orientation-tuned surround suppression is important for figure-ground segmentation and object detection. This dissertation aims to understand the visual characteristics and neural mechanisms of orientation-tuned surround suppression. A shortcoming of previous models of surround suppression is the implicit assumption of a uniform distribution of orientations and the failure to consider possible anisotropies. Thus, the potential impact of the absolute orientation of a target on the orientation profile of surround suppression has been neglected. It is well known that cardinal orientations are overrepresented in the visual system in comparison to oblique orientations. In this dissertation, my aim was to investigate how surround suppression may be modulated by the overrepresentation of cardinal orientations in natural scenes by measuring the tuning of surround suppression across orientation space. In Project 1, I compared the orientation tuning of surround suppression around cardinal and oblique orientations using psychophysical experiments. While human observers were more sensitive to small deviations from cardinal than from oblique orientations (the oblique effect), I observed an inverted oblique effect for surround suppression, with broader tuning of suppression around cardinal than oblique targets. In Project 2, I evaluated multiple computational models with different orientation anisotropies and different divisive normalization procedures. The only model that showed an inverted oblique effect assumed more cardinal-tuned than oblique-tuned neurons, and implemented a non-selective local normalization component prior to orientation-tuned normalization across the target and surround. My results reveal that surround suppression leads to an inverted oblique effect that is likely because of multiple stages of divisive normalization in the visual system. In Project 3, I evaluated whether models that incorporate surround suppression do a better job of predicting neural responses to natural images. I found that implementing surround suppression improved the neural predictivity of Gabor-based V1 models and narrowed the gap between the traditional V1 models and the state-of-the-art convolutional neural network models. The findings from this research provide strong support for the notion that surround suppression is adapted to natural scene statistics and plays a role in extracting behaviorally relevant information from natural scenes. a non-selective local normalization component prior to orientation-tuned normalization across the target and surround. My results reveal that surround suppression leads to an inverted oblique effect that is likely because of multiple stages of divisive normalization in the visual system. In Project 3, I evaluated whether models that incorporate surround suppression do a better job of predicting neural responses to natural images. I found that implementing surround suppression improved the neural predictivity of Gabor-based V1 models and narrowed the gap between the traditional V1 models and the state-of-the-art convolutional neural network models. The findings from this research provide strong support for the notion that surround suppression is adapted to natural scene statistics and plays a role in extracting behaviorally relevant information from natural scenes.