The Neural Representation of Scenes in Visual Cortex

Recent neuroimaging studies have identified a number of regions in the human brain that respond preferentially to visual scenes. These regions are thought to underpin our ability to perceive and interact with our local visual environment. However, the precise stimulus dimensions underlying the function of scene-selective regions remain controversial. Some accounts have proposed an organisation based on relatively high-level semantic or categorical properties of the stimulus. However, other accounts have suggested that lower-level visual features of the stimulus may offer a more parsimonious explanation. This thesis presents a series of fMRI experiments employing multivariate pattern analyses (MVPA) in order to test the role of low-level visual properties in the function of scene-selective regions. The first empirical chapter presents two experiments showing that patterns of neural response to different scene categories can be predicted by a model of the visual properties of scenes (GIST). The second empirical chapter demonstrates that direct manipulations of the spatial frequency content of the image significantly influence the patterns of response, with effects often being comparable to or greater than those of scene category. The third empirical chapter demonstrates that distinct patterns of response can be found to different scene categories even when images are Fourier phase scrambled such that low-level visual features are preserved, but perception of the categories is impaired. The fourth and final empirical chapter presents an experiment using a data-driven method to select clusters of scenes objectively based on their visual properties. These visually defined clusters did not correspond to typical scene categories, but nevertheless elicited distinct patterns of neural response. Taken together, these results support the importance of low-level visual features in the functional organisation of scene-selective regions. Scene-selective responses may arise from the combined sensitivity to multiple visual features that are themselves predictive of scene content.