Adaptive Use of Natural Image Statistics in the Human Visual System
Speaker: Margaret Henderson
Our everyday visual experiences are highly structured to convey essential information about the world around us. For example, the color of an apple informs us about its flavor and ripeness. My research explores how the visual system takes advantage of these regularities to enable our behavior, by integrating knowledge of statistical structure with our current behavioral goals. I will present work that addresses this using experimental and computational techniques. First, I use computational encoding models of functional magnetic resonance imaging (fMRI) responses to establish a link between natural image statistics and the organization of human visual cortex. I demonstrate that neural populations in higher visual cortex are tuned for category-diagnostic features, and demonstrate how a texture statistics encoding model can be used to probe mid-level feature selectivity. Second, I present an fMRI study investigating how the visual system achieves flexible categorization of shapes. These results show that early visual cortex representations are biased during categorization so that stimuli become more discriminable across an active category boundary, suggesting that knowledge of category statistics can modify representations of incoming sensory information. Third, I use convolutional neural network models to explore the relationship between image statistics and category learning. In this work I establish that low-pass filtering (blurring) images can improve basic-level category learning in a model, and discuss these findings in relation to early human category learning. I will close with an overview of my planned research program, which will utilize new experimental and computational approaches to better understand how visual cognition is informed by the structure of the sensory environment.