College of Liberal Arts
There are two mains lines of current research in this lab. The first focuses on developing a compositional, generative model of human vision. This work builds on the lab's previous empirical brain-imaging findings regarding the functions of bidirectional neural processing between cortical areas. Human ability to accurately recognize objects and their local properties relies on both feedforward and feedback (generative) mechanisms operating between cortical areas. The group's results have suggested that the brain's computations solve the inherent local ambiguities in images not only through the well-known "deep" hierarchical organization of the visual system, but also through feedback connections that generate error signals as well as binding signals. In collaboration with computer vision colleagues at Johns Hopkins, this lab is now developing computational models whose goal is to take high-dimensional natural images as input and predict behavioral outputs that match the pattern of successes and failures of human observers. In addition, the models should have components that are consistent with the biology. The particular focus is on the visual perception of human body pose - a domain in which human "expertise" outstrips machine vision algorithms in the ability to deal with occlusions, and critically an enormous range of tasks. Previous work has relied on simple low-dimesional stimuli where predictions and interpretations are straightforward. Because of the complexity of natural images, model development will increasingly rely on large natural image datasets and methods from machine learning, requiring the use of advanced computing facilities.
The lab is just beginning a new, second line of research on the visual perception of natural flows (e.g. smoke, fire, river water, honey, etc.). This is an area of human vision research that we know very little about. Some of the lab's past work has shown how some features of natural flows (e.g. flow sinks and sources, from a shiny vs. matte object) are used by human vision. But shiny vs. matte just scratches the surface of the perceptual inferences that humans make. This group is now making use of a number of existing dynamic video datasets and machine learning methods to discover potential spatio-temporal features for perceptual tasks such as "how viscous?", or "how shiny?". Advances in this line of research will also benefit from high performance computing resources.