Research

My dissertation research (with David Knill at the University of Rochester) focused primarily on how humans use different sources of visual information to control movement. In the lab, we examined how subjects rely on different monocular cues (aspect ratio and texture gradient) and binocular cues (disparity) to manipulate objects and interact with their world. Reliability is probably the most important factor; if one cue describes the world more consistently and accurately than other cues, people will emphasize information from this more reliable cue. However, if the cues are equally reliable, then the speeds at which different types of information are processed make a difference. From our experimental data, the visual system processes binocular cues faster than monocular cues. Since binocular information is processed faster and is therefore available sooner, people rely on it more when controlling their movements. This shows that the temporal dynamics of cue processing are important for determining how we use visual cues.

I also investigated how cue integration strategies change when the information is presented in the periphery and at different depths relative to the fixation plane. Most research has focused on how we use visual information that is presented to the fovea under central fixation, but humans frequently rely on information from other regions of the visual field (like when adjusting the radio while keeping one's eyes on the road). Although visual acuity decreases with distance from the fixation point, the reliabilities of monocular and binocular information decrease at different rates, and we found that this affects how information from these different sources is utilized. A further research question was whether what one is doing affects how one uses available visual information. Subjects performed object placement and prehension tasks using exactly the same visual stimuli. The required judgment about the 3D orientation of a textured disc was the same in both cases, but we found that differences in task demands affected how they used the information provided by the different cues.

Another research direction involved developing computational models of binocular vision that aim to extend our current beliefs about how we use disparity information to estimate depth and 3D orientation. If you have a computer with two adjacent cameras, what does it need to do in order to have effective, biologically-plausible stereo vision? Finding corresponding points and computing raw disparity in the images from each camera can work a lot of the time, but there are clear cases where it fails, and so it is unlikely that the brain uses such an approach. Energy models (e.g. Ohzawa et al.) have approached the correspondence problem by modeling the outputs of neurons in primary visual cortex, and these models and their derivatives are a good starting place for approaching this question. We simulated a population of visual cortical cells tuned to orientation disparity, which has been proposed as a binocular cue for 3D orientation, and analyzed the resulting activity patterns to quantify the amount of information this cue provides.

Other areas of interest include modeling early visual cortical areas, developing biologically-plausible robotic vision algorithms, and investigating neural interfaces (brain-machine interfaces).