Denis G. Pelli, Ph.D. - US grants

More than a million people in the U.S. have severely impaired vision. Their visual impairments are diverse and there is only limited understanding of how these impairments affect performance of daily tasks such as walking and reading. The NEI and several workshops sponsored by it have identified a need for research on the visual requirements of everyday tasks. We seek to determine what aspects of visual capacity can be measured in the laboratory to predict mobility performance. Partially and normally sighted subjects will participate. Normally sighted subjects will wear goggles which restrict vision in various ways (field restriction, contrast reduction, and spatial-frequency cut off) and to various degrees. For all subjects, the residual vision will be characterized by contrast sensitivity and field measurements. Mobility of the subjects is now being tested in an indoor maze of foam rubber columns. The easily randomized maze allows repeated testing under controlled conditions. We will do these experiments at several luminances. Next we will study mobility in three real pedestrian environments: walking on the university campus, on a nature trail, and in a university classroom building. The aim is to determine what measurements of visual capacity will predict mobility performance.

[unreadable] DESCRIPTION (provided by applicant): How do we identify an object from the features that we detect? Understanding how the brain recognizes objects might give insight into how the brain solves problems in general. The object recognition problem has withstood a century of attempts, but we bring new tools - fruits of the last grant period - that allow us to sketch the outlines of a solution. Four approaches, all new and different, converge on one answer. AIM 1. Use crowding, along with other manipulations, to characterize three parallel processes in reading by normal and dyslexic readers. AIM 2. Count features by probability summation. Extending traditional probability summation from explaining just detection to also explain object identification, we acquire a new tool, allowing us to count the number of features the observer must detect in order to identify. AIM 3. Capture the observer's classification algorithm by computer modeling of the observer's responses to thousands of letters in white noise. We use statistical learning theory to build a classifier that accounts for human performance. The observer classifies each of several thousand images of a letter in noise as "a", "b", or "c", etc. These classifications are data that can tell us what the observer is doing. We use a powerful statistical learning algorithm to create a simple classifier that best models human performance. AIM 4. fMRI: Where in the brain are letters identified? Correlate the activation of the "letter" area in the left fusiform gyrus, and elsewhere, with two psychophysically-discovered signatures of letter identification: fast learning and channel frequency. Thus techniques from cognition, perception, statistical learning theory, and physiology together will reveal what is computed where, in the brain, when an observer identifies an object. [unreadable] [unreadable]

Results in hand tell us that letter identification is mediated by features (Pelli, Burns, Farell, and Moore, 1996). The features are bandlimited, with the same tuning as a spatial-frequency channel (Solomon and Pelli, 1994), and they are simple (Pelli, et al., 1966). This is consistent with the popular idea that features are like oriented center-surround receptive fields, or wavelets. We propose to: to use critical band masking in space and spatial frequency to identify the features that mediate visual detection and identification to use the candidate features to model letter identification performance with artificial and traditional alphabets. One reason for confidence in the success of this approach is that we have already succeeded in modeling word identification performance, taking letters as candidate features, showing that human performance approaches but never exceeds the performance of an otherwise-ideal observer that must base its decisions on independent latter identifications. Taking features to be smaller than a letter will bring the upper bound still lower making an even more stringent test. Exceeding that bound would disprove the conjecture that human performance is mediated by independent decisions on that set of features. Three other projects will use letters and noise to: characterize the channel mediating the poor acuity in amblyopic and normal fovea and periphery; (Pilot data indicate that our last-channel hypothesis may explain letter acuity.) measure equivalent noise in the periphery and compare it with predictions based on physiological measurements of ganglion cell noise; make fMRI measurements of contrast response paralleling our measure of efficiency of letter identification.

SUBPROJECT ABSTRACT NOT AVAILABLE

ABSTRACT Our long-term goal is to understand how the human brain recognizes objects. This 3-year project will characterize the computational kernel (computation that is applied independently to many parts of the image data) that is isolated by crowding experiments. We present the discovery that recognition of simple objects is performed by recognition units implementing the same computation at every eccentricity. These units are dense in the fovea and thus hard to isolate there, but they are sparse in the periphery, and easily isolated. Our fMRI & psychophysics pilot data show that each of these units, at every eccentricity, has a circular receptive field with a radius of 2.6±1.5 mm (mean±SD) in human cortical area hV4. Because of cortical magnification, that 2.6 mm corresponds to a tiny 0.05 deg in the fovea, but grows linearly with eccentricity, to a comfortable 3 deg at 10 deg eccentricity. We test this idea by pursuing its implications physiologically (Aim 1), clinically (Aim 2), and psychophysically and computationally (Aim 3). Aim 1. Better noninvasive measures for the health and development of visual cortex are needed. Conservation of crowding distance (in mm) in a particular cortical area (hV4) would validate crowding distance as a quick, noninvasive measure of that area's condition. Aim 2. Huge public interventions seek to help dyslexic children read faster and identify amblyopic children sooner. It would be valuable to know whether crowding contributes to reading problems and provides a basis for effective screening for dyslexia and amblyopia, as it can be measured before children learn to read. Aim 3. Documenting conservation of efficiency gives evidence that the same universal computation recognizes objects at every eccentricity. We are testing the first computational model of object recognition that accounts for many human characteristics of simple-object recognition. The new work extends to effect of receptive field size and learning.