Our main goal is to construct computational models of the human visual system based on its measured abilities (psychophysics) and known architecture (neurophysiology and anatomy). These models use mainly neural network methods which can be simulated with conventional computers and ultimately implemented in specially-built hardware to run in “real time.” The models are designed to mimic as closely as possible the real anatomy and physiology of the human visual system and are not like any of the more conventional neural networks in common use. They are expected to give insights into the workings of the human brain when it is engaged in visual tasks and to make predictions about the response of the human visual system to real world images as well as to artificial laboratory stimuli. These models can also serve to endow robots or computer vision systems with the same or better abilities, especially if the models are implemented in special-purpose highly parallel hardware using architectures found in the brain.
Psychophysical experiments usually involve presenting carefully designed patterns on computer monitors to observers who are asked to report their judgments of such things as depth, color, velocity, flow, identity of an object, texture boundaries, or simply the presence or absence of a signal. Recently we have found that both depth and speed of motion judgments in a small local region are grossly affected by image features quite far away from the local region of attention. Similar effects are known for color and texture. We are therefore putting much effort into exploring these local-global interactions and building neural net models in which purely local node-to-node interactions have global perceptual consequences and in which information gathered from the whole image affects local perceptions.
Since these and many other visual processes are very complex and highly non-linear, conventional mathematical methods offer little hope of providing a useful bridge between psychophysics and neurobiology. Neural network models are therefore constructed using as much as practical of all we know of the psychophysics and neurophysiology of the situation. They are further provided with the ability to learn or adapt by changing their structure in response to repeated exposure to the relevant visual stimuli. In this way the nets “improve themselves” beyond the explicit knowledge used to construct them and sometimes give further insights which we hope will be useful in guiding further psychophysical and neurophysiological studies.
T. Kumar, B. Beutter, and D. A. Glaser, “Perceived motion of a colored spot in a noisy achromatic background,” Perception 22 (10), 1205, (1993).
T. Kumar, P. Zhou, and D. A. Glaser, “Comparison of human performance with algorithms for estimating fractal dimension of fractional Brownian statistics,” J. Opt. Soc. Am. A 10 (6), 1136, (1993).
T. Kumar and D. A. Glaser, “Initial performance, learning, and observer variability for hyperacuity tasks,” Vision Research 33 (16), 2287, (1993).
T. Kumar and D. A. Glaser, “Some temporal aspects of stereoacuity,” Vision Research 34 (7), 913, (1993).
T. Kumar and D.A. Glaser, “Depth discrimination of a crowded line is better when it is more luminant than the lines crowding it,” Vision Research, 35 (5), 657, (1995).
D. A. Glaser and D. Barch, “Motion detection and characterization by an excitable membrane: The “Bow Wave” model”, in Press, Computational Neuroscience (1998).