Dynamical, statistical, and computational properties of biological and artificial neural networks are being studied with the aims of (i) elucidating the ways in which information processing is conducted in the brain and (ii) developing new tools for solving classification and function approximation problems in the biomedical and physical sciences.
Neurobiological Computation
Professor Clark and his students are engaged in a program of research in computational neuroscience in collaboration with Charles H. Anderson and other faculty in the School of Medicine. The central hypothesis being explored by the computational neuroscience group is that ensembles of neurons, through their collective activities, encode and process functions of analog variables pertaining to the organism's environment or internal states. Examples of such variables are limb orientation and velocity, image contrast and color, and measures of depth and optic flow. The formulation being developed embodies Bayesian principles of probability theory and statistical inference advanced by the late Edwin Jaynes, formerly Wayman Crow Professor of Physics. Accordingly, it is envisioned that the neural computational system will process uncertainty estimates along with estimates of mean values. An extension of Bayesian belief nets, together with the mathematics of signal processing, neural coding, and overcomplete functional representation, provides a general framework for the design of large-scale cortical circuits in which top-down models influence bottom-up processing of external stimuli. These ideas, along with more traditional modeling approaches such as dynamical systems theory, are being applied to problems in vision, sensory-motor processing, synaptic competition, and learning. Current projects of the group include the incorporation of realistic models of spiking neurons within the probabilistic description based on population coding; the study of continuous attractors as vehicles for short-term storage of information on analog quantities; investigation of the form and function of nonlinear spatio-temporal interactions on dendritic trees; and simulation of neural behaviors in which expectation- or model-driven top-down information influences input-driven bottom-up processing of external stimuli.
Artificial Neural Networks
Professor Clark and his co-workers have also been developing and applying advanced methods for classification and function approximation involving artificial neural networks that can learn by example. Significant algorithmic innovations include the introduction of an extended version of backpropagation learning that is optimally suited for training on noisy data and the design of two-layer networks (higher-order probabilistic perceptrons) that are capable of modeling all correlations among a finite number of input variables. These methods along with other artificial-intelligence algorithms are being employed to create global statistical models of the properties of complex physical systems, including atomic nuclei and crystalline materials of technological interest. The same techniques may be applied to medical diagnosis and analysis of biomedical images.
Research Publications
Dellen BK, Clark JW, Wessel R (2005 Mar-Apr). Computing relative motion with complex cells. Vis Neurosci. 22 (2): 225-36. Full Article >
Dellen BK, Clark JW, Wessel R (2004 Sep). Motion-contrast computation without directionally ive motion sensors. Phys Rev E Stat Nonlin Soft Matter Phys. 70 (3 Pt 1): 031907. Full Article >
Khodel VA, Clark JW, Zverev MV (2001 Jul 16). Superfluid phase transitions in dense neutron matter. Phys Rev Lett. 87 (3): 031103. Full Article >
Clark JW, Gernoth KA, Dittmar S, Ristig ML (1999 May). Higher-order probabilistic perceptrons as Bayesian inference engines. Phys Rev E Stat Phys Plasmas Fluids Relat Interdiscip Topics. 59 (5 Pt B): 6161-74. Full Article >
Clark JW (1991 Oct). Neural network modelling. Phys Med Biol. 36 (10): 1259-317. Full Article >
Witt JC, Clark JW (1990 Apr). Experiments in artificial psychology: conditioning of asynchronous neural network models. Math Biosci. 99 (1): 77-104. Full Article >
Contact Info
John Clark, Ph.D.
Office Location: 351 Compton Hall (Hilltop)
Office Phone: 314-935-6208
Campus Box: 1105
Fax: 314-935-6219
jwc@howdy.wustl.edu
