Symposium Explores Paradigm Shifts in Cognitive Modeling

At the recent theme symposium of the 1999 International Joint Conference on Neural Networks held in Washington, D.C., neurobiologists, psychologists, and engineers presented a wide variety of research projects attempting to make sense of the human brain. But despite the variance among disciplines and approaches, the symposium, entitled "Neuronal Ensembles: Paradigm Shifts in Cognitive Modeling," had a common thread: elucidating brain function by examining the behavior of groups of neurons rather than single neuronal cells. "Most biologists are concerned with specific molecular processes in, say, individual cells," explains symposium coorganizer Yan Yufik, president of the Institute of Medical Cybernetics in Gaithersburg, Md. "[That's] very important, but [it] doesn't give you an insight into how processes in the human nervous system are different from what's happening in, say, the fish." Experiments at the cellular level cannot, says Yufik, speak to the complexity of the human nervous system. In fact, many neurobiologists agree that the elementary unit of brain function is actually an ensemble, a functional group of neurons, and that the brain is a complex system of systems composed of such functional groups. Characterizing these ensembles as functional units in a complex system may require the work of both neurobiologists and engineers. Engineers' "top-down" perspective, their biology-inspired attempts to artificially model the brain, could greatly assist biologists in their attempts to deduce brain function based on the observed molecular biology of different brain regions. In his plenary lecture, Lawrence Fogel, president of the La Jolla, Calif.-based Natural Selection Inc., argued that if the function of a brain segment were defined well enough, engineers should be able to develop an off-line mechanism that mimics that function and corresponds to what the biologist has discovered independently. "It becomes a model for the biologist to look at," says Fogel. "Let's not disregard what the biologist has to say bottom-up. [Biologists are] doing reverse engineering. They're looking at the pieces and saying how do these fit together. And I'm saying, for what function? If you don't know what function you're looking for, you'll never find it." Yufik presented the merits of a modeling approach called the Virtual Associative Network (VAN), which attempts to capture the essence of brain function and goes against the prevailing modeling notions. First developed in collaboration with Tom Sheridan of Massachusetts Institute of Technology in the early 1990s for NASA and the National Science Foundation, and now supported by the U.S. Army and Navy, the VAN has had practical applications as an aid in high-pressure, multiple-variable decision making of the kind soldiers, sailors, or pilots might face during battle. But, says Yufik, agencies are more generally interested in the VAN because it's in complete contradiction to traditional cognitive paradigm models. "The umbrella theme is to try to take diverse biological data about human performance combined with data about actual biological processes that take place in the human brain and put them into a coherent, unified framework," he comments. "This is really the overarching interest: to understand how things work in biological reality." The traditional cognitive modeling paradigm holds that humans learn by acquiring "templates" of experience and perform by matching the present situation to an archived template that is not causally connected to the others. Therefore, learning consists of building an archive of templates, and performance consists of pulling an element from that archive. Each template contains action recipes, which explains, the model holds, how experts make their decisions. According to this model, the brain maximizes speed and efficiency by assigning a number of these different templates, or "specialists," to perform certain tasks that can be performed in parallel since the templates might not be causally connected. This concept has been the basis for technologies such as neural networks, which have had minimal success at explaining both brain function and human behavior. The VAN model implies that these traditional notions mostly apply to animal learning and not to significantly more complex human learning. Its contention: The key to modeling brain function is mimicking our integration of experiences. Virtual networks are composed of integrated components that are mutually constraining, a feature that enables the VAN to quickly zero in on the relevant pieces of the assigned task; the traditional notion of parallelism isn't applicable. Neuronal ensembles are the substrate of such components. Experience influences contents and the order in which components are processed, but the decision maker can combine components and adjust them to the present situation. In this way, human behavior is not stereotyped, as the template model implies, but it can take advantage of past experiences without replicating them. The capability to apply past experiences in novel ways under unfamiliar and changing conditions is, in fact, the central feature of human intelligence largely absent in animals.1 The VAN model has some similarities with a theory presented at the symposium by nonengineer Gerald Edelman, director of the Neurosciences Institute in San Diego and winner of the 1972 Nobel Prize in physiology or medicine. First published in December of 1998,2 the theory, called the "dynamic core hypothesis," is derived from Edelman's earlier theory of "neuronal group selection"3 and is part of his group's continuing efforts to ascertain the roots of human consciousness. It contends that the key to elucidating the poorly understood phenomenon of human consciousness is focusing on certain properties of the neural process that underlies consciousness, namely differentiation (the number of conscious states is infinite) and integration (these states are intimately connected and cannot be reduced to constitutive parts).4 According to Edelman, underlying consciousness is a distributed neural process that involves major regions of the thalamocortical system. "It isn't any named group of neurons or any particular neuron that has a property that accounts for consciousness," Edelman suggested at the symposium. "Instead, it's ... a particular kind of neuronal ensemble, which we call the dynamic core." Edelman emphasized that according to the hypothesis, templates are not necessary, as in the VAN model. Edelman's theory differs from the VAN in that it's mostly concerned with the brain's biology while the VAN focuses on the functional architecture responsible for complex performance. This approach allows the VAN to model complex capabilities such as logical reasoning and compare human performance--in, for example, a chess game--to that predicted by the model. Engineering-oriented top-down approaches such as Yufik's, as well as those such as Edelman's with a more biological bent, may increasingly supplement basic-science, bottom-up approaches in the future. Says Fogel, "I hope that the top-down approach will couple with the bottom-up approach and benefit both." Eugene Russo can be contacted at erusso@the-scientist.com. Y.M. Yufik, "Virtual associative networks: a framework for cognitive modeling," In: Pribram, K.H., ed., Brain and Values, Mahwah,N.J., Lawrence Erlbaum Associates, 1998, pages 109-79. G. Tononi, G.M. Edelman, "Consciousness and complexity," Science, 282:1846-51, Dec. 4, 1998. G.M. Edelman, The Remembered Present. A Biological Theory of Consciousness, New York, Basic Books, 1989. E. Russo, "Consciousness studies: birth of an empirical discipline?" The Scientist, 13[10]:14, May 10, 1999.

Symposium Explores Paradigm Shifts in Cognitive Modeling

Interested in reading more?

Become a Member of