BIO-CONTROL BY NEURAL NETWORKS Summary of a Workshop supported by the National Science Foundation George A. Bekey Computer Scince Department Uniersity of Southern California and Peter G. Katna Program Director Bioengineering and Aiding the Disabled National Science Foundation Alexandria, Virginia May 16-18, 1990 Participating NSF Programs: Behavioral and Neural Sciences Bioengineering and Aiding the Disabled Engineering Systems Neuroengineering TABLE OF CONTENTS I. Introduction II. Workshop Agenda III. Summary of Presentations IV. Summary of Recommendations V. List of Attendees VI. References I. INTRODUCTION In the view of a number of investigators, there is an increasing dichotomy between engineering research in artificial neural networks and physiological research on neural control mechanisms. In order to determine the state of the art in both the biological and engineering view of bio-control by neural networks, to isolate the major difficulties that hinder communication and block progress in the field and to identify those areas where focused research might be most beneficial, NSF sponsored a small invitational workshop. The specific goals of the workshop were as follows: 1. To determine the state of the art in control of physiological systems by neural networks. How mature is this field? Can current models yield any insight into the structure and function of living control systems, or should they be viewed as input-output models, with little or no isomorphism to the nervous system? 2. To determine whether artificial neural networks, intended to mimic natural control systems, can be used to control systems that include biological components. Are we ready to design control systems that draw upon our knowledge of how natural systems behave? 3. To identify major difficulties that block progress in this field. Are the difficulties conceptual or experimental? Do we lack mathematical, computational, or experimental tools? Are there fundamental gaps in knowledge which hinder further application of artificial neural nets to living systems, either for model-building or for artificial control systes? The workshop was held on May 16-18, 1990 in Alexandria, Virgini. The 32 participants included six NSFprogram directors, two representatives from NIH, and 24 neural network researchers from both the biological and engineering communities. The conference was chaired by Dr. George Bekey, and sponsored by thesday, May 17, 1990 8:30 am Introductions and Presentation of Workshop Goals George Bekey, University of Southern California, Conference Chairman NSF Program Directors: Peter Katona Lazzaro, California Institute of Technology Chi-Sang Poon, Massachusetts Institute of Technology 12:00 pm Lunch 1:30 pm Process Control by Neural Networks Lyle Ungar, University of Pennsylvania T. J. McAvoy Grillner, Karolinska Institute 11:15 am Methodology and Trends in Modeling Herb Rauch, Lockheed 12:00 pm Lunch 1:00 pm Grup Discussions 2:30 pm Presentatons from Groups; Summary of Recommendations 4:00 pm Adjourpper and midde layers of the frog's spinal ord (while the leg is placed in differnt positions) generated a force field with an equilibrium point. The implications of this field on the organization of the spinal cord were disassachusetts) are using these ideas for the design of a new model of cerebellar function. [3] Issues involving the neural control of locomotion were also discussed by Hillel Chiel and Sten Grillner. Hillel Chiel (Biolog For example, some of the model neurons showed rhythmic bursts of activity ("pacemaker neurons") which were modulated by input from other model neurons. In addition, the architecture of the neural net controlling locomotion was ynaptic connections was capable of exhibiting surprisingly complex behavior patterns. [7] Sten Grillner (Nobel Institute for Neurophyiology, Stockholm) Locomotion Control in the Swimming Eel Thelocomotor control s that without simulation, it was not possible to evaluateif the experimentally established network could account for the known locomotor behavior in terms f segmental and intersegmental coordination. [8] [9] [10] Te autohis system, presented by Wade Rogers (DuPont Neural Computation roup), the vagal baroceptor reflex has also been modeled in VLSI by John Lazzaro (Department of Electrical Computer Engineering, University of Colorado- Boulder). . Feldman then discussed certain aspects of the control of respiration, primarily the generation of respiratory rhythms and the importance of various properties of the neurons involved in these systems. Distributed networks of coupled model of the respiratory control system in which the input-output relationship of the brainstem respiratory controller was governed by an optimality criterion. The latter measured both deviation from steady state values of arte the cerebral cortex, which served as a "proxy" of the brainstem neural network. [15] The results suggested that such compound optimization behavior was quite feasible within the CNS, both at the level of the brain stem and higher br neural nets in both feedforward and feedback control, inverse model adaptive control and other control algorithms were discussed. [17] [18] [19] Andrew Barto (Computer Science Department, Univ. of Mass.) On Compute views on some of the important research issues in the field of modeling of neural networks. These included questions on: (1) convergence properties of networks, (2) heuristic architectures for specific tasks, (3) adaptive archIV. RECOMMENDATIONS Much of the work of the workshop was accomplished in three subgroups which met following the major presentations. The groups first discussed the need for new biological data in engineering models of neural networks, as weligator support. b) Post-doctoral/sabbatical support could be used to place biologists in engineering labs and vice versa; perhaps these could be supported as supplements to existing projects. 2.are needed for artificial neural networks: Model neurons should capture more of the richness of behavior patterns seen in biological experiments than the simple weighted-summer-with-sigmoid-nonlinearity thaccount for emergent behavior patterns as those found in living sysems (e.g.: sensory-motor interactions, plant-controller interctions, distributed control paradigms). c) Improved mehods for idenof new engineering adaptive control systems based onbiological prototypes should be pursued: Enhancing living systems, e.g., prosthetics. Chemical process control, control of bioreactors. 3. Ways inrding electrodes. Muscle-type actuators. Better motion monitoring equipment; tendon and contact force gauge implants and joint-angle monitoring implantstems methodologies are needed: System concepts; ssteresis. System level hypotheses to direct experiments. V. LIST OF ATTENDEES Dr. Panos J. Antsaklis Department of Electrical and Computer Engineering Universityy Computer Science Department University of Southern California Los Angeles, CA 90089 (213) 740-4501 (213) 740-7285 (FAX) Dr. Emilio Bizzi Department of Brain & Cognitive Sciences E25-526 Maic Institute San Luis Obispo, CA 93407 (805) 756-2131 Dr. Daniel Bullock Center for Adaptive Systems Boston University 11 Cunnington Street Boston, MA 02215 (617) 353-9486 (617) 353-2more, MD 21205 (301) 955-8334 (301) 955-3623 (FAX) Dr. Sten Grillner Karolinska Institute The Nobel Institute for Neurophysiology Box 60400, S-104 Stockholm, Sweden 011-46-8-336059 011-46-8- Department of Physiology Ward Building 5-319 Northwestern University Medical School 303 E Chicago Avenue Chicago, IL 60611 (312) 503-8219 (312) 503-5101 (FAX) Dr. Peter Katona Bioengineering Cambridge, MA 02439 (617) 253-5769 (617) 253-8000 (FAX) Dr. Thomas McAvoy Department of Chemical Engineering University of Maryland College Park, MD 20742 (301) 454-2432 (301) 454-0855 (FAX) 8-5405 (617) 253-2514 Dr. Herb Rauch Palo Alto Research Lab Lockheed 92-20/254E 3251 Hanover Street Palo Alto, CA 94304 (415) 424-2704 (415) 424-2662 (FAX) Dr. David A. Robinson Rootn, DE 19880-0352 (302) 695-7136 (302) 695-9631 (FAX) Dr. Robert J. Sclabassi Department of Neurosurgery Universiy of Pittsburgh Pittsburgh, PA 15213 (412) 692-5093 (412) 692-5287 (FAX) tion Rom 1151, ECS/ENG 1800 G Street, N.W. Washington, DC 20550 (202) 357-9618 VI. REFERENCES 1. Massone, L., and Bizzi, E., "A neural network model for the cerebellum," Neural Networks for Control, Chapter 15, W.T. Miller, R.S. Sutton and P. J Werbos, (EdD., "A lesion study of a heterogenous artificial neural network for hexapod locomotion," Proc. IJCNN, I:n in bipeds, tetrapods and fish," The Handbook of Physiology, Sec. 1, Vol. 2: The Nervous System, Motor Control, pp. 1179- 1236, V.B. Brooks, (Ed.), Maryland: Waverly Press, 1981. 9. Matsushima, T. and GrillneIT Press, Chap: Silicon Ba receptors modeling cardiovascular pressure transduction in ANALOG VLSI, Lazarro, John, Schwaber, James and Rogers, Wade. 12. Schwaber, J.S., Paton, J.F., Spyer, K.M., and Rogers, 9, 1987. 15. Poon, C.S. and Younes, M., "Optimization on, C.S., "Adaptive neural network that subserves optimal homeostatic control of breathing," (submitted). 17. McAvoy, T.J., "Modeling chemical process systems via Networks for Control, T. Miller, R.S. Sutton, and P.J. Werbos (Eds), Cambridge, MIT Press, 1990. 21. Iberall, T., Liu, H., and Bekey, G.A., "Building a generic architecture for robot hand control," IEEE es during trajectory formation," Psychological Review, 95, pp. 49-90, 1988. 24. Bullock, D. and Grossberg, S., "Spinal network computations enable independet control of muscle length and joint compliance," Adand Suzuki, R., "A hierarchical neural-network model for control and learning of voluntary movement," Biological Cybernetics, 57, pp. 169- 185, 1987. 28 Massone, L. and Bizzi, E., "On the role of input