IJCNN INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS
NEUROCONTROL AND NEUROBIOLOGY WITH AEROSPACE APPLICATIONS
SPECIAL SESSION PRESENTATIONS
IEEE Catalog Number: 92CH3114-6 ISBN: Softbound Edition 0-7803-0559-0
Casebound Edition 0-7803-0560-4 Microfiche Edition 0-7803-0561-2 Library of Congress Catalog No 91-59048
*This paper gives personal views, not official positions of NSF of NASA. All ideas expressed are within the public domain.
Andras J. Pellionisz
Charles C. Jorgensen
NASA Ames Research Center, Rm 261-3*,
Moffett Field, CA 94035
and
Paul J. Werbos
National Science Foundation, Rm 1151*,
Washington, D.C. 20550
ABSTRACT
Neurocomputing has entered a critical period in its development as a field. The driver is that mathematical brain theory emerging from Computational Neuroscience needs to be incorporated with algorithm development and implementation characteristic of neuroengineering. A key question involves how to utilize civilian government agencies along with an industrial consortium so as to successfully complement the so far primarily defense-oriented neural network research. Civilian Artificial Neural System projects, such as Artificial (Vestibulo)-Cerebellar Neurocontroller's, aimed at duplicating Nature's existing neural network solutions for adaptive sensorimotor coordination, are proposed for such a synthesis. The cerebellum provides an intelligent interface between higher possibly symbolic levels of human intelligence and repetitious demands of real world control problems, where complexity, nonlinearity and noise are far beyond the capacity of conventional controllers. The generation of such intelligent interfaces could be crucial to the economic feasibility of the human settlement of space [1] which will require new control technologies in transportation [28] and an improvement in telerobotics techniques so as to permit the cost-effective exploitation of nonterrestrial materials and planetary exploration and monitoring. We propose a scientific framework within which such interagency activities could effectively cooperate.
A PERSPECTIVE ON NEUROCOMPUTER INDUSTRY: MUST SCIENCE PRECEDE TECHNOLOGY?
Half a century ago, the Nuclear Industry was founded on a scientific useful fact that nuclear energy is released during fission or fusion. A sequence of developmental steps led to this result. First, basic research (Nuclear Physics) was required; to understand mathematical principles underlying the new phenomena (quantum mechanics). Second, technology development was needed, such that well-understood phenomena could be controlled and exploited. Existing research efforts in neurocomputing attempt to take a similar path. First, they support basic research to develop theories of the brain, in order to scientifically understand the underlying mathematical principles. Then they try to generate application efforts to translate these principles directly into artificial neural networks. In practice, this approach has not been the most fruitful way to develop the necessary mathematics. In reaction some engineers have gone to an extreme of developing ANNs without any biological inspiration. However, real brains demonstrate capabilities far beyond what conventional controllers have yet been able to capture. Nonetheless even early attempts to imitate biological principles have led to remarkable success in applications, c.f [28]
It is evident in retrospect that even nuclear technology could not have been developed by physicists without first establishing a scientific foundation i.e. nuclear physics. With brain like machines, however, responsibility for technology development rests with engineers, while providing understanding of the biological brain is the responsibility of a different community of neuroscientzsts. Since the sixties, the U.S. made an unprecedented effort in experimental neuroscience, which in turn evolved into a thriving research branch and reflected the investment of billions of research dollars. If we fail to exploit this wealth of knowledge, we could easily turn out a very sterile and limited technology. Indeed, the history of Cybernetics and Artificial Intelligence shows how overspecialization and ignorance of the biological brain can be a self defeating long term strategy. It is important to ensure that the proclamation of the "Decade of the Brain", and its passing as U.S. law, is coordinated with ongoing efforts of putting knowledge and understanding of biological neural networks into practical use in advanced computer research and development.
An alternative to the existing separation of "wet" and "dry" neural network research, is to closely tie experimental and theoretical neurobiological efforts to technological development of brain-like machines [22),[23]. This latter course does have disadvantages, too. Establishing such a bond is extremely difficult given major differences between neuroscience and neural net implementation in the extent of mathematical formalism, culture, funding structure and working philosophy. Nonetheless, a synthesis may be the only way to create a unified discipline. While many workers in "neural network" research tend to ignore the issue of integrating neuroscience and mathematical neural net theory, those responsible for this third wave (after Cybernetics and Artificial Intelligence) must make it totally explicit that we do not have a concise mathematical theory of biological neural net function. We do not even have a consensus about the kind of mathematics that is intrinsic to brain function.
BRAIN THEORY: MATHEMATICS OF NEURAL NETS, LEARNED FROM NATURE
Brain theory will not be a mature discipline without the creation of its own mathematics, and this can only be learnt from Nature (in the present case, with the help of Neuroscience). Therefore, efforts towards technology development should be integrated with brain theory development. Towards this goal, a geometrical approach to neurobiology-based neural net theory was initiated more than a decade ago [20]. However, because of the many challenges involved in initiating this approach [5], its full potential was not generally recognized until a decade later. Today, an increasingly wide group of scientists is convinced that geometrical principles of neural net function can be found and experimentally verified by means of neurobiology [20],[10],[13],[16],[23],[9],[21]. Delay was unavoidable since absorption of novel initiatives into traditional approaches is shown by the literature to be an arduous process [24],[26] versus [25],[4], or [5] versus [3]). Whether geometrical brain theory will win out (or one based on other intrinsic mathematics) remains to be seen. One thing seems likely, with no brain theory at all, "neural networks" may evolve into a disconnected subfield of mathematics, control engineering, and computer science.
ARTIFICIAL NEURAL SYSTEMS PROIECTS: CEREBELLAR NEUROCONTROLLER
Once mathematical theories of specific neural systems and accompanying computer models are available, a key questions is how these theories and models will be incorporated into applied technologies. Perhaps the best answer is to very briefly review the modern history of cerebellar theory [18],[17]. Ever since the classical book "Cerebellum as a Neuronal Machine" [8] it was evident that the cerebellum is probably the top candidate for a prototype of a "brain-like machine". Early mathematical "theory" of the cerebellar cortex [15) appeared long before the official ascent of neurocomputing. Two drastically different strategies were formulated for the use of cerebellar theory and modeling. (1) In one approach some theories were used in neuroscience to conglomerate unwieldy bodies of data on the physiology and/or morphology of this neural apparatus [12],[7],[14],[11] and to advance one group of experimental neurobiology over the other. (2) The other approach was to actually put the mathematical theory into use and thus directly check performance of theory. For instance, Marr's "Purkinje cell learning algorithm", even restricted to cerebellar cortex, was put to actual test in a robotic controller [2] even before the onset of the trend towards "neural networks".
Werbos suggested [27], an "Artificial (Vestibulo)Cerebellar Neurocontroller, should have specific R&D goals, that (1) ensure that theories and models are tested not only by their conformity to experimental facts, but also by the actual performance they are capable of, (2) ensure that there is an active dialogue between the engineering community and neuroscientists.
U.S. CIVILIAN NEUROCOMPUTING INITIATIVE
At the level of R&D organization and management one key question is how to create the framework of civilian government agencies (backed by an industrial consortium) that could successfully complement the so far dominantly defense-oriented neural network research. A historical argument pointing out the necessity of a civilian governmental program for R&D of brain-like computers is based on contrasting the organizational structure of earlier stages of development of the computer industry with the new area of "Neural Networks". Classical computers, the serially organized von-Neumann mainframe machines, were developed for a strategical purpose (sufficiently fast calculation of ballistic trajectories), and thus their development was organized and managed essentially by the defense establishment. The most recent stage of major evolution of the computer industry is personal computers. This revolution was driven by small entrepreneurship, entirely in the industrial commercial domain. Almost exclusively commercial organization of development was made possible by not requiring basic research. Neurocomputer development appears to spontaneously be following these two past successful trends. The present organizational structure of neurocomputing rests on two pillars; on one hand on a defense-oriented program, and on another support drawn from small commercial startup companies.
This dual structure of neurocomputing is inherently faulty, since neither will support neuroscience-inspired basic research without which sustained healthy growth is doubtful. Therefore, it was argued in an interagency-report [19] that establishment of a third civilian governmental component of neurocomputing should take care of the needs of basic research and should contribute to the strengths of the existing two pillars. Defense programs will not be able to support the neuroscience-mathematical theory-technology connection, necessary for the scientific maturation of this interdisciplinary field. Likewise, the commercial market for neurocomputing will have great difficulties in sustaining the "neurocomputer revolution" if developments begin to require drawing heavily on basic neurobiology research. The present structure of neurocomputing leaves therefore a gap between R&D and marketing of technology on one hand, and development of the underlying science based on neuroscience. It is not that the funding structure for strengthening the neurobiological knowledge-base does not exist within agencies such as NIH, NSF and NASA; the problem is the lack of structural connection between knowledge and its utilization. A coherent civilian government program generated under the auspices of US Law "Decade of the Brain" could be most effective towards this goal, especially if industrial partners (most particularly aerospace companies) create an "Industrial Consortium" for neurocomputing, acting as a natural partner for the government program.
NASA lNVOLVEMENT: AEROSPACE APPLICATIONS OF ADAPTIVE SENSORIMOTOR CONTROL
In the National Space Society analysis of settling space [1] teleoperation is listed as one of the three chief bottlenecks which must be overcome before large-scale economical human settlement of space is possible. The National Space Council has taken major initiatives in recent years to urge NASA for long-term activities of this sort, but cost has been a primary obstacle. Improved teleoperation could solve this cost problem, while still keeping man in the loop. How we arrive at an effective interface between a higher intelligence and a physical motor system may be to imitate the neurocomputer-interface that nature developed for precisely this role, the cerebellum. The same approach may have application in other areas, like NASP control, where a human is still in the loop, though it is too early to know which part of human system will in fact prove most useful in a fast, electronic implementation. As a result a fairly broad spectrum of alternative neural control approaches need to be compared and contrasted with the biological methods. Given our ignorance, prudence suggests a multipronged effort, which does at least include an attempt to imitate the one natural system which does perform the task at hand in nature, the cerebellum. An Artificial Cerebellar Neurocontroller could be one of the methods used for aerospace applications where error-tolerance, gracefully degrading architecture, and massively parallel computation are required.
The primary scientific justification of an "Artificial Cerebellar Neurocontroller" project is that the vestibulo-cerebellar neuronal nets offer among the best and clearest examples of an actually performing neurocomputer. It is known that during evolution nature developed this system for maintaining stable position, posture and movement of bodies such that they can perform coordinated fast action in a turbulent environment. Accordingly, the vestibulo-cerebellum occupies an outstandingly high proportion of the brain in birds. Thus, they are masters of flying; rapidly changing their body-geometry to adapt to turbulent conditions. Flight is controlled by an "on-board, real-time neurocomputer" that relies on error-tolerant, gracefully degrading, massively parallel neural network. In addition, it need not rely on supercomplex software that characterizes, and causes most of the breakdowns, of present-day serial computer systems.
The Cerebellar Neurocontroller" provides one good example where research-oriented cooperation between NIH(NIMH)-NSF and NASA can occur. Fast and precise ilight control is fully in line with the purposes of NASA and an Artificial Vestibulo-Cerebellum prototype is to quickly find its way to specific applications in novel flight control systems. While learning nature's neurocomputer secrets for flight control might not yield necessarily better solutions than those provided by traditional engineering approaches, it appears prudent to hedge classical engineering solutions especially since they seem to have reached certain limits, where sheer redundancy of computer systems is too heavy and complexity of software leads to verification difficulties, as well as some intricate theoretical problems, such as sensory fusion, remain a problem although nature evidently resolved them successfully.
As the Chief Scientist of NASA Ames remarked [6]: "Approving flight-critical hardware usually involves meeting a seemingly draconian requirement: the likelihood that all the computers will fail should be no more than one chance in a billion hours of flying...There have been fewer than 20 million hours since the birth of Christ...A comparable requirement for software [verification] does not exist... if you know you really need 10 to the minus nine reliability in software...then you have to worry about what's meant by safety". The classical definition of safety is zero error. The "neural network" definition of safety is "acceptable degree of graceful degradation". It is of some importance that the classical safety techniques "guarantee" safety only under highly unrealistic assumptions, while neural nets may often offer weaker-looking "graceful degradation" which, however, is a more realistic and valid "guarantee".
Government support (most particularly by NIH-NIMH) has already resulted in vast knowledge and some mathematical understanding of biological neural networks. Likewise, civilian neuro-engineering programs (most particularly by NSF) are also in place to provide the technology needed for implementation of artificial neural nets. Specific applicationoriented NASA programs require that we determine the most promising international initiatives that can integrate the hitherto separate civilian efforts. This could best begin by R&D projects, such as aiming at an "Artificial Cerebellar Neurocontroller, under the auspices of an interdisciplinary and interagency program by NASA-NSF-NIH-NIMH.
REFERENCES
1. "Settling Space." America's Future in Space: a Briefing Book. 1989 National Space Society. Washington, D.C.
2. Albus, J. A model of the brain for robot control. Byte (June): 10-34, 1979.
3. Amari, S. Dualistic Geometry of the Manifold of Higher-Order Neurons. Neural Networks. 4(4): 443-451, 1991.
4. Anastasio, T. J. and D. A. Robinson. Distributed parallel processing in the vestibulo-oculomotor system. Neural Computation 1(2): 230-241, 1989.
5. Arbib, M. and S. I. Amari. Sensorimotor transformations in the brain with a critique of tensor network theory of the cerebellum. J. Theor. Biol. 112: 123-155, 1985.
6. Billings, C. and G. Stix. Trends in Transportation. Scientific American (July): 95-106, 1991.
7 . Eccles, J. An instruction-selection theory of learning in the cerebellar cortex. Brain Res. 127: 327-352, 1977.
8. Eccles, J. C., M. Ito and J. Szentágothai. "The cerebellum as a neuronal machine." 1967 Springer-Verlag. Berlin, New York, Heidelberg.
9. Eckmiller, R. "Concerning the emerging role of geometry in neuroinformatics.In: "Parallel Processing in Neural Systems and Computers. Eckmiller, Hartmann and Hauske eds. 1990 Elsevier Science Publ. B.V (North Holland). Amsterdam.pp. 5-8.
10. Gielen, C. C. A. M. and E. J. van Zuylen. Coordination of arm muscles during flexion and supination: Application of the tensor analysis approach. Neuroscience. 17: 527-539, 1986.
11. Houk, J. C. Model of the cerebellum as an array of adjustable pattern generators. Neural Networks. (Suppl. 1): 339, 1988.
12. Ito, M. "The cerebellum and neural control." 1984 Raven Press. New York. 13. Koenderink, J. J. Geometrical structures determined by the functional order in nervous nets. Biol. Cybern. 50: 43-50, 1984.
14. Llinas, R. R. The cortex of the cerebellum. Scientific American 232(1): 56-71, 1975. 15. Marr, D. A theory of cerebellar cortex. J. Physiol.(Lond.) 202: 437-470, 1969.
16. Pellionisz, A. "Tensorial aspects of the multidimensional approach to the vestibulooculomotor reflex and gaze." Adaptive Mechanisms in Gaze Control. Berthoz and Melvill-Jones ed. 1985 Elsevier. Amsterdam. pp.281-296
17. Pellionisz, A. Tensorial brain theory in cerebellar modeling. In: "Cerebellar Functions". Bloedel, J., Dichgans, J and Precht, W. eds. 201-229, 1985.
18. Pellionisz, A. "David Marr's theory of the cerebellar cortex: A model in brain theory for the "Galilean combination of simplification, unification and mathematization".In: "Brain Theory". G. and Aertsen eds. 1986 Springer Verlag. Berlin-Heidelberg-New York. pp. 253-257.
19. Pellionisz, A. U.S. Civilian Neurocomputing in the Decade of the Brain. A NASAMH-NSF Initiative. Report commissioned by NIH-NASA, 1990.
20. Pellionisz, A. and R. Llinas. Tensorial approach to the geometry of brain function: Cerebellar coordination via metric tensor. Neurosci. 5: 1125-1136, 1980.
21. Pellionisz, A. J. and J. R. Bloedel. Functional Geometry of Purkinje Cell Population Responses as Revealed by Neurocomputer Analysis of Multi-Unit Recordings. Soc. Neurosci Absts. 21: 920, 1991.
22. Pellionisz, A. J., B. W. Peterson and D. L. Tomko. "Vestibular Head-Eye Coordination: A Geometrical Sensorimotor Neurocom uter Paradigm" In; "Advanced Neurocomputing". Eckmiller ed. 1990 Elsevier, Norch Holland. Amsterdam. pp. 126145
23. Peterson, B. W. Modeling of Spatial Transformations in Vestibular Reflex Systems. International Joint Conference for Neural Networks. l: 152-156, 1990.
24. Robinson, D. The use of control systems analysis in the neurophysiology of eye movements. Ann Rev Neurosci. 4: 463-503, 1981.
25. Robinson, D. The coordinates of neurons in the vestibulo-ocular reflex. Adaptive Mechanisms in Gaze Control Facts and Theories. Berthoz, A. and Melvill-Jones G. eds. 1985. Elsevier. Amsterdam, pp. 297-311.
26. Robinson, D. A. The use of matrices in analyzing the three-dimensional behavior of the vestibulo-ocular reflex. Biol. Cybern. 46: 53-66, 1982.
27. Werbos, P. J. Backpropagation and neurocontrol: a review and prospectus. IJCNN. I-209, 1989.
28 Werbos, P. J. and A. J. Pellionisz. Neurocontrol and Neurobiology: New Developments and Connections. IJCNN 1992 Baltimore. 1992.
***
III-384