Flight Control by Neural Nets: A Challenge to Government/Industry/Academia.
International Conference on Artificial Neural Networks
Paris, FRANCE
(Ed. F. Fogelman)
Flight Control by Neural Nets: A Challenge to Government/Industry/Academia Andra´s J. Pellionisz, Director Silicon Valley Net Institute 1030 E. El Camino Real 393 Sunnyvale, California 94087, USA director@svni.siliconvalley.com www.siliconvalley.com ftp.siliconvalley.com Chapter 1. Flight Control by Neural Nets: Challenges 1.1An ultimate challenge to neurocomputing: Flight control Humans have eternally admired birds. They can fly - we can't. Moreover, their brain, supposedly inferior to ours, enables their awkward bodies, not primarily made by evolution for aviation, with an astounding flying ability. A hummingbird's dynamic flight is controlled by a brain no bigger than a speckle. Flight is generated by the constantly changing body geometry, speed of reconfiguration surpassing the velocity of signal propagation along nerves! This tiny marvel of nature, using nothing else but neural nets for flight control, provides aerospace engineers with a depressing reminder of the limits of "leading edge technology" of this outgoing century. It also sets a daring challenge for the next millennium, however [1]. We already build airframes designed not for flight-efficiency but for radar-invisibility (stealth). Accordingly, they cannot fly without an on-board computer for flight control - to constantly adjust geometrical configuration, achieving dynamic stability rather than the static stability of old-fashioned airframes with fixed body geometry. The challenge of "flight control by neurochips" has hardly been answered - some countries are not even sure to have accepted it yet. However, some hesitant positioning is already noticeable towards the awesome and intricate challenge for which "existence proofs" we see flying around us every day. 1.2. A challenge to defense organization (Government/Industry ) Defense implications of superior flight control are too obvious to belabor. Given that defense is perhaps the most important function of government, "neuroengineering applied for flight control" emerged over the past half a dozen years an issue to be considered by governments and key R&D agencies (NASA & NSF in USA, DLR in Germany, MITI in Japan, etc). It also increasingly concerns highly competitive aerospace industries in the USA, that are closely tied to the government by defense contracts (Boeing, McDonnell-Douglas, Lockheed, etc). However, the past half a dozen years also represent a shaky transition after the intrinsically stable Cold War, towards a multilateral industrial competition on an unprecedented global scale and fierce intensity. Not so long ago, knee-jerk reflexes of Cold War caused strategic issues of R&D (including submarine voice-pattern recognition by neural nets) to be generously nurtured in seclusion of suitable institutions, with defense-agencies supportively looking on (ARPA). Today's very different times, however, call for pathbreaking solutions (see later in the paper), also opening a (transient, perhaps) "window of opportunity"; based on private enterprise. 1.3. Neurobiological challenge (Academia) "Fight control by neural nets" poses a challenge to Academia as well, as it specifically calls for an integration of branches of basic research; wet (experimental) and dry (theoretical) neurocomputing. It is evident, in retrospect, that nuclear technology could not have been developed without an integration (by a homogeneous set of physicists) of experimental and theoretical nuclear physics. With brain-like machines, however, responsibility for technology development rests with engineers, while providing understanding of the biological brain is the task for an entirely different community; neuroscientists. 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 tens of billions of research dollars. If we fail to exploit this wealth of knowledge, "dry" engineers could easily turn out a very sterile and limited technology, while "wet" neuroscientists could end not even being able to justify which animal experimentation (if any) is truly necessitated by brain theory. The history of Cybernetics and Artificial Intelligence shows how overspecialization and ignorance of the biological brain can be a self-defeating long term strategy. Integration of "wet" and "dry" neural network activity would best start at the basic research level, in Academia. To bring together animal experimentation with mathematical theory and engineering applications is extremely difficult at Universities, however. Not only because of major differences between neuroscience and neural net implementation of mathematical formalism. Culture, funding structure and working philosophy are also different, but most important is the simple structural obstacle that wet and dry activities fall into Medical School and School of Engineering Faculties, respectively. Not all Universities have both schools, and when they do, Faculties are then often far apart (NYU, Cornell). Nonetheless, as it is underscored by management [2] of this third wave (Neural Nets, after Cybernetics and Artificial Intelligence) 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, it should be clear that we do not even have yet a concise mathematical theory of biological neural net function. Worse yet, we do not even have a consensus about the kind of mathematics that is intrinsic to brain function [3]! Because of compartmentalization of University structure (in the USA), this worker sought out NASA Ames Research Center (ARC), in 1990, as the pivotal institution to foster integration of neural and computation sciences [4]. At ARC, in the heart of Silicon Valley, both animal experimentation is conducted (over 90% of NASA's animal experimentation, concentrating on nature's navigation, the so-called vestibular system [5]), but of course hard-core information systems research and its application to flight control are also long-term programmatic efforts [6]. An initial in-depth study for NASA[4], and the"Flight Control by Neural Nets" Symposium last year (organized by Silicon Valley Neurocomputing Institute, dir. A.J.P.) has presented this and related Agencies a blueprint by experts in Government/ Industry/Academia to recommend solutions ([7], see in particular [8]) This Symposium was one of the first actual products of a future "Institute Structure" that is now an established feature of NASA, see later in the paper). Fact remains, however, that because of the major and drawn-out restructuring process of USA Administration and Government, its government agencies (NASA) and reorganization of its Centers, a five year effort has so far been insufficient not only to adapt the NASA/NSF joint interagency plan for neurocomputing, but appeared too short even to fight out an integrated management-structure for animal experimentation and information science; dangerously exposing the Agency to questioning if animal experimentation is necessitated by any hard-core theory. 1.4. Technological challenge of "pushing the envelope": Flight control by neural nets" is a daunting task for Aerospace Industry The "Symposium for Flight Control" at NASA ARC [7] has evidenced that the most important challenge is comprised into a single word: "reconfiguration". Just as it is obvious in natural flight, neurocomputing is of the essence whenever the airframe has a dynamically changing geometrical configuration. This should happen both by design (to keep an aerodynamically unstable stealth aircraft, flying by constant dynamic adjustments of control surfaces) and/or it could also occur if body geometry is rapidly altered by sustained damage and the task is "fault tolerance" under fire, on a massive scale. This worker 's "geometrical approach" to neural net [9-10] has long concluded in a Patent ([11] presently protected only in the USA) that established an industrial foundation of a neural net theory that is directly aimed at the solution the classical problem in biological flight control. Basic tenets are that the geometry (space-time metric tensor) of the body and its dynamic environment is brought into homeomorphism by the metaorganization of a neural geometry (in a vestibulo-cerebellar neural net controller) [10]. The tensor-approach is one of the few neurobiologically proven neural net theories [12]. It is also one that addresses the problem of "reconfiguration" in a mathematically advanced manner, with geometry established as the emerging theme of neurocomputing [13], and the tensorial formalism is now emulated even by its early opponent [14]. Most important, the approach is also technologically applicable in the defense-oriented aerospace environment of NASA [1,4,7]. R&D challenge to industry is vastly complicated, however, by the competitive stance of its major players (Boeing, McDonnell-Douglas, Lockheed). Their standoff is difficult to bridge even by government initiatives such as a NASA-sponsored neural net R&D projects. If such initiative is established with one giant, government is vulnerable to criticism for not being even-handed, if identical or overlapping initiatives are established with several or all industrial players, government project becomes both redundant and a conduit for leaks among competitors. Chapter 2 Neural Net R&D: Answers to Challenges 2.1 Institute structure: Incubated at NASA ARC NASA ARC has increasingly became aware of the manifold structural problems referred to above, and "Silicon Valley Neurocomputing Institute", established by this worker in 1993 on NASA sponsorship was incubated as a solution. First, such "Government/Academia/ Industry integration was based in Academia, but for limitations of the hosting University and competitive frictions among Silicon Valley Universities the Institute became independent in early 1995. It was renamed "Silicon Valley Net Institute". This reflects the phenomenon that the "Information Technology Revolution in Silicon Valley" concentrates also on another "net" (Internet) which in fact is to be greatly improved by "neural net" technologies. The contribution is (also in Nature!) to provide with "information on-ramps" and advanced information processing such as fractal image compression by neural nets, voice and speech recognition, etc [6]) Although the "institute structure" has now became an officially designated formula for advanced planning at NASA ARC, full establishment of its proposed consortium with Stanford/U.C. System is likely to take very considerable time and effort to unfold. 2.2 Domestic and international cooperation in Private Enterprise Meanwhile, many tasks can be immediately addressed in the framework of a private-domain Institute, fully utilizing Silicon Valley physical location combined with full-fledged global Internet- presence. International cooperations, for instance, which are inordinately complex under the aegis of government, can be seeded in the private sector that is free of the sometimes cumbersome government protocol. Even with domestic Universities, for instance, exclusion-policies could prevent them from cooperating, forcing their competitive alienation. Such problems and solutions surfaced at the "Symposium on Neural Nets for Flight Control", where there were many very good, relatively risk-free ANN projects presented. These projects are appropriate for the mission offices that focus on relatively isolated applications [e.g. 15]. Until a NASA leadership is developed, where a government agency leads in developing more advanced designs than off-the-shelf NN software-applications, institutions in the private domain can provide a unifying linkage, tilted towards the biggest-payoff projects. 2.3. "Institute Structure" as a long-term solution Success of an Institute-structure also requires a network of properly trained domestic and international students, to work with relevant companies and agencies. (This worker, as a Humboldt-Prize awardee, is eligible to receive postdocs from Germany supported by the Feodor Lynen Fellowship). For a long-term solution, university basic research efforts in Silicon Valley are to be tied into the effort at NASA ARC, with a consortium of universities in the making. One helpful mechanism to even better accomplish this would be a joint NASA-NSF initiative [4], fashioned after the many existing DOD-NSF joint initiatives. NSF management supports interagency collaborations, especially when they tend to set up a network tying together undergraduate education and basic research through to measurable impacts on national strategic priorities. NSF has several cross-Directorate initiatives to fund engineering-neuroscience collaborations. The neuroscience community expressed serious interest in working together with engineers; the challenge to engineers is to develop the institutions for a more serious approach to neural net solutions seen in the biological brain. A NASA-NSF joint initiative, seeded by an Institute-agenda, could naturally tie in with these existing initiatives, and bolster their engineering content, if NASA should to be a stable long-term player. 2.4. Institute-structure as an "interim reservoir" Uncertain as global politics are, a relapse can occur at any time, even if not to Cold War but to Cold Peace (either with Russia or China or both, possibly even with Japan). As a result, USA and Europe could possibly opt in the future for a tighter cooperation in defense-related industrial R&D where no single geopolitical region would wish to "go it alone". Future is cloudy if and when such happening might occur. Until more settled trends, private domain "Institute-structure" could serve as a hedge, and also an "interim reservoir", where communication, exchange, cooperation and mutual investment is put to use immediately, without undue waste and frustrations in neural networks R&D. *** References [1] Pellionisz, AJ, Jorgensen, C (1992) Cerebellar Neurocontroller Project, for Aerospace Applications, in a Civilian Neurocomputing Initiative in the "Decade of the Brain" IEEE Proc. of IJCNN 1992, III- 379-384 [2] Werbos, P. J. and A. J. Pellionisz. Neurocontrol and Neurobiology: New Developments and Connections. IJCNN 1992 Baltimore. 1992 [3] Anderson J, Pellionisz A,Rosenfeld E. (1990). Neurocomputing-2: Directions of Research. Cambridge, MA: MIT Press [4] Pellionisz, A. U.S. Civilian Neurocomputing in the Decade of the Brain. A NASA-NIH-NSF Initiative. Report commissioned by NIH- NASA, 1990. [5] Pellionisz, A., B. Peterson and D. L. Tomko. "Vestibular Head-Eye Coordination: A Geometrical Sensorimotor Neurocomputer Paradigm" In: "Advanced Neuro-computing". Eckmiller ed. 1990 Elsevier, North-Holland. Amsterdam. pp. 126-145 [6] Pellionisz, A.J. (1994) From Geometrical Foundations of NN Research to Lead-Roles in Silicon Valley Information Industry in Flight Control and Infohighway Interface. Invited Lecture, Proc. of. Korean International Neural Network Conf. [7] Jorgensen, C. and Pellionisz AJ. (eds) (1995). Flight Control by Neural Nets. (Proc. of a 3-day Government/Academia/Industry Symposium, org. by Silicon Valley Neurocomputing Institute, dir. A. Pellionisz, held at NASA Ames Res. Ctr., Aug 18-22, 1994, in prep.) [8] Werbos, P. (1995) Neural Networks for Flight Control: A Strategic and Scientific Assessmen. In: (Proc. of a 3- dayGovernment/Academia/Industry Symposium, organized by Silicon Valley Neurocomputing Institute, dir. A. Pellionisz, meeting held at NASA Ames Res. Ctr., aug 18-22, 1994, in prep.) [9] Pellionisz, A. and R. Llina´s. Tensorial approach to the geometry of brain function: Cerebellar coordination via metric tensor. Neurosci. 5: 1125-1136, 1980. [10] Pellionisz A, Llina´s R. (1985). Tensor Network Theory of the metaorganization of functional geometries in the CNS. Neurosci.; 16:245-274 [11] Pellionisz, A. (1984) Sensorimotor coordinator: United States Patent, # 4,450,530. [12] 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. [13] 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. [14] Amari, S. Dualistic Geometry of the Manifold of Higher-Order Neurons. Neural Networks. 4(4): 443-451, 1991. [15] Adkins, M, Cox, C, Pap R.., Thomas C and Saeks, R (1992) Neural joint control for space station robotic manipulator system, Proc. of 1992 IEEE/RSJ Int. Conf. on Intelligent Robots.