1.0 INTRODUCTION

1.1 Goals and Objectives of this Report

As proposed originally, the goals and objectives of this report are to: (1) describe existing program elements within NASA-ARC and the NIH-NIMH which are relevant to a potential jointly sponsored NASA- NIH Neurocomputing Program (Section 2); (2) discuss potential advantages and disadvantages of such a joint venture (Section 3); and (3) provide conclusions and recommendations to the two organi zations (Sections 4&5). In Appendix, the author's credentials in this area are outlined, background information is given on some of the historical aspects of neurocomputing and its importance and world wide status, followed by references.

1.2 The "Decade of the Brain": Neurocomputing is a Pivotal Challenge for
Neuroscience and Life Science

On July 25, 1989, the President of the United States signed into law House Joint Resolution #174 declaring the 1990s the "Decade of the Brain". The joint resolution signalled a new federal government commitment to neuroscience research. In the spirit of this law, two federal Agencies (NIH-NIMH and NASA) jointly sponsored the present study of whether and how these agencies, specifically NASA Ames Research Center (NASA-ARC) and the NIH National Institute of Mental Health (NIH-NIMH) might benefit from a cooperative program in neuroscience with particular attention to neurocomputing.

Neuroscience, the study of the anatomy, physiology and chemistry of brain and behavior, is evolving from an almost exclusively experimental interdisciplinary science into a discipline with a solid theoretical-mathematical framework. The central theme of the present report is that neurocomputing may be profitably used as a unifying theme in this evolution of Life and Computer Sciences concerned with the brain. Because most studies and approaches to nature evolve towards mathematical or computational models, mathematical and theoretical constructs of brain function provide a point of convergence in neuroscience that can be followed by applications. Neurocomputing is therefore a vital link between knowledge of the nervous system generated by basic research, and utilization of our mathematical understanding of brain function. However, it is also a major challenge to scientific management since although technology cannot be safely and economically developed without understanding of brain function essential mechanisms to integrate neural science and neurocomputer technology are yet to be created. Initiatives for neurocomputing often fall either entirely into the realm of technology or solely of neuroscience leaving a potentially fatal gap until strategies and measures are implemented for systemic integration.

There are currently no formal federal programs that combine unique resources from more than one agency to form a focal point for neurocomputing. Consequently it is suggested by this report that it would be advantageous for NASA and NIH, together with NSF, each with specific areas of interest and expertise, to cooperate in an initiative to coordinate their activities in the field of neurocomputing. Such cooperation would yield more than specific immediate advantages for these agencies. It would also be a suitable platform for a civilian neurocomputing initiative. Such a program is necessary to integrate neuroscience efforts into neurocomputing research and to create an organizational structure in the US that is most competitive with neurocomputer research and development worldwide.

This conclusion is reached after a long period of preparation. Substantial investment in experimental neuroscience since the nineteen sixties, predominantly by NIH, has lead to development of a large body of data on the anatomy, physiology and chemistry of the brain and behavior. Descriptive analyses have accompanied specific findings, but neuroscience should formally incorporate integration of function. Theory must provide both further testable neuroscience hypotheses and mathematical formalisms and concepts for scientists and engineers who make artificial electronic neurocomputers guided by principles of brain organization.

It is becoming accepted that neither pure experimentation nor pure theory is alone sufficient to accomplish the task of proper understanding the brain. Increasingly, mathematical/computational/theoretical neuroscience is a partner- candidate assuring a more rounded, distinguished and fundable research environment. This point is illustrated by a statement of a research director of a newly established Institute of Neurological Sciences at the University of Pennsylvania: "As a molecular neurobiologist, I have lived by the reductionalist credo that understanding how the brain functions requires probing ever deeper into the structural details of its cellular machinery. It was a rude awakening to realize that the integrative analysis of brain function involves concepts for which these molecular details are largely irrelevant. In the same way that rules governing problem solving in a complex computer can be understood without detailed knowledge of each semiconductor junction, the paradigms used by the brain to process information can probably be understood without explicitly defining the molecular architecture of each receptor and ion channel." [1]. Barchi goes on to define "computational neuroscience" as "...those aspects of neural modelling that have their foundations in known facts of neuronal organization rather than the more loosely defined approaches used in the artificial intelligence community; the "how does it really happen" rather than the "how else could it happen" approach..."

---

2.0 Existing Components of a Neurocomputing Program within NIH- NIHM and NASA-ARC

2.1 Mathematical/Theoretical/Computational (MTC) Neuroscience Program of NIH-NIMH

In response to this new trend and in order to support theoretical studies in neuroscience, a new program called Theoretical/Mathematical/Computational Neuroscience was created jointly by NIH-NIMH and NINDS. The program invited research and research training grant applications (individual and institutional) for studies using mathematical, computational, or theoretical approaches to understanding the fundamental mechanisms underlying behavior. The purpose of this program is to place additional emphasis on the use of quantitative tools in solving basic problems in the neurosciences. The program is roughly a year old, received the first batch of applications by February 1st 1989. The program is headed by Dr. Richard Nakamura at NIHM and Dr. Herbert Lansdell at NINDS. Proposals are evaluated by an ad hoc Study Section, and the program operates without a set budget. In theory it has an open access, on a competitive basis with all other grants proposals, to all programs of these agencies. In 1989 no funds were allocated by this program, in 1990 about 6 proposals were funded to date, amounting close to $1M(illion).

This MTC program, although presently minuscle, is of extreme significance for bootstrapping theory into the almost exclusively experimental neuroscience, and is a major qualitative improvement after long progress. Thus, in order to pinpoint how this program could be strengthened, a perspective on this conceptual evolution of neuroscience towards theory will be given below. The trend makes it evident that MTC brain modeling should be upgraded from subserving experiments to partnership with experimental approaches. This is an extremely difficult task of organization, however. Therefore, the initiative of National Institute of Mental Health (NIMH) deserves unparalleled credit for creating a forum for overseeing and sponsoring such research activities. A chief goal of the present report is to derive recommendations based on experience accumulated while working in this field for over two decades. The report is aimed at (a) strengthening the Mathematical/Theoretical/Computational (MTC) Neuroscience program of NIH, and (b) making recommendations how this program could provide the much needed neuroscienceliaison to NASA, leading to (c) a proposed government sponsored US Civilian Neurocomputing Program.

2.1.1 Recommended Frontline Research for MTC Program: Evolution from Phenomenological Modelling to Rigorous and Tested Theory

Difficulties in assessing which fields of neuroscience are most in need of support by the Mathematical/ Theoretical/Computational Neuroscience Program stem from the heterogeneity of levels of approaches. In this segment of the report the basic principle will be behind specific recommendations which area of activity would most cost and timeeffectively coalesce "bits and pieces" of knowledge into broader mathematical theory. Theory should quantitatively unify hitherto disparate approaches, and therefore resources spent on an MTC program should increase coherence and effectiveness of neuroscience research.

1) Single cell modeling and theory of neural coding. The model of a single neuron is a linchpin in understanding neural function. It connects subcellular channel-, membrane-, synaptic levels with net work organization and overall behavioral function. According to its central role, scientific background of single cell Brain Theory and Modeling (brain theory and modeling ) is extensive, as elaborated in several review-books [2], [3], [4], [5], [6]. This author has also contributed with a longer review [7]. Therefore, only select issues are dealt with below.

Single unit models are excellent examples of the contrasting phenomenological versus conceptual modeling. The classical model of the neuron [8] was based on a clear and simple (even simplistic) concept, leading to a classical school of research of neural modeling [9]. Likewise, the conceptually lucid rule [10] that hypothesized synaptic modification according to input fiber activity and firing frequency of the given neuron, became a basis of network as sociation and learning-paradigms [11],[12]. In contrast, a quite dif ferent school was based on phenomenological modeling of membrane events [13]. This provided a more accurate representation of neuronal phenomenology, yet yielded few clues to computational properties of neurons, not to speak of networks. There were successful examples of synthesis; most notably the analytical yet quantitatively elaborate representation of membrane-phenomena of passive dendritic trees [14]. Lastly, at the whole neuron level, membrane equations, single cell morphological and electrophysiological data could be corroborated [15], [16], [7], [17], [18], [19], [20], [21], [22].

For a long period of time, without explicit support for theory, electrophysiology of single cells was virtually the only drive behind neuronal modeling. The enormous dataturnout of this field of research that has been massively supported for decades necessitated such quantitative accounts. However, the trend was to zoom to finer and finer focus and more and more detailed models, which led to an everincreasing number of free parameters. Thus, in compartmental single cell modeling the drive is towards literarily thousands of compartments [20], and to getting down to not only the level of dendrites [23], but to dendritic spines [24],[25],[26] and to channel-models [26],[21]. Modelists, with an eye on synthesis, attempt to close this gap [22]. Nevertheless, the inadequacy of "computational units" in network paradigms (that are still based either on McCulloch-Pitts, or Hebbian hypotheses, while contemporary neuron models develop in leaps and bounds) is becoming ever more glaring. The significance of single cell modeling is critical in both fields of single cell electrophysiology and neural- net computation, in order to base network paradigms on conceptually proper "computational units" and, ultimately, to be able to manufacture E.g. silicon-based "neuromimes". Single cell modeling is vital if single cell electrophysiology and neural-net computation are ever to be connected.

2) Sensorimotor Coordinate System Transformations. Ever since the times of Descartes [27], motor control, and specifically, sensorimotor coordination have been in the forefront of research aimed at un derstanding principles of brain function [28]. A sensorimotor act is a behavioral function of the CNS that can be directly observed and precisely measured by physical means. Basic reflexes, such as gaze- stabilization mechanisms rely on a relatively simple underlying neural network which is accessible to direct neuroanatomical and neurophysiological investigation [29],[30],[31]. Thus, quantitative system neuroscience has focussed on (sensori)motor control [32],[33],[34]. Neurocomputing, which aims at implementing applications of neural control such as in motor control for the purposes of rehabilitative medicine, prosthetics, and even robotics provides a strong additional impetus forrevealing principles of (sensori)motor performance [35],[36],[37],[38],[39].

Motor control research has traditionally revolved around powerful, though not necessarily neural-level, central concepts. A classical view pointed out the massively parallel nature of CNS function and emphasized that the brain is a hierarchy of interconnected re flexes; "the enchanted loom"; [40]. This approach has gradually been oversimplified to motor control formulated in terms of agonist-an tagonist pairs of muscles. This classical approach is represented today in the form of systemtheory applied to oculomotor function [32] [41],[42],[43]. A concept with great integrative potential was brought into the field by a pioneer of of motor control [44], who is credited with the idea of "synergies", a pattern of many muscles acting in concert. Such qualitative tenets are instrumental for their intuitive power but the essence of motor control is quantization and ultimately carrying it to neural level. In his seminal work, Bernstein (1935) attempted to carry the qualitative descriptive notion of "synergies" towards a quantitative concept couched in geometry. A geometer, F. Klein, pointed out that "geometry is the theory of invariants" [45]. Thus, it is logical that the early geometrical view of motor control led more and more to research concentrating on invariants that can be observed such as the trajectory-pattern in arm and hand movements [46],[47]. With the two lines of thoughts combined, phenomenological observations of trajectories from an equilibrium developed as a research-school [48],[49],[50],[51] also in combination with rehabilitative medicine and robotics [35],[37].

Connecting motor geometries with the underlying neural network and the physical structure of sensori-motor systems, is difficult. Lately, a multidimensional coordinate system approach has been adopted by several schools [46],[52],[38],[53] and in sensorimotor research, in general, a multidimensional coordinate system ap proach is taking hold [42],[43],[2],[54],[55],[56]. The promise of a (generalized) coordinate system approach is that it is inherently multidimensional, geometrical and at the same time is likely to pro vide both a mathematical grip on holistic motor control problems such as trajectories, posture and motor strategies, while explaining them in terms intrinsic to CNS neural net function. Since a coordinate system approach is by definition quantitative, it provides the needed link between biological motor control and neurocomputation aspects.

3) Multielectrode Recording Technology and Theory. The historic opportunity inherent in the advent of multielectrode recording techniques has been extensively discussed elsewhere [57], and thus will only be encapsulated below. Although classical neurobiological experimental methods of investigation (electrophysiology) have been developed for single units, it has long been evident that, given the axiom that the CNS is a massively parallel system, new methods needed to be in-vented to access a multitude of neurons simultaneously. Multielectrode recording techniques have been pio- neered through the past decade, cf. [58],[59],[60], [61],[62],[63],[64],[65]. In part because such revolutionary methods are exceed-ingly demanding in terms of human and material resources, attention only recently focused on the further, and equally difficult question of how to theoretically interpret the arrays of data made avail-able by such parallel methods.

The practical significance of a theoretical model-interpretation of multielectrode signals is that great sums are being in-vested to develop multielectrode recording techniques. In contrast, theoretical invest-ment to work out methods of how such data could be analyzed is grossly underrepresented in this spe-cific case, and in modern science perhaps also in general. For instance, this author is closely familiar with several laboratories engaged in multielectrode recording. Still, there appear to be precious few theoretically based pilot-studies that would make experimentally testable predictions how an emerging multielectrode tech-nique should yield important new insights to CNS function although it is uniformly believed in the neu-roscience community that such breakthroughs are imminent. The strategy implicit in developing multielectrode techniques appears to be that experimentation ought to proceed on a serendipitous path of simply probing neurons and thinking about the obtained results perhaps later. But a major difference is that tra- ditional electrophysiology started with answering utterly simple "theoretical hypotheses"; E.g. whether an external stimulus is connected by a pathway to a probed neuron or not. Modern network hypotheses are infinitely more com-plex, rendering this old strategy obsolete. This contrast is comparable to two periods of nuclear physics: at its dawn it was sufficient to serendipitously mea-sure radiation-properties of different chemical elements, where some atoms exhibited intriguing spontaneous fission, some did not. In con- temporary research, however, when a super-collider costs tens of billions of dollars and maddeningly complex patterns emerge from smashing atoms to myriads of flickering particles, no one would dare proposing (let alone funding) such research without massive background investment into theory-based computer model simulations. Sound theoretical modeling should generate well posed questions for an experimental technique, even before the device is built, e.g. for multielectrode recording.

4) Brain and Body Imaging. Modern computerized technologies such as Computerized Axial Tomography [66], Positron Emission Tomography [67],[68], and Magnetic Resonance Imaging [69] provide a formerly unimagin-able wealth of information, permitting researchers to observe the living brain and body in a noninvasive manner and in much higher detail than ever before, especially with MRI. It is only the most direct and useful clinical application of such imaging technologies to identify brain systems that are implicated in specific disorders and disabilities. Neither the development of technology (not even mathematical-com-putational aspects of imaging itself) nor the obviously exceptional clinical applications belong to the di-rect funding responsibilities of the MTC research program. It is, however, a major mathe-matical/ computational and mostly theoretical challenge of how such a wealth of information can be intelli-gently and costeffectively utilized for scientific understanding.. At least two different kinds of impact of modern (together with the classical) imaging are very likely. In a negative sense, the age-old desire of pre-cise lo-calization of various CNS functions will probably have to be given up in view of rapidly accumulat-ing evidence that the instinctively assumed "principle of localization of function" (that each cortical territory belongs to specific function) is an oversimplification. As the time-resolution of imaging techniques will catch up with the ever increasing spatial resolution, dynamically changing macro and micro-patterns are likely to be revealed. These findings will drive researchers to give up the implicit theoretical axiom of territorial organization of CNS function. Thus, by denying neuroscience one of its most ancient implicit ax-ioms, imaging techniques will greatly stimulate a theoretical search aimed at replacing this fundamental assumption. This is a major theoretical challenge which is presently hardly faced as most workers still hope that the age-old "homunculus" could be refined into a modern spatial representation model of CNS function.

5) Functional Neuromuscular Stimulation. The above mentioned body imaging has an important ap-plication also in Functional Neuromuscular Stimulation. Research of the morphological organization of structure (as anatomical basis of function) already directly benefits from development of imaging tech-niques. On one hand quantitative computerized histology of neural elements of the CNS [70],[71] will enormously profit from introducing into "brain mapping" the latest techniques of computerized imagery. On the other hand, the so far rather neglected complementary mapping, that of the body (which the CNS uses) will undoubtedly have to emerge as a main user of present and future techniques of modern imaging. Indeed, projects aimed at complete mapping of the CNS, and even complete mapping of the human genome are on the drawing board, yet a computerized image-based anatomical map of the human body is not only nonexistent, but hardly planned. This is in spite of the fact that not only does neuroscience re-quire quantitative knowledge-base of sensorimotor systems that CNS operates with, but other fields of re-search and technology such as sports and rehabili- tation medicine, ergonomy, kinesiology, functional de-sign of man- machine systems such as air and space-planes, terrestrial and aquatic vehicles, all badly need and should co-sponsor development of a computerized body mapping project. Such anatomical mapping (e.g.of obvious sensorimotor body mechanisms such as vestibulo- ocular systems) goes back more than a century [72], and recently there is a clear trend of evolving from classical methods [73],[74], [75],[76],[77] towards the use of more modern imaging techniques such as MRI for such pur-poses [78]. An advantage of using brain and body mapping together with sensorimotor models (such as transformation of generalized body coordinates) is that mathematical and computer modeling techniques can be developed and used jointly in a manner that data-procurement proceeds jointly with the development of neural mathematical theory of such sensorimotor transformations [2],[79].

Functional neuromuscular stimulation [80],[81],[82] is a particularly important field of research that also stands to benefit from computerized body mapping. In case of paraplegics disabled by spinal cord trauma (due to injuries afflicted by wars, or by car, swimming pool and sport accidents) the musculature is initially intact but the neural control and coordination is gone. Artificial replacement of this function by a "neurocomputer" prosthesis that can electrically stimulate muscles (either transcutaneously or by electrodes attached to nerves) is envisioned here as a prime application of neurocomputer research and development. The significance of such a project is manifold. First, such a civilian, highly visible socially beneficial application of neurocomputers is much needed, both to demonstrate the impact of neurocomputers on society as well as to help generate Congressional support for a civilian neurocomputer initiative. Replacement of rudimentary spinal cord function with that of a "neurocomputer" (even if not producing ambulation but only a tool for paraplegics to stand up) is an obvious explanation and showcase of the utility of neurocomputers. Second, coordinated motor control of typically a dozen of muscles (acting with rather slow biological speed) is technologically much less challenging and thus much more feasible in the present state of art of neurocomputing than e.g. recognition of electromagnetic patterns in typical defense applications, where both the number of controlled variables and their speed of change is several orders of magnitude higher. Third, functional neuromuscular stimulation application would work with computerized anatomical models based on data acquired by body imaging (see above); thus the cohesion of these biologically oriented projects would be enhanced.

6) Mathematical Theory of CNS Function. This activity must be the backbone of any theoretical mathematical/computational neuroscience program. This report draws close comparison throughout its arguments between the present evolution of neurosciences and a similar earlier revolution in nuclear physics and technology. As any field of basic science, neuroscience will at one point have to develop its mathematical discipline (just as nuclear physics did it with quantum mechanics). In neuroscience, mathematical psychology became closest to this goal [83] , but it was arguably an application of mathematicsto neuroscience rather than discerning the mathematics intrinsic to brain function from neuroscience data. The difficulty here is that the enormity of the scientific challenge of creating the mathematical discipline of brain function is generally not even realized yet by either the neu roscience community or that of pure mathematicians. It is essential therefore that the MTC program create outstanding precedents which can draw attention and attract the best workers to this centrally important field of research. Only such a concerted ap proach can create the discipline of CNS mathematics. It must be emphasized again, that mere "mathematization" of existing subproblems of data collection is an important, but qualitatively different exercise from developing mathematical axioms, formalisms, theories and testable predictions. In order to help to distinguish such extreme classes of phenomenological modeling and general mathematical theory, it is recommended that the MTC program formally separate research proposals by putting them into groups with rather different priorities. Mathematization of phe nomenology should be of lower priority than establishment of mathematical theories (especially if they provide quantitative experimentally testable predictions).

The main difficulty in defining a mathematical theory of CNS is that the intrinsic mathematics of brain function is unknown. It is even arguable which branch of mathematics is the most relevant. In very recent times, a school of researchers is emerging who maintain that the brain is a geometrical device. Such geometrical approaches to brain function fall roughly into two classes, one which uses classical (Euclidean) geometrical axioms and formalism such as linear Cartesian vector analysis; [11],[84],[85] and a more novel class in which workers emphasize that brain function and structure (similar to other manifestations of natural evolution) is based on profoundly different geometries such as that of noneuclidean metrical manifolds, [86],[87],[88],[89] chaotic geometries [90],[91],[92],[93],[94],[95] and fractal geometries [96],[97],[98]. Several comments on these novel trends appear necessary. Geometrical approaches of metrical manifolds yield mathematical formalisms for the analysis of either classical Euclidean geometries see Cartesian tensor analysis; [99],[100] as well as the analysis of curved noneuclidean CNS functional spaces [101]. Subtle differences in mathematical axioms however are more difficult neuroscientists not trained in mathematics to evaluate, than superficially grasping spectacular chaotic and fractal features that nonlinear dynamics displays. It is quite telling of this effect that although the founder of fractal geometry only conjectured that brain cells were fractals [96], a statement that "Neuron exemplifies fractal structure" already found its way in an affirmative manner even into the most recent issue of Scientific American [93], p.49, although support of actual research of substantiating the basic conjecture of fractal neuronal morphology is just about to emerge.

In a larger sense, there is a danger of concentrating on trees and not seeing the woods by looking at chaos and not seeing geometry. Different manifestations of neural geometry (metrical, fractal and chaotic) are all interrelated. It is less important therefore to boost a trend of data-collection on any particular manifestation (e.g., chaos) but it is more important to spend more on theoretical research of discerning mathematical laws of phenomena underlying nonlinear dynamics displayed by CNS (e.g. how a chaotic behavior relapses from a metrical, locally linear, geometry). Just as in the functioning of a human society, chaos is often a transient stage from one higher order (geometry) to another. Although the rule of local laws (chaos) is surely the most dramatic and colorful dynamic stage, knowledge of the more general laws of order that emerge from such disorganized stages are often more important and useful. It is quite evident that we are very far from broad-based theories that could shed light on fundamental mathematical laws of operation of CNS with various but related noneuclidean geometries. To help mathematical and theoretical research on this important problem, and not merely to collect more data, is recommended to be a primary task of MTC program of NIH-NIMH.

7) Theory of Unified Spacetime Geometry of CNS. An important role of brain theory and modeling is to consolidate the axiomatic basis of neuroscience. This is much needed, since some of its (implicit) a sumptions are based on demonstrably untenable axioms. An example of this is the traditional space and time separation, that is the result of an uncritical adoption in neuroscience of Newtonian mechanics. There, light serves as an agent that ensures simultaneity of (much slower changing) space-coordinates but the CNS cannot use that principle since in neural network function, the axiom of simultaneity (which can only be established by a superfast agent) is inapplicable. Replacement of scientific axioms may take centuries (cf. Galilei), and even in modern physics it took decades from initiation to textbooks [102],[103],[104]. Therefore, it is not unusual that a concept of unified spacetime coordination in CNS introduced in [105] awaits detailed models in order to be practically useful for a spacetime-analysis of specific sensorimotor systems not just being an unsettling new axiom. The vestibulo-collic head stabilization reflex in cat is a system in which space and time information routinely converges (vestibular canals and otolith report on different time-derivatives which merges in the three neuron arc; [106],[107],[108]). Also, such spacetime derivatives can be accessed at neurons. Elaboration of a model of VCR with spacetime-performance, however, is impractical without the availability of proper units (such as a neuron-model which is capable of producing time-derivatives). A significant practical aspect of spacetime analysis is that electrophysiologists routinely use sinusoidal testsignals, although theoretically predicted time derivatives (and delays or inhibitory effects) all yield a non discriminative sinewave whose phase confuses all the above effects. By proposing more effective methods of spacetime analysis, theory and modeling, again, could save both animal lives and costs.

8) Theory of the Mind-Body Problem. The chain of questions of whether we can understand details without a more general framework, ultimately leads to the problem whether the broadest (philosophical) approach of neuroscience is appropriate. It is characteristic of modern funding structure, aimed principally at data collection, that theoretical problems bordering on philosophy are generally regarded as unfundable (see an elaboration of this point in [109]). Again, the strongest recommendation that one can give in this regard is to "take the bull by its horns", and fund those workers who have the preparedness and will to directly address this theoretical issue. Merely hoping that "accumulation of experimental data will finally tilt the balance", is likely to be futile in view of experience. Since this field lies on the borderline of theory and philosophy, it appears particularly important to increase their overlap. This could be done effectively by jointly sponsoring meetings or establishing sponsorship of sabbaticals (as recommended in Sect. 4.7).

9) Theory and Modeling of CNS Systems. Because of the analytical nature of experimentation and nascent stage of MTC activities, there is a shortage in integrative theories and models of CNS systems. Most mathematical modeling and theory concerns subproblems of channels, synaptic mechanisms, single cells and small circuits. Efforts should be encouraged to outline broad and more general theories for larger subsystems and then entire systems. Modeling and theory of the cerebellar system is perhaps the most advanced in this regard [12],[110],[111]. Vestibular system models have also evolved since the thirties [29], [30],[32],[112],[77],[113],[56]. Time is now to integrate such models to representations of the vestibulo- cerebellum, with later integration of the colliculus into a sensorimotor system theory and model. Similarly, the retina, the hippocampus, the olfactory bulb, pyriform cortex, etc. are entire systems that are ripe for integrative theory and modeling. Since such efforts are sizable and draw manifold criticisms, workers attempting such integration have to take a more than a usual risk. Thus, to grow the first crop of such integrative theories and models, encouragement and strong support is vital.

10) Research on Emergent Properties. This is an area in the forefront of curiosity. It is indicative of the theoretical weakness of neuroscience, however, that certain categories are created as umbrellaterms, serving purposes of identification rather than explanation. The term "emergent property" is an eminent example of this effect. This author would like to compare to the category of "emergent properties" to "synergy" (of e.g. muscle activities as mentioned in part 2 of Sect. 2.1), or to the term "pattern" (of elec trical activities of e.g. brain cells). All these categories are loosely indicative of the general intuition that the "whole is more than its parts". It is the task of mathematical theory to turn these intuitions into novel concepts elaborated by rigorous formalism. From the point of view of mathematics, it is remarkable that e.g. x,y,z components of a mathematical vector constitute parts of an entity whose existence is only evident if all parts (coordinates) are taken together. This primitive mathematical example illustrates that mys terious cooperative features of components become selfevident in retrospect, once we know the entity itself whose components we are dealing with. This author maintains the view, therefore, that whenever the category of "emergent properties" is raised it indicates the need of a theoretical integration of bits of knowledge into a more general (mathematical) understanding. "Emergent properties" are not much more than calls for theoreticians to explicitly come into play. As this author likes to identify the origin of "emergent properties" to McCulloch's classical hypothetical question ("Where is fancy bread?" [114]), it is perhaps appropriate to suggest that funding efforts should be oriented towards explicit theoretical studies to generate the missing overall understanding (just as McCulloch did so successfully in his pioneering; [114]) rather than towards efforts engrossed in the mysticism of the category itself. It may be appropriate for Sperry to argue in 1980 that mental states are emergent [115]. It is expected, however, that in the future neuroscience will have the mathematical/ theoretical means to explain how such mental states emerge from specific neural net structure and function. Until then, umbrella terms as "emergent properties" are only useful indicators where problems in our theoretical understanding hide and thus where should we focus targeted basic research to shed light on them.

11) Research on Neural Networks (Neurocomputing). This is a central area in MTC. It is mentioned last, since the main argument of this report is that neurocomputing is the field of research that could quite effectively coalesce several of the above approaches in MTC. It will be shown that certain specific problems that are regarded as separate fields of interest (e.g. Fault Tolerance, Neural Modularity and Modulation) arise as specific features of Neural Networks. Because of the importance and great potential impact of "neural network research" elaborate background information is provided on neurocomputing in the Appendix.

2.2 Existing Programs Relevant to Neurocomputing at NASAARC

Just as NIH is coping with many aspects of neurocomputation, a large number of elements of neurocomputer research already exist throughout NASA research facilities; see an early compilation in [116]. The efforts are scattered however throughout all research bases and within a particular center (for instance ARC, see below). Different approaches are often quite isolated from one another. Further, in most facilities neural net research is pursued from an almost entirely technological viewpoint, with only a token presence of the "neural" part of neurocomputing. It is found noteworthy that current attempts to launch largescale neurocomputing programs, such as the Jet Propulsion Laboratory's proposed "Neural Information Processing Systems Program" ($95M/10 YR), presently considered by NASA, are not directly coordinated with Life Science (Neuroscience) Programs, in order to integrate them with Technology Research and Development Programs. It is recommended (see in 5.4.) that an integration mechanism be established before neurocomputing programs of such dimension and significance are launched, to ensure that (a) each half of neurocomputing receives its adequate share (or at least a third) of support, (b) a NASA neurocomputing program is coordinated with all research centers, (c) the NASA neurocomputing program is coordinated with similar programs of other US gov ernment agencies.

This report finds it a remarkably unique feature that NASA Ames Research Center (NASA-ARC) combines both technological and life science oriented (neuroscience) aspects of this field of research. However, in part because of the local organizational structure of being composed of different directorates, it is found that different aspects are often pursued in separate groups with fairly weak interconnections. Although an integration of these hitherto relatively isolated efforts is clearly desirable and often actively sought by individual scientists and engineers, the efforts have not yet conglomerated into projects which "could be more than the sum of components". To illustrate this point, a brief survey of the NASA- ARC neurocomputing related research efforts is given below. Clearly, some efforts involve explicit commitments but take a technological approach. Others are life-science-oriented, with an evident impact on neurocomputing but without either explicit commitment to neurocomputing or adequate wherewithals for such task. All would clearly benefit from interactions with each other in come sort of structured neurocomputing program.

2.2.1 Technological Approaches

(A) Software research by Drs. Michael Raugh and Pentti Kanerva (at Research Institute for Advanced Computer Science; RIACS a contractor to NASA under Universities Space Research Association). This is essentially an effort in software design aimed at developing parallel computational paradigms based on the Sparse Distributed Memory mathematical algorithm [117]. Extensive efforts have been made to relate this mathematical concept both to actual neural mechanisms of the cerebellum as well as to engage, at Stanford University, in a machine implementation of the computational paradigm.

(B) Hardware research and development by Drs. Gary Hill and Lloyd Corliss (at Code FAM, Aerospace System Directorate, Aircraft Technology Division, Military Technology Office). This is a Northrop/DARPA supported effort concluding in hardware for massively parallel computations used in aircraft design.

(C) Parallel computing for high-speed computer architectures for space based intelligent systems by the group of Dr. Henry Lum (at Code RI, Aerophysics Directorate, Information Sciences Division). This is an ef-fort to use parallel computing, with an interest in neurocomputing, for tasks in intelligent systems research, automation and robotics. Applications include high-speed 3 dimensional graphics and pattern recognition as well as control problems in aero-space crafts and robots. Members of the team, specifically Mr. Coe Miles, maintain close cooperation with RIACS (see A).


2.2.2 Life Science Approaches

(D) Physiology of the vestibular system. This research is based on Dr. David Tomko's group at Code SL, Vestibular Research Facility, VRF [118; 119]. Physiological studies are determining the way in which gravity receptors function to control eye and head movements during linear acceleration on Earth and in space. The author of this report contributes to this vestibular system research performed by this group by mathematical modeling and theory. Neurocomputing impli-cations are at least threefold. First, mathematical modeling of the vestibular control of eye-head-neck system could obviously greatly benefit from the high-speed 3 dimensional graphics research above (see C). Second, the effort to-wards discern-ing the mathematical computational paradigm inherent in senso-rimotor neural net transformations would benefit from contribution by activities in (A). Third, the hard-ware effort (B) could be instrumental for a pilot project funded by the Director of NASA-ARC, of using at VRF a hardware neurocomputer and paral-lel processing software development for the analysis of multi electrode recordings from the vestibular system (see point 5.3.3.1).

(E) Morphology of the vestibular system. The approach by Dr. Muriel Ross (at Code SL, Biocomputation Center) is aimed at grounding mathematical & hardware neu-rocomputing research in neuroanatomical realities. Quantitative histological analysis and anatomical imaging were proposed as techniques most suitable for discerning mathematical principles of biocomputation [120]. These projects stand to benefit from ties de-veloped with project (C). An integration of anatomical models e.g. of the otolith system with systems physiology of the vestibulareye head-neck system (research D) is also both desirable and eminently attainable, an such a combination is likely to lead to learning mathe-matical and computational paradigms from actual biological organisms. Quantitative histologi-cal analysis provides the anatomical basis of neural network research (e.g demonstrating the mas-sively parallel con-nections already at the primary sensory level [71]. It is quite likely that existing and future anatomical imaging techniques will be suitable for providing a data-base for discerning mathematical principles of the structural geometry of neurons, with the strong possibility that such geometries are fractal [96],[98],[93]; see point 5.5.4.2. Therefore, it is highly desirable as well as eminently feasible to estab-lish an automated and fully computer-ized technology in the form of a quantitative histology laboratory. This may become a center for those in-volved in identifying the anatomical basis of biocomputation.

---

3.0 Advantages and Disadvantages of a Joint NASA/NIH Neurocomputing Program

Both the possibilities and challenges to the development of joint NASA/NIH neurocomputing program stem from the interdisciplinary nature of neurocomputing. The strongest argument of this report is that this fledging field might fail if in order to avoid difficulties the two sides neuroscience and com-puting technology are not brought together at an institutional level.

The disadvantages are obvious, and are almost exclusively logistical. Bringing such a cooperation to life requires a major organizational task, as the two separate government agencies have differing charters, goals, structure and politics. These obstacles and difficulties are apparent even in the scope of compiling this initial report. These logistical problems led to the conclusion and recommendation (Sect. 5.2.) that the best way to minimize such disadvantages may be to create a US governmental Neurocomputing Advisory Committee which could most effectively coordinate such rapprochement.

The advantage of a joint program is manifold, and mainly scientific. Since neurocomputing may not suc-ceed without integration of its two basic aspects, creating a coordinated joint project is scientifically ad-vantageous for both institutions as they become the institutional guarantor of such linkage. For NIH, a main advantage is that neurocomputer applications will greatly increase the accountability of the many billions of dollar of investment into neuroscience research. Indeed, this is not only an advantage, but also a respon- sibility, as trustees of this massive investment that generated a great body of knowledge, are expected not to stay on the sidelines witnessing how "neurocomputer" research strays to directions of research that have no validation by actual knowledge of the biological brain. A further advantage of a joint program for NIH is that theory, validated or falsified by actual technological implementation will rejuvenate the so far almost exclusively experimental and descriptive neuroscience. Among the advantages of bringing mathe-matical and computational techniques are decreasing the dependency of neuro-science research on animal experimentation and by selecting experimental proposals help to contain the escalation of the overall bud-get. Other advantages are pointed out throughout this report.

As for NASA, the main scientific/technological advantage stems from the fact that neurocomputers are the computers for future aerospace applications as they are gracefully degradable, fast, and simple to program. For example, neurocomputing stands to contribute to development of computers capable of controlling dy-namically unstable aircraft, or computers capable of controlling remote landing craft with adaptive characteristics that normally can be accomplished only by human operator. Thus, NASA has an extremely seri-ous stake in success of this new field of research and development, and NASA's Life Science program can significantly contribute to the acquisition of basic knowledge necessary for devel- opment of such brain-like computers. If NASA did not have a Life Science Program of its own, it might be conceivable that neuro- computer research would be pursued from a purely technological viewpoint. Joint neurocomputer re-search, however, presents the advantage that NASA's Life Science Program can be the linchpin connect-ing the NIH-sponsored basic research in the neurosciences with actual technolog-ical applications in the realm of NASA. Although neurocomputing is not mentioned specifically in the Robbins committee report [121] it is obvious that achievement of many of the goals specified by that study would require a new gen- eration of computers.

3.1 CONCLUSION: Consolidation of Compartmentalized Organization
is Difficult but Permits Launching Integrated Projects Necessary
for Sustained Neurocomputing Research

A central scientific advantage of a joint neurocomputer initiative is that it permits launching inte-grated projects, such as an artificial vestibulo-cerebellum project (see Section 5.4.3) that by their nature re-quires both neuroscience basic research and technological applications. It is argued in Sect. 4.3 of this re-port, that such integrated projects have numerous advantages, but so far they could not be launched be-cause of compartmental organization.

---

4.0 A Neurocomputing Program: Major Tasks to be Accomplished

4.1 Secure Funding for Theoretical Neuroscience

Theoretical proposals compete with all (experimental) proposals. The present arrangement in which theoretical/mathematical/computational neuroscience proposals theoretically have "unlimited" ac-cess to the total pool of NIH resources is appealing at first sight. However, any tightening of resources almost immediately freezes new, creative and thus controversial ideas, as they by necessity attract dis-senting opinion. Theoretical neuroscience is far too young, controversial and politically heterogeneous compared to most wellestablished, coherent and mutually supportive fields of experimental neuro- science. Thus, "free competition access" practically guarantees that few theoretical proposals will suc-ceed in getting funded, and thus reputable theoretical workers might loose trust in this new NIH pro- gram. A decreasing interest on the part of extramural scientists could weaken or permit elimination of this vitally important and potentially extraordinarily influential program.

Suggested countermeasure. Convert the presently ad hoc Study Section into a permanent committee and develop a definite (even if relatively small) budget specifically allocated to this program. "Equal access" of both theory and experimentation to taxpayers' research dollars will be reality, rather than illusion, if this program will have equal opportunity for having its permanent regular Study Section and its estab-lished pool of own resources, as other "regular" experimenter programs do. Even with such "equality" in the funding mechanism, the actual balance of funds will more than likely be grossly biased towards experi-mentation although it is increasingly evident that a growing and very vocal portion of taxpayers would prefer replacing animal experimentation with alternative approaches whenever it is scientifically sound.

4.2 Establish Accountability of Theoretical Research

Theoretical work is not accountable. Mathematical/Theoretical/ Computational Neuroscience activities do not presently enjoy a status that is commensurate with the difficulties involved with conceiv-ing, incubating, nurturing and fully developing mathematically proven theories that can lead to com-putational models which provide hypotheses for experimentation. One of the strongest reasons for this defi-ciency is that it is difficult to measure the value of theory before it has been applied.

Suggested countermeasure: Funding theory and modeling establishes an accountability. In the best tradition of the NIH funding system, evaluation of research efforts can be squarely based on the return of the agency's in-vestment. Many theorists took considerable "controversy" and "turf-protection antag-onism", most of the time not coming from peers, before, or instead of receiving direct support. It is im-portant and fair to establish an NIH policy to fund theory first, then evaluate performance rather than using upfront criticism to deny funding.

4.3 Define and Fund Integrative Projects

Compartmentalized organization prevents launching integrated projects. As it was pointed out throughout this report, one of the main difficulties of any NASA-NIH(NIMH) cooperation in neuro-com- puting is their organizational independence.

Suggested countermeasure: As a means of minimizing such problems, it is suggested that the Separate agencies involved identify those research projects that are co-fundable. Launching such integrative pro-jects would permit concentrating only on minimal organizational effort allowing creativity and resources to be focused on the research projects themselves.

Such an approach requires the definition of those most important research areas that are co-fundable. It is suggested here that major neurocomputing efforts should be centered on specific identifiable organisms of the CNS, aiming at transferring the available body of knowledge into a theoretical/mathematical un-der-standing of the specific system, and further, concluding in an actual implementation (utilization) of such understanding by means of engineering. Such projects, in effect combining neuroscience research of CNS systems with the creation of their artificial (electronic) equivalent offer several advantages, as shown in 5.4.3. and can be verified from a few existing examples (most particularly from the Artificial Retina Project by Dr. Carver Mead [122]).

The main advantage of "Artificial CNS Organism" projects (such as Retina, Vestibulocerebellum, Hippocampus, Colliculus) is that they provide not only an organizational but more importantly a scientific anchorage of neural network research. While theories and models of any specific part of the brain (e.g. of the cerebellum) may be inappropriate for the biological brain, or even outright misdirected, the underlying validity of the actual biological solution will always loom large in front of the researchers and ultimately guide researchers back to the right nature-proven track. Thus, many million years of natural evolution will guarantee the scientific soundness and feasibility of such "Artificial CNS Subsystem" projects.

Actual utilization by engineering of a mathematical understanding of CNS function will serve a multiple and central role. First, experimentally based data-gathering will be purposefully guided towards those missing pieces of information that can lead most efficiently to a utilizable understanding. Second, uti-lization is always a practical check of the soundness of understanding. Having an "artificial cerebellum" prototype at hand, cerebellar models and theories may be pooled and compared, including those whose pri- mary thrust is not towards contributing to an understanding of a biological organization (but which may well yield extremely useful technological inventions) and those which provide specific guidance as to how to construct electronic equivalents of actual biological systems. Third, even models and theories which will turn out to be different from the actual biological organism would be utilized since a project can be even more important if it "concludes in the design of a fast wheel although it was originally aimed at mimicking a slow leg". Therefore, artificial CNS system projects will be conglomerating rather than divi-sive. Fourth, such projects could far better be justified by socioeconomical arguments, as they will not be perceived as pure expenditure but as projects with potential for generating actual returns. Fifth, such pro-jects will not only bring and keep researchers with rather different background and exper- tise together but are likely to develop those communication interfaces (jargon, formalisms, even underly-ing mathematical theory) that can serve as effective bridges among the contributing communities. Sixth, as will be shown in Sect. 5.4.3. such projects, as the Artificial Vestibulo-Cerebellum Project, can serve as effective scientific platforms even in cases where organizational structure is very unlikely to play such a role. Finally, the common interest of funding such integrative projects will naturally increase the coherence among dif-ferent Agencies, thereby ensuring the continuity of funding of such projects as they evolve from being mostly neuroscience-oriented at first and become mostly engineering-oriented later.

4.4 Improve Evaluation Mechanisms for Theoretical Proposals

Theoretical proposals rarely get a true peer review. With few full- fledged theorists around the re-viewing process can hardly be completed by peers only. In addition to direct peers, review can rely on ex-perimentertheorists (morphologists, electrophysiologists, psychologists, etc with an active interest in the-ory). Such reviewers may lack appropriately strong mathematical/theoretical/computational back-ground, or may be biased towards approaches that serve interests of particular experimentation, however. A third class, that of pure mathematicians, physicists and technologists can also be used for re- view. Such workers may be less afflicted by the above disadvantages, but may not necessarily have contributed themselves to theoretical neuroscience; thus may exercise judgement rather than acting as true peers.

Suggested countermeasure. A real solution will only come when, and if, theoretical neuroscience reaches critical momentum. Until then, the reviewing process may wish to balance the use of potentially biased ex-perimentertheorists by employing workers as reviewers whose best interest is to use theoretical (mathematical) understanding, not just burgeoning knowledge, gained from neuroscience research. These are users, for instance basic scientists in the field of neurocomputing, who are ready and eager to employ any hard understand-ing emanating from experimental brain research. These scientists would certainly en-courage new ap- proaches that may bring them closer to turning understanding into utilization. This ar-rangement would be similar to NASA's Universities Space Research Association (USRA), an organization that looks for, and regularly supports, University scientists who possess scientific under-standing that NASA could turn into useful application. It is recommended in Sect. 5.5.3. that NIH-NIMH create a Neurocomputing Research Association, either as a branch of NASA- affiliated USRA, or simply an NIH affiliated pool of University-based scientists similar to USRA. Such a body could help make the Study Section permanent and much closely resembling a truly peer-review mechanism.

Other than improving the composition and status of the Study Section, it should adopt routine tests for screening proposals. Theories applicable to an MTC program should qualify by passing the "three Galilean tests; of simplification, unification and mathematization" [109]. Finalists are expected to pass seven tests: 1) Proper philosophical foundation (e.g. do not try to theoretically frame the brain as a ma-chine), 2) Solid axiomatic conceptology (should be free of axioms that are proven inappropriate), 3) Suitable mathematical formalism (method must not be demonstrably too limited or rigid), 4) Internally self-consistent exposition (theory should be contradiction-free), 5) Demonstration by computer simulation (should accompany any theoretical proposal), 6) Experimentally testable predictions (their existence should qualify a proposal to a higher class), 7) Tested, confirmed and used by ex- perimentalists (this is the ultimate test of scientific theory). True values can also be spotted by laymen; checking not only for a) users and enthusiasts, but b) imitators and c) appropriate antagonists, since "great ideas often meet violent opposition"

Three further tests should determine who are most NIH-fundable modelers and theorists. To qualify, one should document 1) expressed commitment to understanding the biological brain, 2) willingness to com-municate with "wet" neurobiologists, 3) ability to produce computer models that can provide ex-per-imentally testable predictions.

4.5 Alleviate Dependency of Math Modelers and Theoreticians
on Experimentalists

It has been and is difficult to obtain independent funding for theoretical studies in Neuroscience. As a result, most modelists or theorists are sup-ported by experimentalists, often using "bootlegged" re-sources for such purposes. Dependency can result in abuse of either mathematics and/or personnel. It is pointed out elsewhere [109] that one of the more obvious results of such a funding structure is that per-sonnel capable of creatively using mathematical/theoretical/computational techniques may be employed for "mathematization" of rather minute subproblems. A mathematization or computer modeling of phe-nomenological details will not spontaneously coalesce, however, into more general models, much less into true theory. In part because of such problems, workers with mathemati-cal/computer science expertise often depart from neuroscience towards "dry" applications typically found in computer industry. In either case, theoretical neuroscience is shortchanged.

The significance of helping theo-rists escape their dependence for funding on experimenters is important for the following reasons. 1) Serving dayto-day pressures, brain theory and modeling can become a crop that is constantly har-vested and not let grow its own roots. Thus, brain theory and modeling may concentrate only on high-yield expediency, and may tend to get used before ready, generating contro-versies stemming from misunderstandings due to premature promulgation of ideas. 2) It may be in the interest of employers of theorists, especially if they use theory as a leading edge, to guard their posses-sion and be reluctant to share such resources with competitor experimenters. The interest may be to iso- late or alienate brain theory and modeling from the experimental community. Career development and growth of brain theory and modeling workers is potentially against the interest of their employer; daunt-ing their growth requirements. 3) Access to captive brain theory and modeling, when it is ready for a wider use in experimentation, can be difficult. When external experimenters succeed to enter into cooperation, proposals compete for what is perceived as "experimenters' money", and proposals are re-viewed not by theorist-peers but by experi-menters. Direct users may strongly favor such a proposal be-cause it gives their group a lead, while com-petitors are likely to oppose it for the same reason. It is the duty of theorist peers to sort out competent proposals with maximal potential to generate progress in brain theory and modeling. Once existence and growth of brain theory and modeling is granted, tests and thus survival of models and theory should be up to experimenters.

Suggested countermeasure: Since nascent theoretical ideas are often immature, it seems prudent to create a pool of "seed-funds" for developing creative theoretical ideas to a stage where they can produce exper-i-mentally testable hypotheses.. NIH claims (e.g.in the Neuroscience meeting in 1989 Phoenix) that it is "Seeking New Ideas". The present crisis condition of close to single digit funding percentage virtually guarantees that new and creative ideas quickly dry up. Unless some countermeasures are implemented, novel concepts are unlikely to get funded as any new idea (which is by definition controversial) rarely gets the unanimous consensus necessary for obtaining such priority levels. As elaborated in Sect. 5.3, to re-solve this problem, NIH may wish to seek new ideas" by setting aside a pool of funds to provide "seeding" support for theoreticians/modelists, joint proposals written by a theorist PI and co-signed by an experimentalist who wishes to risk such a test (or vice versa), should get that modest yearly 50-100k that such theoretical efforts require.

To establish independent existence is important not only for professional theorists, but also to experi-mentertheorists, who also pursue a difficult and controversial task. They must face the dilemma of de-vel-oping and verifying their own theory. Both tasks require fulltime effort, however. It is difficult to properly fulfil even one of the tasks. Also, a bias may be unavoidable in trying to prove one's own the-ory. Not engaging in tests, on the other hand, would contradict to commitment to the-ory. An opportunity to get in-dependent funds for theory should let experimentertheorists come out to the open, announce their theo-ries-and we all shall see how genuine, impartial experimenters check them out.

The experimental/theoretical balance at NIH is an estimated 99%-1% (exact figures are required), which is an untenable, unhealthy and ultimately extremely expensive imbalance. If a theo-rist/modelist brings his/her own money from funds designated for theory to an experimental neuro-science lab, the danger is lessened that mathematical/computer modeling ends up subserving the sup- porting experimental-ist's sole jurisdiction. Such two-sided arrangement would be similar to NASA's National Research Council Associateship, which provides a pool of money for bringing in scientists with creative ideas to work in already estab-lished NASA laboratories. Approval presupposes mutual agreement by the recipient and contributor, and is administered by the impartial body such as National Research Council (NRC). It is suggested therefore that the MTC program establish (or share) such a system as recommended in Sect. 5.3.or creates a system similar to that of NASA/NRC.

4.6 Refine the Balance of Experimental and Theoretical Research

A healthy triad of experimental-, theoretical and joint proposals is not in place. The main sig-nifi-cance of funding theory would be that such programs establish existence of three classes of research grants, beyond the presently almost exclusively experimental grants. Experimental proposals will, of course, continue to be the mainstay. The class of experimentaltheoretical cooperative research grants, although still a rarity, already has some precedents. Funds for theory itself, would complete the full spectrum. Such a mechanism has long been in place for example in physics.

Suggested countermeasure: Purposefully use the MTC neuroscience program to establish the triad of ex-perimental-, theoretical and joint proposals. The practical significance of actively building up such a "triad" with checks and balances is potentially enormous. Its main strength lies in its pluralist nature. Experimental proposals are based on the doctrine in modern science that all questions to nature should rest on theoretical hypotheses which experimentation is supposed to verify or reject. One reason this doctrine is so frequently violated in neuroscience is that professional theoretical hypotheses cannot be explicitly de-manded if theory is nonexistent, in short supply or inaccessible. If theory is put on a supply and demand basis of the free market, every experimental-proposal can be judged by the importance of the theoretical hypothesis (freely chosen from the market) that the proposed experimentation is capable of answering. Also, use of animals will be better justified since market forces will en-sure that every (theoretical) investi-gation that can be conducted without the absolute necessity of animal experimentation, will be accom-plished by that much less expensive method.

For the overall NIH budget, establishing the proposed "triadic" funding strategy would actually provide a mechanism to effectively save expenditure by helping to screen out those animal research proposals whose theoretical foundation is less than professionally established. With all of the disadvantages, it is realized that the introduction of such "triad" (a) can only be gradual as MTC is still too young to support the whole experimental field, (b) must not be forced, as it will trigger even harsher turf-protection in-stincts (not just from pure experimentalists for whom such precedents are danger-ous, but also from pure theorists for whom the precedent of a theorist and an experimentalist working together is equally threaten-ing), (c) could be best accomplished by insisting on rewarding attempts with initial success in integrating theory and experimentation by cooperation. Even with only a few "success stories" created the trend will certainly spread in the experimental community, especially because of the overcrowding of experimental field with each approach yearning to gain a competitive edge. As for brain theory and modeling, the bene-ficial aspects are twofold. First, the theoretical and cooperative legs of such a triad provide a sound basis for (sometimes decades-long) efforts that so far have been conducted on devotion, under inappropriate and dangerous existential arrangements as appendages to experimentation. Second, with a balance of the-oretical, ex-perimental and experimental- theoretical proposals competing models and theories can be objec- tively evaluated by measuring their impact on experimentation.

4.7 Intensify Interaction of Experimental and Theoretical Research

Theory and experimentation does not at present have enough cross- fertilization. The difference in the intellectual and cultural milieu of biological experimentation and mathematical/computational the-ory, and the meagerness of mechanisms promoting their cross- fertilization greatly inhibits the process of scientific synthesis.

Suggested countermeasure. Create mechanisms for interdisciplinary access. Two particularly important possibilities appear feasible.Some independent funding agencies (e.g. Sloan Foundation) provide partial support for Sabbatical leave in certain fields of science. This author is in the process of trying to per-suade the President of the Sloan Foundation and officials of the McDonnell Foundation to extend such a sabbati-cal program to the field of neurocomputing. It is known that some distinguished workers (e.g. neurobiol-ogists) take sabbaticalleave to Depts. of Engineering or Computer Science (and vice versa) in order to synthesize these different disciplines. Such cross-fertilization is particularly important in order to establish Neurocomputing Ph.D. programs at Universities, which can only be accomplished if dif-ferent Departments (often in different Schools such as Engineering and Medicine) actively interrelate. While private foundations could set up such program, because of their typically small size there remains a need for a similar national program. It is recommended in Sect. 5.5.3. therefore that a sabbatical pro-gram be co-sponsored by NIH and NASA, to encour-age University scientists to attempt to practically utilize their ideas. Another need for interdisciplinary access is created by the importance of using some specialized equipment that is often required for neurocomputing. For instance, massively parallel com- puters, such as the Connection Machine, or supercomputers such as Cray, and other experimental facilities that are too ex-pensive either to establish or to run by regular University Departments. To this cate-gory belong equipment for parallel recording from many brain cells simultaneously by multiple electrodes or equipment suitable for computerized quantitative histology and imaging/modeling techniques (see point 5.5.3. of this report).

4.8 Facilitate the Link of Basic Research with Technological Development

Basic research and technological development aspects of neurocomputing are isolated from one another. Because of differences in goals, methods and support systems of neuroscience and research and development of parallel computing, these efforts are typically isolated even if physical proximity would permit and personnel actively seek such integration.

Suggested countermeasure. Seek and fund (preferably in a joint fashion with a complementary agency) projects that involve both aspects of neurocomputing. Encourage neuroscientists with theories and models of specific CNS subsystems to test their understanding by electronic implementation. In turn, workers with mathematical "neural net" algorithms and hardware implementation of parallel computing should be encouraged to check how similar or different are solutions actually employed by real neural networks. Since these aspects require different skills, such linkage is best achieved by cooperative research part-ners. A particularly useful mechanism for encouraging such cooperative efforts may be the launching of "Artificial CNS System" projects.

---

5.0 Conclusion and Recommendations

5.1 CONCLUSION: A US Civilian Neurocomputer Initiative from the Government
is Needed to Establish Coordinated Basic Research Foundation for
Neurocomputing

Neurocomputing research in the US needs a government supported and coordinated civilian pro-gram complementary to the DoD (DARPA) neurocomputing initiative. Such a program is necessary, for the following reasons.
(a) The DARPA initiative is not as successful alone as planned since geopolitical conditions have drastically changed.
(b) Civilian Government Agencies, such as NIH, NASA (and NSF) all have neurocomputing initiatives in a formative (planning) stage, but they are not coordinated with one another, all are presently subcritical in dimensions and face different structural difficulties imposed by the specific character of each Agency. Developing a synergist strategy by which these efforts would be coordinated, mutually reinforc-ing and balancing one another is deemed critical.
(c) A civilian program is necessary to estab-lish the basic research foundation that is vital for neu-rocomputing but is not provided by the DoD initiative. The basic research leg is presently missing, and thus the current twotiered structure of neurocomputing, based on defense and business, is inherently faulty and may lead to a repetition of the earlier cycles of Cybernetics and Artificial Intelligence. This diffi-culty arose because of the slow evolution of theoretical neuroscience. The problem has historical prece-dents and has geopolitical implications (see also Appendix).

5.1.1Slow Evolution of brain theory and modeling in Neuroscience

The crucial question of theory and modeling in neuroscience has been extensively treated else-where; cf. [123],[124],[125],[126],[109],[5],[127]. This author also contributed with reviews; [128], [36], [129]. Therefore, this section will comment on selected issues only.

Several stages in the evolution of modeling and theory can be discerned in all fields of science, including neuroscience. "Modeling" often commences with the most primitive expediency, in which numbers are used for hardly more than pseudo numerical ornamentation of data. The first real stage, phenomenolog-ical modeling, where the model is a quantitatively concise presentation of data is rather ubiquitous in all sciences. Later, phenomenological representations may evolve into conceptual modeling where the model is a specific quantitative elaboration of a concept. With broader ideas put into increasingly rigor-ous formal-ism that is endogenous to the scientific problem, theories of subsystems can evolve. Ultimately, with the emergence of concepts based on axiomatic and comprehensive foundation, couched in a homogeneous and genuine formalism (that is both powerful and general) there is hope for scientific theories. It is expected that neuroscience, just as other branches of natural sciences, will arrive at this advanced stage. Ideas and formalisms will compete for ex-perimental testing as in physics, where cor-puscular and wavetheories of light, or the Schršdinger versus Heisenberg approaches to quantum me-chanics provided useful alternatives.

Conceptual evolution is slow, since theory is a difficult goal to attain in any field of science. In neuro-science, the challenge is particularly serious since the two basic roles of brain theory and modeling (maximize rigor and minimize discontinuities in our under-standing) are almost hopelessly contradictory. Starting at the most natural level of neurons, maximiz-ing rigor drives investigation towards analysis; to synaptic-, membrane-, channel and ultimately molecular levels (thus the overdominance of molecular neu-roscience). On the other hand, the task of understanding CNS function at the level of behavior, or at least at overall per-formance of vast (sensorimotor) networks, drives research towards synthesis. One can hope that all pieces of information will "come together by themselves". But the history of science shows otherwise. Particle physics, for instance, had to heavily invest in theory to attain synthesis, and research on super-conductivity presently concludes in massive support of developing theories. Only theory can con-nect microscopic levels, such as the spin of a single electron, with emerging properties such as the resistance-free passing of current. With no theory, vast sums could be misspent trying to achieve super-conductance at room temperature that may not be possible for theoretical reasons. In turn, nuclear tech-nology, without underlying theory, would have been not only wasteful but deadly. With the "Decade of the Brain" already underway one wonders if neuroscience programs will succeed without an appropriate balance between experimental and theoretical sides of this discipline.

Centrifugal effects among various levels of analysis can be illustrated by pioneering examples in brain the-ory and modeling . The simplest assumption was to consider single neurons "all or none" elements [8]. If brains were computers, such elemental mathematical "understanding" of neurons would directly connect to a grand mathematical theory of the whole CNS; see information the-ory, [130],[131]. Moving along this track of equating brains with computers, theorists established simple synaptic "learning rules" [10], which later led to "neuronal" schemes of associative learning [11],[84],[132]. However, neurons were found not to be simple "flip- flops" but units displaying complex membrane potential wave-forms [13]. Also, diagonally op-posite organizational principles of brains and computers were revealed [133]. Levels of investigation, therefore, greatly dispersed. On one hand, information theorists were left with "patterns of excitation" of many cells [134],[135]. On the other hand, classical single cell electro-physiologists turned to electroresponsive membrane phenomena of cells and sy-napses [136]. In a mathe-matical sense great progress was made by finding closed-form analyti-cal equivalents for (passive) dendritic trees [14]. Measurements of many neurons contin-ued following in the footsteps of pioneers [137], although it was difficult to con-nect them to either the underlying anatomical structure or the activity-pattern of many neurons. Therefore, map-ping of neural structure and function continued along separate paths. Overall sensorimotor behavior was quantitatively described in terms borrowed from system analysis of gain and phase control [32],[33],[138],[139]. As these separate lines of research burgeoned, even sym-bolic referencing of today's many levels of investigation is impractical here. Nonetheless, we have to "put pieces to-gether" into models or theories. Is that task important enough to assume such a crushing burden?

Synthesis is an urgent, vital task, not only from a philosophical viewpoint ([109]) but also from the general vantage point of evolutionaryeconomic aspects of science. Presently, brain theory and modeling is a bottleneck both in System Neuroscience (in neuroscience) and Neurocomputing. This author made sustained efforts to represent sensorimotor brain theory and model-ing both in neuroscience; c.f. II. World Congress of IBRO, [140], Multidimensional Sensorimotor System Satellite (1984) and Workshop [2] at Soc. Neurosci. meetings) as well as in Neurocomputer meetings ([2] [140],[141],[142],[57],[39],[143] and also in the Editorial Boards ("J. Theoretical Biol-ogy", "Neuronal Networks", "Neural Computation"). Given the difficulty of the task, such individual efforts are insuffi-cient. To properly ground theoretical neuroscience in experimental neuroscience strong institutional backing is needed.

5.1.2 Historical Precedents: Cybernetics and Artificial Intelligence
Have not Incorporated Neuroscience

The first two attempts at creating brain-like machines, Cybernetics and Artificial Intelligence, waned after initial surges of success (see Appendix). One of the strongest reasons for this is that they could not and would not rely on "wet" brain research. Cybernetics ought not be blamed for a lack of neu-roscience basis as in the forties to fifties modern neuroscience was still in its infancy. Artificial Intelligence explicitly exempted itself from interaction with experimental neuroscience claiming that a reliance on brain research was not necessary. Criticisms that Artificial Intelligence fell short of promises and expectations and the very fact that the alternative approach of "Neural Networks" emerged point out that an attempt to create brain-like machines is unlikely to succeed without understand-ing the brain. Neurocomputer re-search is presently coping with similar historical dilemmas. There are, again, reasons for disassociating the "neurocomputing" strategy of creating brain-like computers from research of wet brain. There are more than equally strong reasons however for strengthening this precious but hitherto weak linkage. The first choice would permit rapid progress in technology cut loose from constraints of the much slower process of theoretical understanding of biological brains. The disadvantage is that "Neural Networks" may thus repeat earlier cycles experienced in Cybernetics and Artificial Intelligence. The strongest contention of this report is that the present "Neural Network" ap-proach to Neurocomputing has severely limited chances to succeed without establishing a strong bond with exper-imental neuroscience.

The above contention is illustrated by both negative and positive historical examples. First, attempts at creating the new technology of brain-like machines without relying on knowledge of biological brain (supplied by neuroscience) have been tried and shown not to have completely fulfilled expectations. For a positive historical example, it is selfevident now in retrospect that nuclear technology could not and thus would not be developed without establishing first its scientific basis, nuclear physics. With the scientifictechnological revolution of the nuclear age, it helped to create such a coherent ap- proach that a single group of academia (of physicists) was singularly responsible for development of both basic science and foundation of technology. With brain-like machines, however, responsibility for technol-ogy development ulti-mately rests with engineers, while providing the understanding of the biological brain (and making sure that this body of knowledge is not ignored) is the responsibility of neuroscientists. In the sixties, the U.S. launched an unprecedented effort of developing experimental neuroscience, which accordingly evolved into a burgeoning and thriving research enterprise reflect- ing the investments of many billions of re-search dollars e.g. from the National Institutes of Health. It will be obvious, if not now then in retrospect, that without putting the accumulated knowledge of the bio-logical brain into use, the technology of brain-like machines will be difficult to develop by Neural Network research. While it is still possible that "Neural Network Research", similarly to Artificial Intelligence or Cybernetics will ultimately opt for tech-nology- development cut off from neuroscience, but if such a de-cision will prevail it should only happen over clear warnings today by those who are responsible to society for having in-vested so much and for so long to experimentaltheoretical neuroscience.

The alternative and much preferred choice is to closely tie experimentaltheoretical neurobiological re-search to technological development of brain-like machines, for which strategy there are already several examples [36],[39],[56],[113],[144],[145]. This latter course also has disadvantages. Establishing such a bond is extremely difficult given major differences between neuroscience and neural net implementation for instance in the extent of mathematical formalism, culture, funding structure and even philosophy. Nonetheless, a synthesis of experimental and theoretical- technological aspects may be the only way to create a discipline from interdisciplinary explorations. While difficult, it should be pos-sible just as it was in case of the process of creating a flourishing nuclear technology. Interdisciplinary studies, between and beyond classical physics and chemistry, led to forging a new discipline of nuclear physics, with its own mathematical understructure (quantum mechanics). Establishment of the basic science could later safely serve as a foundation on which to build a new technology. A similar scientific technological integration, necessary today for putting neural net tech-nology on solid grounds of a basic science, is therefore not un-precedented only difficult.

The potential benefit from such a link is that it provides existence proof of the solution for the field of Neural Net research. Indeed, the presently most exciting three areas of research are "Brain like Machines by Neural Nets", "Superconductivity at Room Temperature", and "Cold Fusion". While some are controversial (since the latter two may not be attainable), Neural Nets certainly exist in the form of how nature perfected the biological brain. Opting for research safely following natural evolution entails two key resolutions, however. The first is to settle down to a much longer haul than most workers or agen-cies would like to commit themselves. The second is to carefully select those systems in the biological brain that provide neural network researchers with the best biological paradigms of neurocomputing.

A second set of historical arguments for emphasizing the necessity of a civilian governmental program for re-search and development of brain-like computers based on contrasting the organizational structure of earlier stages of development of computer industry with the new stage 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 essentially by the defense establishment. An almost exclusively technology- develop-ment was possible, since the mathematical basis of von- Neumann computers was "in the books": Boolean algebra; [131], [130]) and thus development of computers was solely a matter of technology required no basic re-search. The very recent stage of major evolution of computer in-dustry is the development of home (personal) computers. This revolution was driven by small en- trepreneurship, entirely in the industrial commercial domain, so well exemplified by start-ups in garages growing to be a multibillion dollar industry in a decade. Almost exclusively com-mercial organization of development was made possible, again, by not requiring basic research. (Moreover, personal comput-ing did not even require major technology de-velopment, since chips were readily available). Required was creative hardware assembly and innovative software devel-opment fueled by an explosively enlarg- ing commercial market.

Strikingly, neurocomputer development appears to spontaneously follow these two past successful trends although the situation and accordingly the needs are vastly different with "neural nets". The present organizational structure of neurocomputing is apparently intended to rest on two pillars; on one hand it would like to rely on the exclusively defense-oriented DARPA program, and on another it relies support drawn from small commercial startup companies (which follow the trend set by personal com-puting; mar- ket-ing creative hardware assembly of commercial micro-hosted offthe-shelf parallel boards wrapped in an innovative software package).

It must be made crystal clear by this report that this two pillared (de-fense and small business organizational) structure of neurocomputing is inherently faulty, since neither will support the neuroscience-oriented basic research without which sustained healthy growth is virtually impossible. Therefore, the pre-sent report argues for the firm establishment of a third (civilian governmental) pillar of neurocomputing that should take care of the needs of massive basic research. Also, it should contribute to the strengths of the existing two pillars as presently neither is suffi- ciently strong. As judged momentarily, the DARPA program may or may not succeed in obtaining from Congress the funding necessary for devel-oping the technology for DoD applications of neural nets and even if it did, it decidedly will not sup-port the neu-roscience basic research necessary for the scientific maturation of this interdisciplinary field. Likewise, the existing and slowly expanding - but still relatively very meager commercial market for neurocomput-ing will have unsurmountable difficulties in sustaining the neurocomputer revolution and has no interest whatsoever, much less any funding, for supporting basic science. Shaky financing aside, the present structure of neurocomputing leaves a potentially catastrophic gap between research & devel-opment 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 and it rests within a civilian governmental domain with agencies such as NIH, NSF and even NASA. It is argued in this re port that the present task is to get these civilian government institutions together for the purposes of not just supporting but actu-ally guiding neuroscience research into an active interrelation with (DoD- related) technology development in the field of neural network research.

5.1.3 Implication on Worldwide Competition: Europe and Japan Organize
Civilian Neurocomputing Programs

Both European and Japanese neurocomputer devel-opments are fostered entirely by civilian government organizations. In Germany, the Ministry for Research and Development, in Japan the Ministry for International Trade and Industry (MITI) seized initiative for neurocomputer development. In Germany the leading neurocomputer specialist is an engineer-neuroscientist; Dr. R. Eckmiller of Univ. of DŸssel-dorf; [146]), the president of the Japanese Neural Net Society is also an engineer neuroscientist; Dr. K. Fukushima of Univ. of Osaka; [147]. There would be much less need to consider involving civil-ian governmental institutions also in the US if DARPA could generate alone the funds for neural net tech- nology, or free enterprise could alone build up the mar-ket of neurocomputers, and neuroscience support-ing agencies (e.g. NIH and NSF) could firmly estab-lish funding structure for mathemati-cal/ theoretical/ computational approaches amid experimental re-search. However, Congress is more than a year late with the downscaled first batch of the DARPA pro-gram, most startup neurocomputer companies are still operating in the red, and NIH is yet to reward pioneers of mathematical/theoretical/computational neuro-science who are ready and willing to elevate theoretical neuroscience to a partnership of experimentation. Thus, it appears that a concerted effort of civilian agencies initiating a neurocomputer program that would complement DARPA's defense-oriented neurocomputing program is essential.

In a general historical sense, there would be less of a need for the US to consider emulating European-German and Japanese approaches to civilian governmental funding of neurocomputer research & devel- opment if the entire geopolitical sit-uation had not changed dramatically during the past year or so. It was true that the defense establishment could single-handedly underwrite the success of traditional computer development during World War II. In today's World, however, it seems safe to predict (even without the evidence of DARPA's problems in getting funds from Congress) that because of the present political cli-mate it is next to im-possible to guarantee funding of neurocomputing as a defense package alone. As the President of the International Neural Net Society (Dr. B.Widrow of Stanford University) raised this point at the de-fense-panel at IJCNN 89-WASH, reflexes of World War II are no longer automatic to-day. As he maintained, US is presently fighting WW-IV and not yet winning it. Dr. Widrow ex-plained that WWIII (the Cold War) was won without armed conflict just by the sheer economic force of the relentless arms race. However, in a further argument he pointed out; WWIV is already on and it is an economic re-search-development war in which the US is at a great disadvantage over Germany and Japan as its military ex-penditure is far higher. The President of the Neural Net Society argues that in an all-out competition with Germany and Japan the US may wish to adopt some of their strategy e.g. di-verting some of the classical military hardware funding to government-organized and sponsored research and develop-ment, in the style of MITI in Japan or the Ministry of Research and Development in Germany. Although the commission of this report does not seek nor its scope permits specific recommendations towards such national policies, it is interesting to consider that a U.S. Civilian Governmental Program (that complements, and does not compete with DARPA's similar program) would be helpful in providing a balanced civilian-de-fense neurocomputing proposal, presenting a strongly justified case and precedent for Congress to initi-ate a MITI style Research & Development program, in order to improve the US competitive posture in WWIV. With the proposal of ACTA (Advanced Civilian Technology Agency; a "civilian DARPA"; a bill sponsored by Senator John.Glenn, D-Ohio and seven colleagues) presently considered by the House of Representatives, the Civilian Neurocomputing Initiative is a package that should be at the top of the agenda of such a newly created agency.


5.2 RECOMMENDATION: Establish an NIH-NASA-(NSF) US Civilian
Neurocomputing Advisory Committee for Longterm Neurocomputer
Research Initiative and Coordination

The central requirement to implement the above recommendation is to identify the civilian governmental agency that could play the spearheading role in such an initiative. Such an agency should have a clear civilian profile yet strong and natural connections with the defense estab-lishment. Second, neuro-computers should be central to the mission of such an agency. Third, the agency should be able the con-join the technological and basic life science aspects of neurocomputing. While it is fairly obvious that NASA eminently fulfils the first requirement the second, that neurocomputers are of a central interest to NASA, may not be evident, although neurocomputer efforts are numerous at all NASA research facilities [116].

To support the second point it is argued (see also Sect. 3) that neurocomputers are the computers for future aerospace activity. Three most important features of neurocomputers, arising from their massively parallel organization, destine neurocomputers to be ideally suited for aerospace research. First, parallel organi-zation enables faster operation given identical weight and dimensional constraints. Secondly, parallel orga-nization makes neurocomputers hardware errortolerant. "Graceful degra-dation" is of utmost importance with systems exposed to known and unknown harmful in flight effects, with virtually no possibility for repair of even the smallest disability. Presently, some missions are completely disabled by computer errors that may be relatively minor. Computer technology of sixties (employed by the Shuttle) ensuring reliability by simply quintupling computer hardware is a less than optimal solution for this problem. Third, and per-haps most importantly, neu-rocomputers rely on "self-organizing" software that is orders of magnitude simpler than conventional computer software. It is already conspicuous that many mission delays and fail-ures are attributed to inevitable errors in such supercomplex software that is necessary for advanced opera- tions in aerospace research. The "self-organizing" nature of neurocomputer software ("netware") is also of great signifi-cance as it is the basis of autonomous intelligence, that may be the essence for unsupervised or very remotely supervised and thus delay-laden space computer systems. Finally, concerning the main mission of NASA (aerospace flight), it should be evident that atmo-spheric and space flight control could greatly benefit form an understanding of how natural evo-lution worked out biological "neurocomputers" for fast, precise and environment-adaptive control of navigation both in water (vestibulocerebellar control of swimming in fish) and in the atmosphere (vestibulocerebellar coordination of flight of birds).

Additionally, NASA is one of the very few government research institutions that can combine interest and investment both in technology as well as neuroscience research and development. In this regard it is for instance of great potential significance that NASA's existing re-search program has vested interest and significant capital investment and manpower specifically in research of biological subsystems such as the vestibulo cerebellum (nature's neurocomputer for motor coordination) while its main mission is tech-nol-ogy research and development for aerospace flight (see point 5.2).

Finally, while NASA is destined for a central role in neurocomputing among US Civilian Governmental Agencies its liaisons are building up with other agencies playing an active role in neurocomputing. Specifically, NIH (with a Study Section for Mathematical/Computational/ Theoretical Neuroscience pro-gram, headed by Dr. R. Nakamura and H.Lansdell) are presently looking into NIH-NASA cooperation in neurocomputing and NSF (with a special program for neurocomputing, headed by Dr. P. Werbos) is also fostering NFS-NASA joint programs in neurocom-puting. Thus, based in its existing life science program, NASA could initiate a neurocomputing program in a manner that technology devel-opment is coordinated with neuroscience research. This is a unique feature for the U.S. Civilian Governmental Neurocomputing Program, and also enables ties to NIH and NSF neuro-science programs yet it would possess built-in con-nections to technological applications both within NASA and with NSF neurocomputing instrumentation programs pursuing related interests.

Based on the above rationale, NASA appears to be ideally poised to seize the initiative for creating, to-gether with NIH and NSF, a US. Civilian Governmental Neurocomputing Advisory Committee. The committee would serve as an organizational umbrella, both to plan and coordinate the preparatory stage of neurocomputer activities at NIH-NASA-(NSF) as well as to jointly prepare a Civilian Neurocomputer Initiative to obtain (joint) new funding for such activities from Congress, in coordination with DARPA. Many if not all of the recommendations of this report are fairly complex, thus their imple-mentation would have to be followed up. For practical purposes of overall coordination, it appears essential to have such a committee as an umbrella mechanism. Specific recommendations towards attaining the above goals are the following.

With NASA's initiative establish an Interagency Governmental Advisory Committee for U.S. Civilian Neurocomputing. A useful existing mechanism to create such a Committee appears to be the "Interagency Working Group of NASA-NIH-(NSF)". It is recommended that this report be put on the agenda of this Working Group for discus-sion. If the issues raised in this report are considered worthy of more substantial evaluation, the proposed "Interagency Governmental Advisory Committee for U.S. Civilian Neuro-computing" be established.

If NASA declines the role of initiating this organizing and coordinating body, either NIH, or ACTA (if established) could be the sponsor and administrative agency for this initiation. Alternatively, one of the National Laboratories could suitably play the role of initiation of such Committee.

Having created an interagency committee, under its guidance and coordination NIH-NASA-(NSF) would proceed with developing separate (but mergeable) programs in a preparatory 2-year stage. Such a transition is deemed necessary, since presently neurocomputing is not fully established either at NIH or NASA. These neurocomputing programs are still too young and immature to be wedded now. Thus, the recom-mended general strategy is to structure initially separate but linked and strongly coordinated programs such that they competition is minimized and cooperation is maximized. Each program would establish a lever-age, used by the specific Agency, for securing a share in new funds jointly obtained in the future for civil-ian neurocomputing. It is suggested that a coalition of DoD Agencies (with DARPA) and Civilian Agencies (with ACTA) stands a good chance to submit to Congress a balanced new national program for Neurocomputing. With the basic structures and mechanisms established in the preparatory 2-year stage at each civilian agency, they could successfully ask for, and implement, a Civilian Neurocomputing Program as part of a national neurocomputing initiative.

5.3 RECOMMENDATION: Broaden MTC Study Section to an Overall
NIH Review Board in Order to Allocate Centrally Created
Seed-Budget for MTC Research

The MTC Study Section ideally poises NIH for a neurocomputing initiative. It is already evi-dent that beyond Institutes that jointly launched this program (NIMH and NINDS), other Institutes, for instance the Institute on Deafness and Other Communicative Disorders; NIDCD are also coping with the challenges and opportunities of introduction and broadening MTC research. Given the range and seri-ousness of the difficulties with MTC proposals (see Sect. 4) it is unlikely that each Institute could re-solve these problems separately. As one of the key problems is the subcritical dimension of MTC, it would certainly be ill-ad-vised to scatter efforts and resources. Therefore, this report recommends that the MTC Study Section be broadened; to be open to all neurocomputing- related proposals from all NIH Institutes.

The Study Section would receive, therefore, proposals addressed by PIs directly to this program, as well as proposals from all Institutes that decide that a particular proposal of MTC nature could be best re- viewed by mathematical/theoretical/computational experts on this Study Section. While this openness of a specific Study Section to NIH proposals, in theory, is already possible, the present recommendation goes beyond this potential in two aspects. First, it makes deferral of proposals of MTC nature to one Study Section from all Institutes the rule rather than the exception.

Recommendation: Since such established channels from all Institutes would increase the volume and steadiness of proposals flowing into the MTC Study Section, in concert with Sect.4.1. and 4.4. it is rec- ommended that the Study Section be made permanent.

Second, in addition to the scientific rationale for infusing experimental neuroscience research with the-ory, a financial incentive should be created for the Institutes to do so. This should not only permit but encour-age the use of experts on this MTC Study Section by all Institutes, without the fear of particular Institutes of loosing such promising new research initiatives.

Recommendation: MTC Study Section should be appropriated a budget; an amount that NIH centrally designates for new MTC- Neurocomputing related research. Once a proposal gets funded from this pool, that part of budget (and administration of program) should immediately go to the Institute that sent in the proposal. Thus, neurocomputer research should be financed by new money diffused to existing Institutes by the MTC Study Section mechanism. The existence of such program should be a reason for NIH to re-quest new funds, as recommended and supported by the Interagency Committee. Thus, this program will constitute NIH's "seed" planted in a preparatory stage for a later US Civilian Neurocomputer Program.

This "nursery" program for MTC and neurocomputing will be very popular. It will be seen by all Institutes as a mechanism by which they can (a) ensure adequate review to proposals that are notoriously difficult to evaluate, (b) establish new but perhaps risky and controversial research lines that the Institute wants but for structural reasons cannot launch, (c) procure additional new funds to the Institute. An ad-ditional mea-sure would further increase the likelihood that Institutes send the best and most competitive MTC- neuro-computing proposals:

Recommendation: Reviewers of MTC should be named by different Institutes, in proportion of funds awarded to them from this pool. For coordination and balance purposes, appointments should be con- firmed by the Interagency Committee. This structure (similar to the one proposed for NASA in Sect. 5.4) will further facilitate that initially separate neurocomputing review boards of these agencies (and preferably the whole programs) would be merged once the US Civilian Neurocomputer Program is established.

Recommendation: It is estimated that $12 M(illion) per year, set aside for the purposes of MTC is suit-able for the program. This is almost comparable to the dimensions of neuroscience-based neurocomput-ing spending in Europe and Japan (see Appendix) and is hardly more than what a single contractor to NASA proposes for neurocomputer technology development alone. While the current ratio of funded experimentaltheoretical proposals at NIH is not known, it is estimated that even with an MTC budget as recom- mended, spending on theory will continue to be a tiny fraction of the NIH budget.

5.4 RECOMMENDATION: Establish NASA Organization and Seed- Budget
for Neurocomputing to Parallel that of NIH

The main difficulty in coordinating NIH and NASA neurocomputing programs is, that NASA has yet to establish an organization and supply of funds designated for neurocomputing. As current ac- tivities and future programs envisioned for neurocomputing by NASA and its contractors are in a formative stage, it is recommended that (a) the structure parallels that of NIH such that an interaction and pos-sible merger is facilitated, (b) the program be guided and coordinated by the Interagency Committee, (c) be aimed at establishing a seed program that will be used beyond a 2-year preparatory stage as a leverage for taking part in a US Civilian Neurocomputer Program.

5.4.1 Create NASA Neurocomputer Advisory Committee and Program
to Integrate Neurobiological Life Science Research with
Neurocomputer Technology Development

As currently there is no organization at NASA for coordination and longterm scientific planning of neurocomputer research, it is understandable that ongoing activities are scattered, and the Life Science program does not actively participate in existing and planned largely technology-oriented neurocomputer research and development (the "Strategy for Space Life Sciences" Report [121] makes no specific mention-ing of neurocomputing).

To ensure that fundamental scientific and organizational aspects of neurocomputing programs are sound and balanced, a NASA Committee should be created. This coordinating and organizing body should (a) participate in and be guided by the Interagency Committee, (b) should plan and organize neurocomputer research planning at NASA and coordinate it with similar activities with other agencies, (c) ensure, by in-volving both codes RI and code Sl into planning, that Neurobiological Life Science Research compo-nent and Neurocomputer Technology Development components are balanced, (d) structure neurocom-puter re-search at NASA in a two-staged manner: creating "seed"-programs from its own funds for secur-ing a share in a later US Civilian Neurocomputer Program, to be financed directly from Congress.

5.4.2 Organize Short Term Neurocomputer Technology Development at
NASA by Allocating Neurocomputer Component to Specific Missions

A strategy was recommended such that NIH ensures that theory infiltrates the existing traditionally purely experimental approaches. In a parallel fashion, it is recommended that the shortterm goals of neu-rocomputer technology development are achieved by infiltrating existing specific missions with the pro-jected use/development of this new technology. While virtually all existing and planned missions are com-putation-intensive and thus neurocomputer- related, organization of specific missions should establish the extent to which shortterm technology development, necessary for its success, is appropriate. This strategy will ensure that at all times the actual "market conditions" will determine the shortterm expenditure on this new technology for every contractor Ð just as market conditions have al-ways determined the use of traditional computers.

5.4.3 Organize Long Term Neurocomputer Basic Research around
Integrative Artificial CNS System Projects: E.g. Establish a
Neurocomputer Laboratory based on the Artificial
Vestibulo-Cerebellum Project at NASA-ARC

The longterm basic research component of a goal-oriented development project (e.g. nuclear physics necessary to the "Manhattan project", or DNS research necessary for the "Human Genome pro-ject") cannot be "market pulled". Likewise, laboratories have to be set up for that kind of longterm neu-rocomputer research that will institutionally ensure that the neuro and computer-sides of this new field are pursued in a concerted fashion. Just as natural evolution guaranteed that in developing our brain the problem and its solution interacted in a longterm and intensive manner, it is recommended that we take a similar approach to neurocomputer development: Organize neurocomputer basic research around integrative Artificial CNS System projects pursued in laboratories set up for these specific purposes.

This idea is not entirely new, and the success of some precedents testifies to its basic viability Ð although new recommendations are forwarded to substantially improve on organization. An earlier project that can be called "Artificial Cerebellum" [148] was highly productive, as one of the most success-ful present neu-rocomputer project is that of aimed at an "Artificial Retina" [122]. Such "Artificial CNS System Projects" by their nature unite the neurobiology and technology-side of neuro-computing, and by constantly interre-lating the natural and artificial solution keep the approach honest and the solution guaranteed by upholding an "existence proof". It is expected that such projects aimed for instance at artificial analogue of the Vestibulo-Cerebellum, Hippocampus, Olfactory bulb-Pyriform Cortex, Colliculus-Visual Cortex, etc, when established in an environment that ensures access to both neurobiology and hightechnology, will flourish even more than an existing precedent. Dr. Albus launched his "artificial cerebellum" project year before neurocomputing appeared, and the host-institution (Bureau of Standards) did not provide access to neurobiology. Thus, his research did not involve the vestibular system, for instance.

It is estimated that to launch such "Artificial CNS System" projects would require an expenditure less than what was necessary, for instance, to establish the Vestibular Research Facility at NASA- ARC.

The recommended organizational principle can be best illustrated in this report by an outline of an Artificial Vestibulo-Cerebellum project to be pursued at lab in NASA-ARC. The present state of art of neurocomput-ing at NASA-ARC is characterized by virtually all elements of an integrated neurocomput-ing effort such as envisioned by an Artificial Vestibulo-Cerebellum Project. This fact indicates the time-liness and importance of this issue as well as the preparedness and excellence of individuals and groups for potential in-volvement. The present fragmentation is the result of two factors. First and foremost the hierarchical or-ganizational structure of this research center does not spontaneously facilitate such inte-gration. Secondly, spontaneous coalescence and explicit efforts for synthesis could not break through and materialize in or-ganizational integration. It is proposed that rather than attempting an integration of the organizational structure upfront, parties desirous of an integrated effort define and launch integrative research projects first The project itself will then coalesce any structure that may emerge as necessary and justified.

From an organizational viewpoint, all existing elements of neurocomputing related research point to an Artificial Vestibulo- Cerebellum Project that is both eminently feasible and would provide a unique plat-form for any desired integration. It is likely that a life- science-based Neurocomputer Laboratory seeded at NASA-ARC, where such artificial vestibulo-cerebellum could be built, could serve as an instrument con-necting and consolidating ongoing but isolated neurocomputing-related efforts. The positive reinforc-ing cooperation of such a laboratory with research of advanced mathematical neurocomputing paradigms, such as Sparse Distributed Memory Dr. Kanerva in Dr. Raugh's group at RIACS is more than evident, espe-cially since that well-developed model has an obvious relationship to biological cere-bellar networks. While RIACS is actively engaged in hardware implementation of the mathematical paradigm the possibility of an in-house electronic implementation at NASA-ARC appears to be helpful since such would tie-in with both ongoing physiological and morphological research of the vestibulo- cerebellum and also the NASA-ARC. Artificial Cerebellum is destined to lead to actual flight control prototypes. The potential inherent in such cooperative liaison with Dr. Lum's group (RI) is equally prominent, as cerebellum modeling activities by C.Miles are perceived herein as an essential part of such an integrative approach.

A nucleus for an Artificial Vestibulo-Cerebellum project seeded in the Life-Science Division Space Research Directorate of ARC is demonstrably synergistic with the Vestibular Research Facility (D; helping to elevate it into a uniquely equipped laboratory, e.g. by planning to provide hardware for realtime anal-ysis of multi- electrode recordings by means of massively parallel neurocom- puters; c.f. 5.5.4.1.). Also, a neurocomputer laboratory would be synergistic and mutually supportive with neu-roanatomical research (Sect. 5.5.4.2), by providing essential mathematical assistance and soft-ware/hardware contribution to well established efforts in discerning mathematical computational prin-ciples from anatomy. Such synergy could help the realization of the potential inherent in a Biocomputational Center envisioned by its director Dr. Ross to become a cooperative enterprise. In turn, a life-science based neurocomputer laboratory not only would be open to technologically oriented efforts (A-C) but would certainly develop increasingly strong ties with them; e.g. by providing novel neurobiological applications for parallel computing efforts, and thus connecting "dry" neuro-computer development to "wet" neuroscience research. This synergy would be beneficial to draw support for life-science based neurocomputing from sources beyond the means of code OSSA funding (potential sup-port from OAET). This combination of life science based yet "hightechnology friendly" neu-rocom-puter R&D, existing in a scattered form at NASA-ARC, is diffi-cult to find among institutions presently en-gaged in neurocomputing Worldwide. Realization of this out-standing opportunity for a unique life science based conglomeration is undoubtedly difficult, requiring both or- ganizational support within NASA as well as funding support from resources external to NASA (e.g. NIH, see 6.). Because of the unique concentration and composition of neurocomputing R&D that would result from such integration, from a scientific point of view the goal is deemed well worthy of in-vesting siz-able efforts and taking those political and existential risks that such organizational task in- volves.

The above co-fundable project provides an example of research- oriented cooperation of NIH(NIMH) and NASA. Given the interest of the NSF neurocomputing program (under the directorship of Dr. Paul Werbos, inventor of the "back-propagation" neurocomputing paradigm) in technological implementation and electronic application of neurocomputing algorithms, it is quite conceivable that NSF could be ap-proached to co-sponsor of such a project. This assumption is reinforced by the knowledge that NSF is actively seeking an organizational interaction with NASA at the present time. The scientific project oriented approach to interagency cooperation in neurocomputing gains further importance in light of diffi-culties in crafting organizational ties among such agencies. Nonetheless, specific examples and recom-mendations are given below that could be achieved by relying on organizational ties alone. These rec- ommendations rest on the possibility of utilizing the Research Resource Program of NIH for elevating existing NASA ARC laboratories to the level of National Facility.

The fundamental scientific justification of such project is that the cerebellum is regarded ever since the nineteen sixties; [149] as the part of the brain which is best understood, and whose function (sensorimotor coordination) is the clearest. The author contends that vestibulo-cerebellar neuronal nets of-fer among the best and clearest examples of an actual well-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 [150]. Vestibulo-cerebellum first ap-peared in the course of evolution with sharks, pro-viding a performance-margin ensur-ing a remarkable evolutionary survival of this epitome of navigation. Vestibulo-cerebellum attains an out-standingly high proportion of the brain with birds. Accordingly, they are masters of flying rapidly changing their body-shape to adapt to turbulent conditions. Flight is controlled by an "on-board, realtime" neuro- computer that relies on errortolerant, gracefully degrading, massively parallel neu-ral network. In addition, it need not rely on supercomplex software that character-izes, and causes most of the break-downs, of present-day serial computer systems. Technological im-plementation of the nature-perfected vestibulo-cerebellar neurocomputer should be obvious.

With the knowledge that nature developed the vestibulo-cerebellum during evolution for fast sensorimotor control and that the vestibulo-cerebellum, as the most ancient part of the cerebellar system, becomes domi-nant with birds for fast and precise flight control it is virtually guaranteed that development of such an in- strument is fully in line with the purposes of NASA-ARC. Accordingly, an Artificial Vestibulo-Cerebellum prototype will quickly find its way to specific applications in novel flight control systems. It may be mentioned that flight control will become particularly difficult not just far into the future (when it may become a "show stopper" for the space-plane) but is almost unsurmountable with airplanes which were not designed for flying, but were designed for being invisible. Thus, the F117 is known as "Wobbly Goblin" and the B2 (Stealth) is subsonic as stability problems may not be entirely resolved. Biologists will recall that nature successfully resolved comparably formidable problems of making birds the masters of flight, although they developed from bodies that were not originally meant to fly (terrestrial Dinosaurs). While learning nature's neurocomputer secrets for flight control might not yield 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.

It is emphasized that such an Artificial Vestibulo-Cerebellum Project will start being very useful long be-fore its payoff in strategic high- tech applications will be evident. This is important since, because of complexities of biology, presently only preliminary knowledge is available on what mathemati-cal prin-ci-ples vestibulo-cerebellar neural networks operate. A specific subsystem is relatively well known however. The vestibulo-cerebellum is an integral part of gaze-stabilization systems, such as eye and head-move-ments compensating for displacements of the body and thus ensuring a steady gaze [151]. Pursuit and saccadic eye movement systems provide perhaps the fastest and most precise biologi-cal example of targettracking and interception [152]. Because the overall function is so important and well-defined, and because of the underlying biological mechanism is confined and relatively simple, the gaze system has been the target of intense research. In gaze systems many specific solutions by nature can be found for problems that are very difficult for today's engineers to resolve. One of the most clearly identified such subproblem is "sensory fusion", i.e. the mathematical theory of how to integrate information emanating from disparate sensory modalities. In neuroscience it is evident that not only information arising from all vestibular semicircular canals converge on individual neurons of the vestibular nucleus, but any cell also integrates different time-derivative signals arising from the otolith apparatus [106]. Thus, a neuroscience-neurocom-puter approach to Artificial Vestibulocerebellum could target such subproblems first, and on successful resolution proceed to larger- scale problems and applications.

Secondly, there is an even stronger mathematical rationale for seeking analyses of specific natural neu-ro-com-puting systems such as the vestibulo cerebellum. Neurocomputing will succeed, or fail, depend-ing on whether it can meet two crucial challenges. The easier task is the development of the suitable technol-ogy. The more difficult is to discern (and/or develop) the mathematics intrinsic to neural net function. It was argued elsewhere [39] that a departure from the algebraic mathe-matical lan-guage of traditional serial computers [8] is long overdue. An emerging geomet-rical school of brain theory would also have to im-prove upon its classical mathematical formal-ism of Cartesian vectors [11], as evidence accumulates that neural geometry is non Euclidean [143],[97],[98],[95]. These methods go far beyond traditional system- theoretical studies of gaze. A recent study [56] is very telling from this point of view: the authors argue strongly in favor of a drastic change from the earlier approaches to"Distributed Parallel Processing in the Vestibulo-Oculomotor System" since earlier models"have generally taken the form of black-box diagrams (for example, [153]) representing the flow of hypothetical signals between idealized sig-nal-processing blocks. They approximate overall oculomotor behavior but indicate little about how real eye- movement signals would be carried and processed by real neural networks". An alternative concept, inherently based on distributed parallel processing, was proposed a decade ago. Tensor network theory was de-vel-oped for explaining the function of massively parallel neuronal networks in terms of trans-formation of multidimensional generalized vectors [154]. Implications of this princi-ple on the-ory and modeling of gaze reflexes, such as the vestibulo-ocular reflex, were immediately pur-sued; as evidenced by the Chapter "The vestibulo-ocular 'reflex' as a tensorial response" in [112]. Similar to other new shifts of scientific axioms, the approach to vestibular system as a distributed parallel processor encountered a delayed re-sponse. Nevertheless, through a decade, there was much progress in modeling gaze, with a clear trend towards multidimensional approaches; both using extrinsic vectors [42],[43],[54],[56] or intrinsic generalized vectors (tensors); [112],[105],[155],[156],[77],[129],[157],[2],[158],[159],[160],[161],[162], [163],[113]. As for tensor network theory, the fundamental principle of the CNS as a distributed parallel processor that uses a vectorial language was experimentally substantiated in general [164]. The theory has evolved into a cohesive approach [86],[109]. Different generalizations of tensor network theory have also been attempted [89],[165]. Most importantly, the approach concluded in ex-peri- mentally verifiable quantitative theoretical hypotheses that were tested and confirmed both by inde-pen-dent workers [88] as well as in collaborative experimental research [166],[167]. Thus, vestibulo- cerebellar research, together with an advanced mathe-matical- theoretical analysis, is of central significance even before technological applications of neural networking for flight control, in the stage dominated by neuroscience research.

5.5 RECOMMENDATION: Converge NIH-NASA-(NSF) Parallel Neuro-
computer Organizations by Cooperative Agreements between Civilian
Governmental Agencies

The general strategy recommends a 2-year preparatory stage with NIH and NASA establishing parallel organization and "seed" programs as above. During this period, the MTC Neuroscience Program Administration at NIH and NASA's Advisory Committee on Neurocomputing, both guided by the Interagency Advisory Committee, coordinate actions in the preparatory stage and elaborate specific plans for their convergence or merger. Below, recommendations are made for such convergence; starting the list with the most longterm and difficult task and proceeding towards the most imminent and least prob-lematic recommendation.


5.5.1 Use NIH-NASA-(NSF) Neurocomputer Seed-Budgets to Request from
Congress New Funds for a Joint US Civilian Neurocomputing Program

As stated at the outset, this is the most important task both for guaranteeing a future of Neurocomputing different from that of Cybernetics or Artificial Intelligence, as well as for all the specific organizational plans that lead to this goal. This is the main theme throughout this report, thus no separate arguments are listed here rather some practical timing aspects are considered.

Essentially, time would be right now to come forward with a "civilian booster" of DARPA's momentar-ily not fully successful defense neurocomputer initiative. It is considered a great loss of time and momentum that not only its civilian complementary is not ready, but according to the assessment of this report some preparatory period is necessary for NIH-NASA-(NSF) for establishing their programs based on which a joint initiative can be built. Given the agenda for a preparatory period, the tentative 2-year win-dow is con-sidered the minimum during which the envisioned goals can be realistically fulfilled. Therefore, the report requests proceeding with recommendations leading to the preparatory stage with the utmost expediency, and to seek ways by which recommendations can be implemented in a parallel, rather than sequential order.

5.5.2 Facilitate Interactive Research Proposal Evaluation by Merged
Review Board for Allocation of Merged or Separate Funds for
Neurocomputer Basic Research

Implementation of this recommendation can be phased in long before the new financial basis of a joint initiative is established. Just as different NIH Institutes are experiencing problems with peer-re- view-ing neurocomputer proposals, it is expected that NASA will have similar, if complementary, diffi-culties with neurobiology and mathematicaltechnological aspects of proposals. Therefore, as already re-ferred to in Sect. 5.3, NIH and NASA should consider merging their review boards and possibly funds allocated for neurocomputer basic research. It would be the responsibility of the overseeing Interagency Committee to maintain coordination and balance of the two counterparts. For further ensuring evenhand- edness, it is suggested that NASA matches the yearly $12M fund recommended for NIH neurocomput-ing basic re-search, resulting in a size of the civilian neurocomputing initiative that is roughly equivalent to DARPA's defense based neurocomputing program. Since an early joint review-board and joint "seed" funding would be an instantaneous platform and justification for aiming at a US Civilian Neurocomputer Program funds from Congress, it is suggested for consideration whether this recom-mendation could be phased-in during the first or second year of the preparatory period.

5.5.3 Create Mechanisms to Use Manpower in an Interactive Joint Fashion:
Establish Universities Neurocomputer Research Association to Administer
Interchange, Sabbatical and Conference Programs between NASA and NIH

This recommendation can be implemented almost immediately, and requires only a minimal amount of organization, as a mechanism already exists at NASA. It is recommended (see also Sects.4.7 and 4.8) that NIH and NASA establishes a joint mechanism fashioned after NASA's Universities Space Research Association (USRA), either as its division (with NIH participation), or as a similar organization set up separately. This mechanism could very efficiently serve the purposes of creating a pool of re-searchers suitable for peer-review system, manpower interchange for cooperative research, sabbatical and conference programs.

5.5.4 Utilize Intramural Facilities in an Integrative Joint Fashion by NIH-NASA

This recommendation is similar and related to the above, and can also be implemented shortly and with minimal organizational effort. The interactive use of manpower, facilitated by the above rec- ommen-dation, also serves this goal of integrative utilization of intramural NIH and NASA facilities by personnel of the other agency. Beyond manpower exchange, however, the joint utilization and further development of equipment is also a distinct and possibility.


5.5.4.1 Use Vestibular Research Facility at NASA-ARC to Link Neuro-
physiology to Systems Modeling and Neurocomputer Analysis

Suggested Project: 3D Skeletomuscular Systems' Modeling
Suggested Project: Multidimensional Geometry of CNS Revealed by Multielectrode Electrophysiological Analysis

The above recommendation is based on the immediate possibility of extending the use of the Vestibular Research Facility at NASA-ARC to new neurocomputing-related projects such as (a) three di- mensional skeletomuscular system's modeling and (b) the investigation of the multidimensional functional geometry of CNS as revealed by multielectrode electrophysiological analysis. Both projects are eminent examples of the type of research that can most efficiently be performed at a National Facility rather than in University Departments. The rationale is simple, as both projects require equipment of the kind that is normally beyond the means of typical university-based research laboratories, and their estab- lishment, maintenance and regular use requires an integration of physiological-, mathematical-, computer science ex-pertise. It is specifically pointed out that Three Dimensional Skeletomuscular Modeling could be most ef-ficiently achieved by developing and putting into use modern computerized body-imaging technologies (see point.4 of 2.1.1). As for multielectrode analysis, both the electrode-system as well as the neurocom-puter envisioned for fast on-line data analysis are hightech approaches both requiring significant invest-ments and yielding an experimental setup that can be most efficiently used by a network of collaborators.

The approach of generalized coordinates intrinsic to CNS function requires an anatomical data-base. While coordinate-components of head acceleration are measured by the vestibular semicircular canal system whose spatial (boneembedded) orientation-directions can be well established [158],[77], some functional coordinate systems are changeable during movements; e.g. oculomotor and neck-motor muscle-co-ordinates [156],[112],[129],[75],[166],[161] and especially limb muscle coordinates [88],[160],[162]. Quantitative application of the tensorial neurocomputer-paradigm to actual neurobiological sys- tems re-quires, therefore, both the anatomical measurement of such data, as well as the construction (in the tempo-rary absence of suitable biomechanical models) of computer models of skeletomuscular appa-ratus [168],[159]. In an ongoing study of the vestibular eye-head coordination in the monkey, a skeletomuscu-lar model is to be used to quantitatively account for eye and neck muscle pulling directions changing with gaze movements. Such a model of the monkey eye-head system follows similar earlier modeling studies in the cat [168] and the monkey [169].

In turn, at the level of neuronal arrays, it is also a distinct possibility that the multidi-mensional functional space (spun over the firing frequencies of n neurons) is endowed with a structure which may not be re-stricted to a Euclidean geometry (see an elaboration of this argument in [57]. Indeed, the experimental establishment of the metric tensor of such neural n-spaces is a centrally important task for neural net the-ory. Geometries, by definition, are characterized not by a single point, but re-lationships among several points (c.f. triangulation in geodesics). Thus, measure-ment of the firing fre- quency of a single neuron will not suffice to characterize the structure of a multi-dimensional space; multielectrode methods (see [61]) are needed to map point-relationships in the n-space. In a theoretical study [57] specific suggestions were for-warded to use matrices of correlation-coefficients (as the covariant metric tensor) for calculation of both the co and contravariant tensors of the typically curved n-space. Research of the relationships of (Euclidean) external physical spaces to internal neural geometries is a major challenge for both experimental neuro-science and neural network theory. The vestibulo-cerebellar system is also among the best actual neural systems in which many of the different types of investigations necessary to attain such important goal are already in progress.


5.5.4.2 Utilize Biocomputing Center at NASA-ARC to Link Morphology
to Computer Modeling to Discern Neurocomputer Mathematics

Suggested Project: Fractal Structural Geometry of Neurons and Biological Organelles Revealed by Computerized 3D Histology and Modelling

The overall geometrical function of the vestibulo-cerebellum (by developing an internal model of the external physical geometrical spacetime relationships via multisensory mapping) provides impor- tant in-roads to the analysis of neural geometries involved in such CNS function. The physical geometri-cal relationships among sensory and motor coordinates (such as vestibular canal orientation and eye and neck muscle pulling directions) are fairly simple types of such geometries. Mor-phological and physio-logical analysis of neurons of the vestibular nucleus, or of the cerebellum, reveal much more sophisti-cated noneuclidean neural geometries (intimately related to the mapping of the spacetime manifold that is customar-ily described externally as characterized by Euclidean geometry). At a single neuronal level, the dendritic "trees" of brain cells obviously reveal a "natural" geometry suspected to be fractal [96] a conjecture substantiated recently [98].

As pointed out earlier in this report, mathematical analysis of fractal morphology of CNS structure is a rapidly unfolding field of research [93]. The difficulty in gearing up for future re-search is twofold. First, such a research activity requires the close integration of mathematical analysis with modern morphological techniques. Second, many of these technologies are presently being devel-oped. Such technology devel-opment, for instance, for creating the morphological data-base necessary for fractal analysis, could either be envisioned in cooperation with a traditional university department [70]. In this case, the University would have to allocate significant resources to such research, but would emerge as a rather unique center of this modern branch of research. Alternatively (and preferably), National Facilities could be established where both the technology development is much easier, and such facilities could be shared by a rather large number of collaborative scientists. The NASA-ARC Biocomputation Center, where many if not all of such computerized techniques have al-ready been developed, or are being developed for 3-dimensional visualization and computer simulation in the life sciences [120] lends itself as an ideal contender for being brought up to the level of such a National Facility.

---

APPENDIX

Neurocomputing Credentials of the Author

The author of this Report is Dr. A.J. Pellionisz (46), Professor of Physiology and Biophysics at New York University Medical School. He holds degrees in Electrical Engineering, Neuroscience, and Computer Science. He invested two and a half decades into neuronal modeling and brain theory, in the field of what is now called "Neurocomputing". He is an author of well over a hundred publications, articles, book chapters and books in this interdisciplinary research, originator of a mathematical brain theory ("Tensor Network Theory"). The theory resulted in a US patent "Sensorimotor Coordinator" as well as in testable quantitative predictions which have been experimentally confirmed by collaborative as well as in-dependent research. The theory generated a sizable and steadily spreading followership, and is now inter- nationally recognized by the Alexander von Humboldt Prize for Scientific Research by Germany 1989.

The author is heavily concerned with organizational aspects of neurocomputing, especially regarding its connection with the neurosciences. He is a founding member of the Editorial Boards of Journals of "Theoretical Neuroscience", "Neural Networks", "Neural Computation" and "International Journal of Neurocomputing". He was Chairman of a Symposium and a Workshop relating to this area in the Annual Meeting of the Society of Neuroscience (1984, 1986), organized a Symposium at II. World Congress of International Brain Research Organization (1987), organized neurobiologically oriented sessions at the IEEE Neural Net Conference (1988, 1989), and was the Program Chairman for the International Joint Conference on Neural Networks (1990). It can be interpreted as an acknowledge-ment of his personal contribution to and his general concern about this field that he was nominated for President of the International Neural Net Society in 1989.

As a Senior Associate of the National Research Council, the author is about to be involved with Neurocomputation-related research efforts at NASA Ames Research Center in San Francisco Bay Area. Since both NIH and NASA (as well as NSF) are presently exploring the possibility of a coordi-nated effort in neurocomputing by civilian U.S. government agencies, the author was jointly commis-sioned by NASA and NIH to prepare this "think tank"type concept report, leading to specific recommendations of how such initiative could be launched.


Background Information on Neurocomputing

A BRIEF W5 OF WHO, WHAT, WHY, WHEN, WHERE IN NEUROCOMPUTING
THE EMERGENCE OF THE ORGANIZING POWER OF NEUROCOMPUTING

What is Neurocomputing? Neurocomputers are new, non von- Neumann type parallelly and not se-ri-ally organized computers that mimic the functioning of the biological brain. They implement brain-like functions based on the massively parallel organization of neural networks in the central nervous system. Being parallel machines, neurocomputers are fast errortolerant (degrade gracefully with damage) and re-quire a new type of much less complex software. They can perform highly complicated tasks that are based on processing very large numbers of parameters at once. Such functions are the recognition, generation, coordination, prediction of phenomena composed of very many components. In a historical sense, Neurocomputing is the third wave (after Cybernetics and Artificial Intelligence) of concerted attempts to create brain-like machines.

WHO CONTRIBUTED WITH WHAT TO CREATE NEUROCOMPUTING?
A RETROSPECT OF CYBERNETICS, ARTIFICIAl INTELLIGENCE AND NEURAl NETS

"Cybernetics" was the brainchild of MIT Professor Wiener, who published an epoch-making book in the late forties [170]. Cybernetics was made possible by the confluence of enthusiasm generated by a true mathematical understanding of computers in terms of Boolean Algebra and infor-mation theory [130] and a preliminary understanding of the brain that implied that the mathe-matics of biological neural networks may be identical [8]. Cybernetics, as the science of "control and communication in animal and man" was an attempt to integrate control theory and computer science of engineering with theory of the biological brain. While both control system engineer-ing and computer science evolved spectacularly, for some crucial reasons elaborated below, their integration with "mathematical brain theory" was essentially a fail-ure despite some gallant early attempts [11], [171]. However, for about two decades, Cybernetics managed to maintain a rather spectacular momentum, fueled mainly by technology developments in both control system engineering and computer science. The most serious twin reason for not succeeding in inte-grating a true understanding of biological brains with technology of brain-like machines was the relative underde-velopment (in the fifties and sixties) of (experimental) neu-roscience, and the cost-ineffective accessibility of large computer systems for actual implementation of artificial neural systems. The first severe blow to any hope of integration was the slim posthumous book of von Neumann, who as a creator of the conventional computer; the so-called von Neumann machine was an authority to make it rather clear that the fun-damental mathematical principles of the computer and those of the biological brain are di-ametrically oppo-site; the first being a serially organized system, while the latter is a massively parallel or- ganization [133]. The final blow was delivered about a decade later, when some gross inabilities of the principal mathematical paradigm of "neural networks" (the Perceptron) were made clear [172]. Thus, during the sixties Cybernetics di-vorced; one branch adopted a strategy to develop brain-like machines in a totally artificial manner, without reliance on any understanding of the biological brain, merely by develop-ing software for con-ventional computers which would mimic intelligent functions of the brain (AI), while the other branch immersed into neural modeling of actual biological systems, with-out any immediate hope for a more general understanding.

"Artificial Intelligence" developed into a vast field from the initiative by Minsky. According to the Neural Network program administrator at DARPA, Dr. Barbara Yoon, the estimated support of AI from DoD alone, over more than a decade, amounted to half a bil- lion dollars). This investment produced "expert systems", not only useful but essential for controlling modern complex pieces of engineering. On the other hand, neuroscience experienced a massive build-up of knowledge of the bio-logical brain, based on "wet" experimentation at neural system, single brain cell, neuronal membrane, synaptic, and molecular (ion channel) levels. By the eighties, it was fairly clear that Artificial Intelligence did not fully satisfy either users in engineering ("expert systems" for typical DoD applications reach a point of complexity in soft-ware development that seriously affects reliability). In turn, the greatly expanding community of neu-roscientists saw little possibility for integrating knowledge of the biological brain with "dry" approaches taken in Artificial Intelligence. Thus, during the period of Artificial Intelligence, mathematicaltheoretical ap-proaches to CNS function were largely confined to what was called "neural modeling" a service com-ple-menting "wet" experimental neuroscience with occasional "hard" mathematics and computer simulation [14],[173],[16],[17]. By the late seventies, however, neuroscience and mathematical brain theory devel-oped algorithms of massively inter-connected neural nets that laid down the basics for a revival of computational-mathematical brain theory [174],[123],[84],[132],[175],[154]. Also, with the explosive growth of computer industry, hardware means of implementation (microcomputers and microchips) became easily accessible for general use.

Neural Networks initiative was seized in 1982 by a physicist, J. Hopfield of Caltech, in order to trig-ger a new wave of integration of "dry" and "wet" neural network research. His influential paper [176] revitalized interest and has led to a sequence of events generating the new field of "Neural Networks". He presented solid mathematical evidence (based on knowledge available from literature; see a collection [177] that "neural networks" are capable of solving mathemati-cally rather intractable problems without apparent programming, by a simple and im-plementable hardware of massively parallel set of interconnections; c.f. "traveling salesman paradigm". The resulting intellectual fervor led to a set of tumultuous US and international meetings and had created a rapidly increasing fol-lowership of several thousands see di-rections of research in [178]. Rapid develop-ments occurred in the United States basically at two levels. First, in the best tradition of entrepreneurship, dozens of startup companies sprung, a few of which is starting to show a profit by now [179]. Second, at an organizational level, initiative was seized by the de-fense establishment. Organized by Dr. Jasper Lupo at DARPAtTO, a major study was launched in 1987, which resulted in a massive volume of documents on what the defense establishment sees as an "emerging technology" of neural com- puters [180]. A program resulted from this study was presented for congressional ap-proval, with a recommendation of $400 million over the span of 7 years. The much her-alded broad agency announcement by DARPA by March 1st 1989 created high expectations (more than 600 proposals were received). Because of the realities of DARPA research directions and funding interests, it was decided, however, that none of these funds would go supporting neuroscience related neuro-computer research. This decision is presently mute, since at this time (after a delay of one year), a scaled down initial batch of funds for DARPA's neurocomputing initiative ($33 million for 28 months) is still in congressional limbo and no checks will be issued this year

Neurocomputing today is an emerging technology that attracts heated worldwide interest, because of its implications in revolutionizing computer industry (neurochip manufacturing), and its applications in advanced computation in defense as well as in a wide array of civilian projects such as flight control and robotics. Landmarks of the interest and importance are the big share of neurocomputing in the $6,000 mil-lion "Human Frontiers" program by Japan over 20 years, $100 million neurocomputing program of Germany over 10 years, or the DARPA $33 million program for the next 28 months in the US.


WHY IS NEUROCOMPUTING TIMELY AND IMPORTANT?
AN UNFOLDING SCIENTIFICtECHNOLOGICAl REVOLUTION
Neurocomputing is a scientifictechnological revolution. Major breakthroughs in science and technol-ogy have significantly altered the world throughout history. Inventions of gun powder, the steam engine, automobile, airplane, electricity, nuclear energy, rockets, genetic engineering are only a few examples. A relatively recent scientifictechnological breakthrough was the development of computers. Their importance need not be belabored. Suffice to observe that today's world-balance depends not so much on the avail-ability of raw materials or weapons but on the access to superior computer systems to control and automate both production (robotics), and de-fense (complex strategic systems). Therefore, it is of high significance to observe that a scientifictech-nological breakthrough is presently taking place, leading to construction of brain-like computers. Brain-like machines will have profound implications well into the next millennium. It is difficult to overestimate the revolution un-folding today in neurocomputing since it is about to make present-day (von-Neumann type) computers, the traditional ways and means of controlling both production and de-fense, surpassed.

One may wish to point out at the outset that Neurocomputing is only one of several apparent breakthroughs that attract often hyped excitement. Cold Fusion and Superconductivity at Room Temperature are similarly in the forefront of interest. There is a great difference however. It is a scientific truism that in ab-sence of existence proof Cold Fusion or Warm Superconductance may or may not exist. Neurocomputers on the other hand, are ubiquitous as they have been perfected by nature throughout many million years of evolution. Each reader of this report uses an existing portable model as an "existence proof". Compared to "cold" fusion and "warm" superconductivity Neurocomputing is the truly "hot" is-sue, since the question is not whether neurocomputers exist and thus could be mim-icked. The simple question is not whether but why, when, where, and by whom will neurocomputers be developed.


WHEN AND WHERE DOES NEUROCOMPUTING EMERGE?
THE STATE OF AFFAIRS OF NEUROCOMPUTING IN EUROPE, JAPAN AND THE USA

The status-quo of brain theory, leading to neurocomputers, is determined by two fundamental factors. One is a strategic technological need. The other is an impending scientific breakthrough that makes the need fulfillable. As for the need, one might wish to recall that today's computers were developed during World War II in direct response to a demand for controlling fast sin- gle projectiles. It is rapidly be-coming clear that the interdependent, optimized control of a vast array of defense- projectiles alone poses a strategic challenge for developing a new breed of parallelly orga-nized "computers". Arrays of actuators and sen-sors used in robotic production systems also call for means of massively parallel computations. Brain Theory is about to find out how Nature accomplishes par-allel computations, Neurocomputer Technology is well on its way to implement it.

Terms of Neurocomputers, Neurochips, Neurobotics, Neurophilosophy signify that the scientific breakthrough, that appears to make such instrumentation possible, is likely to emerge with the cardinal contribution by Neuroscience. Nature's means of performing computation are the vast interconnected networks of brain cells, the neurons. Theoretical principles of how neuronal net-works operate however have barely be-gun to be revealed in Neuroscience in a clear mathematical manner. A reason for this lag is that Neuroscience has hitherto concentrated on the first step of building up the necessary body of knowledge about the nervous system by experimentally gathering data. To date, only an insignificant fraction of the investments has been aimed at constructing a mathematical understanding of the principles underlying the function of Neuronal Networks. Today, there are few theories available for even the simplest brain function such as sensorimotor operations. Their mathematical understanding is essential for any kind of physical (or computer-software) implementation of brain- like machines. This is crucial for simple robotic motor effectors, let alone robots equipped with vision and other sensation, or intelligence. It may be an ill spent effort to launch full-scale development of Neurocomputers without necessary understanding of the actual computing that neuronal networks perform, or even what neurons and neuronal networks are.

The positive aspects of the dramatic development in neurocomputing are not hailed in this report, as there is more than sufficient coverage of its success in the professional as well as general press. Rather than dwelling on successes of neurocomputing, this report pinpoints some major problems that character-ize the present state of affairs and attempts to formulate recommendations for their resolution. Suffice to mention the existence of three generations of existing neurocomputers; software simulation by traditional von-Neumann computers, microcomputer-hosted parallel boards implementing neurocomputing algo-rithms by a massively parallel processor, and "neurochips" developed by several Japanese and U.S. semiconductor manufacturers. Dozens of startup neurocomputer companies exist, a few of which are starting to make a profit. The US-dominated international Institute of Electrical and Electronics Engineering Society and the American Physics Society are actively sponsoring neurocom-puting. The International Neural Net Society (INNS) and the Japanese Neural Net Society have been set up solely for this purpose. International as well as regional (mostly US and European) meetings more than abound. At least half a dozen new jour-nals ("Neural Networks", "Network", "Neural Computing", "International Journal of Neurocomputing", "IEEE Transactions on Neural Nets", "Biological Cybernetics") and many dozens of books are devoted to this field with many more being in the pipeline.

DIVERGENCES THAT HAMPER CONSOLIDATION Although the interest is heated, and the "Neurocomputing" community grows rapidly, there are three major stumbling blocks on the road to- wards consolidation.

INTERCONTINENTAl TRIANGLE JEOPARDIZES COOPERATION OF THE WESTERN WORLD Since Neurocomputers are important for both the production and defense of the Western World, inter-continental competition and cooperation of the US, Europe and Japan is predictable in the process of getting orga-nized for attaining such strategic goals. An imbalance of assets and somewhat divergent regional goals, however, makes this problem very difficult, if not impossible, to resolve on a short-sighted basis. While the US is extremely strong in experimental Neuroscience and classical (von- Neumann type) Computer Science, it seri-ously neglects Brain Theory.

The international government-funding of neurocomputing was exposed for instance recently at the International Joint Conference for Neural Nets in Washington D.C. [181]. It was pointed out that the sit-uation is different in Europe, which uses to its advantage its well-known favor to-wards lean theoretical approaches. Early advancements in Brain Theory were traditionally accomplished in Europe. Robotics and advanced computer system development lags, however, in Europe. Presently, European EC countries organize neurocomputing in the civilian domain, mostly within academia. The ESPRIT program by the European Community EC allocates between $36 and $48 (M)illion over the next four years on neuro- computing related R&D. This includes $6M over three years for the project ANNIE and $6M over two years for the PYGMALION neurocomputer project. $7.2 M spending is assigned over two years for other ESPRIT basic research. The other European neurocomputer initiative is BRAIN, with current funding of $4.8 M over two years and a plan for total spending of 160M ECU (roughly $200M). For Germany alone, 10.5 M DM (roughly $7M) was allocated for 188- 1990, and approximately $18M is planned for 1991-1993. This program, Information Processing in Neural Architecture, focuses on visual and auditory recognition of patterns, associative memories, control of robots and autonomous vehicles, and neural net hardware. 90% of the funding in Germany goes to academic institutions, including eight en-dowed profes-sorships created for neurocomputing.

Japan, on the other hand, contrasts and complements Europe by thrusting aggressively not only in the field of classical computer- chips and traditional robotics, but also in creative large-scale funding of ad-vanced R&D efforts in the outlined new direction; cf. "Human Frontier Science Program" (HFSP), the 6th Generation Computer Program, and the International Institute for Novel Computing (IINC). Two Japanese governmental agencies oversee the organization, allotment, and distribution of funding in neu- rocomputing the well-known MITI (Ministry for International Trade and Industry) and the STA (Science and Technology Agency). HFSP is seeking to further comprehension of human biological systems, in- cluding efforts to simu-late the human brain. It is operating with $18M for 1990 and $24M for 1991. STA is also funding the Riken Frontier Science Program ($45M for 15 years), the Brain Functions program ($1.5M for three years) and the Biodevice Program ($25M for ten years). The 6th Generation program, replacing the previous one that spent $350M over a decade will focus on the following key areas of com-puting: learn-ing, parallel distributed processing, robotics, and pattern recognition. It is noteworthy that new signs of a "reverse brain drain" can be observed in the international landscape. International Prizes (E.g. Humboldt-Prize, Kyoto-Prize) and International Research Awards (such as Human Frontier Science Program) are exposing and luring abroad some of the best ideas, concepts and plans developed in US.

INTERINSTITUTIONAl TRIANGLE DELAYS ORGANIZATION OF INFRASTRUCTURE The well-known cultural and structural divergences of Academic Industrial and Defense institutions represent another major difficulty of consolidation. Academic organizations such as University Departments could best play the role of becoming the neutral and accessible fo-rum that can consolidate the various research ef-forts. Actual construction of "Neurocomputers" is a much more suitable Research and De- velopment issue for in-dustrial institutions. In turn, a traditionally leading major consumer of new types of computers is likely to be the defense establishment. Thus, the field of neurocomputing is pulled apart by three drastically different organizational and cultural environments.

INTERDISCIPLINAl TRIANGLE MAKES A PROPER UNIVERSITY- EMBEDDING DIFFICULTthe new field of neu-rocomput-ing is expanding on a territory that presently is a "no man's land" or "everybody's turf" among the three major circles of Experimental Neuroscience and Medicine, Mathematical Computer Science and Artificial Intelligence, Basic Physical Sciences, and Robotics Engineering. Since these fields fall into different Schools; Medical, Arts and Sciences and Engineering, the triangle handicaps those Universities that do not have all these faculties. Missing one or two of the three precludes pursu-ing the above goals in a fully integrated manner. The triangle hampers even full Universities, since such emerging interdisciplinary activities are parcelled and fragmented in the three different Schools. Having to break new ground between established territories is sometimes seen not only as a problem but also as a chal-lenge that can open up new possibilities. Neurocomputing Centers are presently formed either as divi-sions of exist-ing Departments of Medical School; (e.g. of Neuroscience, Biophysics, Rehabilitation Medicine, Physiology and Biophysics Depts.) or as part of Basic Science School Departments (e.g. of Physics, Computer Science, even Mathematics Depts.) or as part of Engineering School Departments (e.g. of Robotics, Biomedical Engineering Depts.). While Neurocomputing could be best established as a sepa-rate Department,with full and optimal growing potential, because of the above problems no University in the U.S.is known to the author to have established such a Department as yet. This situation is to be com-pared to the emergence of computer science departments. In the forties and fifties few universities had them, while today they are a must. Situation was much simpler though, as computer science departments traditionally emerged from departments of electrical engineering.

---

REFERENCES


[1] Barchi, R. L. (1990) Director's Column. Forum, Mahoney Institute of Neurological Sciences 1(2): 2.

[2] Pellionisz, A., J. Soechting, C. Gielen, J. Simpson, B. Peterson and A. Georgopoulos (1986) Workshop: Multidimensional analyses of sensorimotor systems. Soc. Neurosci. Absts. 12: 1.

[3] Cole, K. (1972) Membranes, ions and impulses. Univ. Calif. Press,

[4] Jack, J. J. B., Noble, D. and Tsien, R.W. (1975) . Clarendon Press, Oxford (1975) Electric Current Flow in Excitable Cells. :

[5] McGregor, R. J. (1988) Neural and Brain Modeling. Academic Press,

[6] Koch, C. and I. Segev (1989) Methods in neural modeling. MIT Press, Cambridge, MA,

[7] Pellionisz, A. (1979) Modeling of Neurons and Neuronal Networks. The Neurosciences: IVth Study Program, F. Schmitt and F. Worden, eds., MIT Press, Boston, 525-546.

[8] McCulloch, W. S. and W. Pitts (1943) A logical calculus of the ideas immanent in nervous activity. Bull. Math. Biophys. : 5, 115- 133.

[9] Caianiello, E. R. (1968) Proceedings of the school on neuronal networks. Neuronal Networks, E. R. Caianiello, eds., Springer Verlag, New York,

[10] Hebb, D. O. (1949) The Organization of Behaviour. John Wiley, New York,

[11] Steinbuch, K. (1961) Die Lernmatrix. Kybernetik : 1, 36-45.

[12] Marr, D. (1969) A theory of cerebellar cortex. J. Physiol. 202: 437-470.

[13] Hodgkin, A. and A. Huxley (1952) A quantitative description of membrane current and its application to conduction and excitation in nerve. J. Physiol. (London) : 1176:500-544.

[14] Rall, W. (1964) Theoretical significance of dendritic trees for neuronal input output relations. Neural Theory and Modeling, R. F. Reiss, eds., Stanford Univ. Press, Stanford, CA, 7397.

[15] Cooley, J. W. and F. A. Dodge (1966) Digital computer solutions for excitation and propagation of the nerve impulse. Biophys. J. 6: 583-599.

[16] Pellionisz, A. and R. Llinás (1977) Computer Model of Cerebellar Purkinje Cells. Neuroscience 2: 37-48.

[17] Traub, R. D. and R. Llinás (1977) The spatial distribution of ionic conductances in normal and axotomized motor neurons.ÿ. Neuroscience 2(6): 829-850.

[18] Traub, R. (1982) Simulation of intrinsic bursting in CA3 hippocampal neurons. Neurosci. 7: 1233-1242.

[19] Koch, C., T. Poggio and V. Torre (1982) Retinal ganglion cells: a functional interpretation of dendritic morphology. Phil. Trans. R. Soc. Concl. B. 298: 227-264.

[20] Shelton, D. P. (1985) Membrane resistivity estimated for the Purkinje neuron by means of a passive Comp. model. Neurosci. 14: 111-131.

[21] Borg-Graham, L. (1988) Modeling the electrical behavior of cortical neurons simulation of hippocampal pyramidal cells. Computer simulation in brain science, R. M. J. Cotterill, eds., Cambridge Univ. Press, Cambridge, UK, 384-404.

[22] Segev, I. and W. Rall (1988) Computational study of an excitable dendritic spine. J. Neurophysiol. 60: 499-523.

[23] Sheperd, G., R. Brayton, J. Miller, I. Segev, J. Rinzel and W. Rall (1985) Signal enhancement in distal cortical dendrites by means of interactions between active dendr. spines. Proc. Natl. Acad. Sci. USA 82: 2192-2195.

[24] Koch, D. and T. Poggio (1983) A theoretical analysis of electrical properties of spines. Proc. R. Soc. London. Series B 218: 455-477.

[25] Koch, C. and T. Poggio (1983) Electrical properties of dendritic spines. Trends Neurosci. 6: 80-83.

[26] Crill, W. and P. Schwindt (1983) Active currents in mammalian central neurons. Trends Neurosci. 6: 236-240.

[27] Descartes, R. (1972) Treatise of Man (Original in 1629 in Latin and French). Harvard Univ. Press, Cambridge, MA,

[28] Towe, A. and E. Luschei (1981) Handbook of behavioral neurobiology. Motor coordination, eds., Plenum, New York,

[29] Lorente de No, R. (1933) Vestibulo-ocular reflex arc. Arch. Neurol. Psychiatry 30(2): 245-291.

[30] Szentágothai, J. (1950) The elementary vestibulo-ocular reflex arc. J. Neurophysiol. 13: 395-407.

[31] Peterson, B. W. and F. J. R. Richmond (1988) Control of head movements. Oxford University Press,

[32] Robinson, D. (1968) The oculomotor control system: A review. Proc. IEEE 56: 1032-1049.

[33] Stark, L. (1968) Neurological Control Systems. Plenum Press, New York,

[34] Houk, J. (1980) Principles of system theory as applied to physiology: Systems and models. Medical Physiology, V. Mountcastle, eds., Mosby CV and Co, St Louis, Chapter 7.

[35] Hollerbach, J. and T. Flash (1982) Biol Cybernetics, 44:67-77 :

[36] Pellionisz, A. (1983) Brain Theory: Connecting Neurobiology to Robotics. Tensor Analysis: Utilizing Intrinsic Coordinates to Describe, Understand and Engineer Functional Geometries to Intelligent Organisms. J. Theoret. Neurobiol. : 185-211.

[37] Hogan, N. (1984) Mechanical impedance control in assistive devices and manipulators. Robot control, J. Brady, J. Hollerbach, T. Johnson, T. Lozano-Perez and M. T. Mason, eds., MIT Press, Cambridge, MA, 361-370.

[38] Gurfinkel, V. S. (1987) Robotics and biological motor control. Proc. of II. IBRO World Cong. Neurosci, Suppl., eds., S381.

[39] Pellionisz, A. (1988) Tensor geometry: A language of brains and neurocomputers. Generalized coordinates in neuroscience and robotics. Neural Computers, DŸsseldorf, R. Eckmiller and C. v. Malsburg, eds., Springer Verlag, Berlin, 381-391.

[40] Sherrington, C. (1906) The Integrative Action of the Nervous System. Scribner, New York,

[41] Robinson, D. (1976) Adaptive gain control of vestibulo-ocular reflex by the cerebellum. J. Neurophysiol. 39: 954-969.

[42] Robinson, D. A. (1982) The use of matrices in analyzing the three-dimensional behavior of the vestibulo-ocular reflex. Biol. Cybern. 46: 53-66.

[43] Robinson, D. (1985) The coordinates of neurons in the vestibulo-ocular reflex. Adaptive Mechanisms in Gaze Control Facts and Theories, A. Berthoz and G. Melvill-Jones, eds., Elsevier, Amsterdam, 297-311.

[44] Bernstein, N. (1935) The problem of the interrelation of coordination and localization. Arch. Biol. Sci. 38: 15-59.

[45] Klein, F. (1939) Elementary Mathematics form an Advanced Viewpoint. Dover, New York,

[46] Soechting, J. and B. Ross (1984) Psychophysical determination of coordinate representation of human arm orientation. Neuroscience 13(2): 595-604.

[47] Brooks, V. B. (1985) How are "move" and "hold" programs matched. Cerebellar Functions, J. R. Bloedel, J. Dichgans and W. Precht, eds., Springer, Berlin, 1-23.

[48] Bizzi, E. (1981) Visuomotor control as a computational problem. Theoretical Approaches in Neurobiology, W. E. a. P. Reichardt T. eds, eds., MIT Press, Cambridge, MA, 177-184.

[49] Bizzi, E., N. Accornero, W. Chapple and N. Hogan (1984) J Neurosci 4: 2738-2744.

[50] Massone, L. and E. Bizzi (1989) Generation of limb trajectories with a sequential network. Proc. of Intl. Joint Conf. on Neural Nets, Washington, D.C., eds., 345-349.

[51] Mussa Ivaldi, F., J. McIntyre and E. Bizzi (1988) Theoretical and Experimental Perspectives on Arm Trajectory Formation: A distributed Model for Motor Redundancy. Biological and Artificial Intelligence Systems, E. Clementi Chin, S., eds., Escom, 563-577.

[52] Georgopoulos, A., A. Schwartz and R. Kettner (1986) Neuronal population coding of movement direction. Science 233: 1416- 1419.

[53] Bizzi, E., F. Mussa-Ivaldi and N. Hogan (1987) Neurosci. Suppl. 22: 1145S.

[54] Henn, W. and B. Cohen (1988) Proc. of the Symp: Representation of 3-Dimensional Space in the Vestibular, Oculomotor and Visual Systems. Org.by the Bárány Society, Bologna, Italy, eds.,

[55] Deno, D. C. and E. Keller (1989) Visual pursuit errors: Implications for network models. International Joint Conf. on Neural Nets, Washington, D.C., eds., I(717-722).

[56] Anastasio, T. J. and D. A. Robinson (1989) Distributed parallel processing in the vestibulo-oculomotor system. Neural Comp. 1(2): 230-241.

[57] Pellionisz, A. (1988) Vistas from Tensor Network Theory: A horizon from reductionalist neurophilosophy to the geometry of multi-unit recordings. Computer Simulation in Brain Science, R. Cotterill, eds., Cambridge Univ. Press, Cambridge, 44-73.

[58] Freeman, W. (1975) Mass action in the nervous system. Academic Press, New York,

[59] Kruger, J. (1982) A 12-fold microelectrode for recording from vertically aligned cortical neurons. J. Neurosci. Methods 6: 347- 350.

[60] Gernstein, G. L., M. J. Bloom, I. E. Espinosa, S. Evanczuk and M. R. Turner (1983) Design of a laboratory for multineuron studies. IEEE Systems, Man and Cybernetics 13: 668-676.

[61] Reitbšck, H. J. P. (1983) A 19-channel matrix drive with individually controllable fiber microelectrodes for neurophysiological applications. IEEE Trans. on System, Man and Cybernetics, SMC-13 5: 676-682.

[62] Bower, J. and R. Llinás (1983) Simultaneous sampling of the responses of multiple, closely adjacent, Purkinje cells responding to climbing fiber activation. Soc. Neurosci. Abst. 9: 607.

[63] Carman, G. J., B. Rasnow and J. M. Bower (1986) Analysis of the dynamics of activity in ensembles of neurons recorded simultaneously in cerebellar cortex. Society for Neuroscience Abstracts 12: 1417.

[64] Fukuda, M., T. Yamamoto and R. Llinás (1987) Simultaneous recordings from Purkinje cells of different folia in the rat cerebellum and their relation to movement. Abst Soc Neurosci 13: 603.

[65] Nelson, M., B. Rasnow, J. Banik and J. Bower (1988) Spatial & temporal patterns of Purkinje and granule cell activity recorded simultaneously in the rat cerebellum using silicon microelectrodes. Soc. Neurosci Absts. : 759.

[66] Lassen, N. A., D. H. Ingvar and E. Skinhoj (1978) Brain function and blood flow. Sci. Amer. 239: 62-71.

[67] Sokoloff, L. (1981) The relationship between funciton and energy metabolism: Its use in the localization of functional activity in the nervous system. Neurosci. Res. Progr. Bull. 19: 159-210.

[68] Phelps, M. E. and J. C. Mazziotta (1985) Positron emmision tomography: Human brain function and biochemistry. Science 228: 799-809.

[69] Edelman, R. (1984) The clinician's guide to the theory and practice of NMR scanning. Discussions in Neurosciences 1: 9-48.

[70] Hillman, D. E., G. Mahoney, R. Llinás and M. Canady (1989) Basic concepts and methodologies of image acquisition, frame analysis, and image data presentation. Three-Dimensional Neuroimaging, A. Toga, eds., Raven Press, New York,

[71] Ross, M. D. (1988) Morphological evidence for parallel processing of information in rat macula. Acta Otolaryngol. 106: 213- 218.

[72] Helmholtz, H. (1896) Handbuch der Physiologischen Optik. Zweite Auflage. Voss, Leipzig,

[73] Curthoys, I. S., R. H. I. Blanks and C. H. Markham (1977) Semicircular canal functional anatomy in cat, guinea pig and man. Acta Otolaryngol. 83: 258-265.

[74] Blanks, R., I. Curthoys and C. Markham (1975) Planar relationships of the semicircular canals in man. Acta Otolaryngol. 80: 185-196.

[75] Simpson, J., D. Rudinger, H. Reisine and V. Henn (1986) Geometry of extraocular muscles of the rhesus monkey. Soc. Neurosci. Absts. 12: 1186.

[76] Ezure, K. and W. Graf (1984) II. A quantitative analysis of the spatial organization of the vestibulo-ocular reflexes in lateral and frontaleyed animals. II. Neuronal networks underlying vestibulo- oculomotor coordination. Neuroscience 12: 95-109.

[77] Graf, W. and A. Pellionisz (1985) Morphology and a tensor network model of the three-neuron vestibulo-ocular reflex arc in cat. Soc. Neurosci. Abst 11: 1041.

[78] Miller, J. (1990) Functional Anatomy of Normal Human Rectus Muscles. Vision Research (in Press):

[79] Pellionisz, A. (1987) Sensorimotor operations: a ground for the coevolution of brain theory with neurobotics and neurocomputers. IEEE First Annual International Conference on Neural Networks, San Diego, eds., 593-600.

[80] Mauritz, K. H. (1986) Restoration of posture and gait by functional neuromuscular stimulation (FNS). Disorders of posture and gait, W. Bles and T. Brandt, eds., Elsevier, Amsterdam, 367-385.

[81] Kralj, A. and S. Grobelnik (1973) Functional electrical stimulation A new hope for paraplegic patients? Bulletin of Prosthesis Research 10-20: 75-102.

[82] Gruner, J. A. (1986) Considerations in designing acceptable neuromuscular stimulation systems for restoring function in paralyzed limbs. Central Nervous System Trauma 3(1): 37-47.

[83] Grossberg, S. (1982) Studies of mind and brain. Dordrecht,

[84] Anderson, J., J. Silverstein, S. Ritz and J. Randall (1977) Distinctive features, categorical perception and probability learning: Some applications of a neural model. Psychol. Rev. 84: 413-451.

[85] Hinton, G. E. and J. A. Anderson (1981) Parallel models of associative memory. Lawrence Erlbaum Assoc., Hillsdale, New Jersey,

[86] Pellionisz, A. (1987) Tensor network theory of the central nervous system. Encyclopaedia of Neuroscience, G. Adelman, eds., BirkhŠuser, Boston, 1196-1198.

[87] Llinás, R. R. (1988) The intrinsic electrophysiological properties of mammalian neurons: Insights into central nervous system function. Science 242: 1654-1664.

[88] Gielen, C. C. A. M. and E. J. van Zuylen (1986) Coordination of arm muscles during flexion and supination: Application of the tensor analysis approach. Neuroscience 17: 527-539.

[89] Gaál, G. (1987) Non-orthogonal hormonal coordinate axes calculated from correlation matrices according to tensor network theory. Neurosci. Suppl. 22 S146:

[90] Harth, E. (1983) Order and chaos in neural systems: An approach to the dynamics of higher brain functions. IEEE SMC-13 13(5): 782-789.

[91] Skarda, C. and W. Freeman (1987) How brains make chaos to make sense of the world. Behavioral and Brain Sciences (June):

[92] Guevara, M., L. Glass, M. Mackey and A. Shrier (1983) Chaos in neurobiology. IEEE Trans. Systm.Man.Cybernet. 13: 790-798.

[93] Goldberger, A. L., D. R. Rigney and B. J. West (1990) Chaos and fractals in human physiology. Sci.Amer. 1990(2): 42-49.

[94] Chay, T. (1984) Abnormal discharges and chaos in a neural model system. Biol. Cybern. 50: 301-311.

[95] Freeman, W. (1987) Simulation of chaotic EEG patterns with a dynamic model of the olfactory system. Biol. Cybern. 55:

[96] Mandelbrot, B. (1977) The fractal geometry of nature. W.H. Freeman, New York,

[97] Pellionisz, A. (1989) Fractal geometry of Purkinje neurons: Relationships among Metrical and non-metrical neural geometries. Soc. Neurosci. Absts. 15(I.): 180.

[98] Pellionisz, A. (1990) Neural geometry: Towards a fractal model of neurons. Models in Brain Function, R. M. Cotterill, eds., Cambridge Univ. Press, Cambridge,

[99] Temple, G. (1960) Cartesian Tensors. John Wiley, New York,

[100] Koenderink, J. J. (1984) Geometrical structures determined by the functional order in nervous nets. Biol. Cybern 50: 43-50.

[101] Pellionisz, A. J. (1990) About the geometry intrinsic to neural nets. International Joint Conference on Neural Nets, Washington D.C., eds., Lawrence Erlbaum, (in press).

[102] Minkowski, H. and H. Weyl (1908) Space and time. The Principle of Relativity, E. Lorentz H.A. A, eds., Dover, Toronto, 75- 91.

[103] Schršdinger, E. (1950) Spacetime structure. Cambridge Univ. Press,

[104] Weyl, R. (1952) Spacetime-matter. Dover, New York,

[105] Pellionisz, A. and R. Llinás (1982) Spacetime representation in the brain. The cerebellum as a predictive spacetime metric tensor. Neurosci. : 2949-2970.

[106] Baker, J., J. Goldberg, G. Hermann and B. W. Peterson (1984) Spatial and temporal response properties of secondary neurons that receive convergent input in vestibular nuclei of alert cats. Brain Res. 294: 138-143.

[107] Graf, W. and K. Ezure (1986) Morphology of vertical canal related second order vestibular neurons in the cat. Exp. Brain Res. 63: 35-48.

[108] Markham, C. H. and I. S. Curthoys (1972) Convergence of labyrinthine influences on units in the vestibular nuclei of the cat. II. Electrical stimulation. Brain Res. 43: 383-396.

[109] Churchland, P. (1985) Neurophilosophy: Towards a unified understanding of the mind-brain. Bradford Books/MIT Press, Boston,

[110] Albus, J. S. (1971) A theory of cerebellar function. Math. Biosci. 10: 25-61.

[111] Pellionisz, A. (1985) Tensorial brain theory in cerebellar modeling. Cerebellar Functions, J. Bloedel, J. Dichgans and W. Precht, eds., Springer, Heidelberg, 201-229.

[112] Pellionisz, A. and R. Llinás (1980) Tensorial approach to the geometry of brain function: Cerebellar coordination via metric tensor. Neurosci. 5: 1125-1136.

[113] Peterson, B. W. (1990) Modeling of Spatial Transformations in Vestibular Reflex Systems. International Joint Conference for Neural Networks, Washington, D.C., M. Caudill, eds., Lawrence Erlbaum, 152- 156.

[114] McCulloch, W. S. (1965) Embodimentf of mind. MIT Press, Cambridge, MA,

[115] Sperry, R. W. (1980) Mind-brain interaction: Mentalism, yes; dualism, no. Neuroscience 5(195-206):

[116] Posner, E. (1987) Workshop on Neural Network Devices and Applications. Sponsored by SDIO, NASA, DARPA, Pasadena, eds., Jet Propulsion Laboratory, 1-422.

[117] Kanerva, P. (1988) Sparse distributed memory. MIT Press, Boston, MA,

[118] Tomko, D. L. and G. D. Paige (1988) Linear vestibular-ocular reflex (LVOR) in squirrel monkey: I. Basic characteristics. Soc. Neurosci. Abstr. 14(1): 332.

[119] Paige, G. D. and D. L. Tomko (1989) Eye movements and visual- vestibular interactions during linear head motion. Symposium on head movement control, Paris, A. Berthoz, eds., Centre National de le Recherche Scientific, (in press).

[120] Ross, M. D. (1987) A Center for Biocomputation Program Design and Initiation Proposal. NASA-Ames Res. Ctr. :

[121] Robbins, F. C. (1988) Exploring the living universe: A strategy for space life sciences.

[122] Sivilotti, M. A., M. A. Mahowald and C. A. Mead (1987) Real- time visual computations using analog CMOS processing arrays. Advanced Research in VLSI, Stanford, CA, P. Losleben, eds., MIT Press, 295-312.

[123] Cooper, L. (1974) A possible organization of animal memory and learning. In: Proceedings of Nobel Symposium on collective properties of physical systems. Lundquist B, Lundquist S (eds), Academic Press, New York pp 252-264 :

[124] Crick, F. (1979) Thinking about the brain. Sci. Amer. 241: 219-232.

[125] Edelman, G. (1979) Group selection and phasic reentrant signaling: A theory of higher brain function. The neurosciences, IV study program, F. Schmitt and F. Worden, eds., Cambridge, MA, 1115- 1139.

[126] Lewin, R. (1982) Neuroscientists look for theories. Science 216: 507.

[127] Arduini, A. (1987) Principles of Neurophysiology. Springer Verlag,

[128] Pellionisz, A. and R. Llinás (1978) A Formal Theory for Cerebellar Function: The Predictive Distributed Property of the Cortico-Nuclear Cerebellar System as Described by Tensor Network Theory and Computer Simulation. Soc. Neurosci. Absts. 4: 68.

[129] Pellionisz, A. and R. Llinás (1985) Tensor Network Theory of the metaorganization of functional geometries in the CNS. Neurosci. 16: 245-274.

[130] Shannon, C. (1948) A mathematical theory of communication. Bell System Technology J : 27, 3-4.

[131] Turing, A. (1948) Intelligent machinery. Rect. Natl. Phys. Lab. Teddington :

[132] Kohonen, T. (1977) Associative memory. A systemtheoretical approach. Springer, Berlin, Heidelberg, New York,

[133] Neumann, J., von (1957) The computer and the brain. Yale University Pres, New Haven,

[134] Farley, B. G. and W. A. Clark (1961) Activity in networks in neuronlike elements. Information Theory (Fourth London Symposium), C. e. Cherry, eds., Butterworth Scientific Publications, London,

[135] Beurle, R. (1962) Functional organization in random networks. Principles of self-organization, H. Foerster and G. Zopf, eds., Pergamon, New York, 291.

[136] Eccles, J. C. (1957) The Physiology of Nerve Cells. Johns Hopkins Press, Baltimore,

[137] Adrian, E., F. Bremer and H. Jasper (1954) Brain mechanisms and consciousness. Blackwell, Oxford,

[138] Young, L. R. (1969) Biocybernetics of the vestibular system. L. D. Proctor, eds., Little, Brown & Comp. Boston, MA, Washington, D.C., 79-117.

[139] Ito, M. (1970) Neurophysiological aspects of the cerebellar motor control system. Int. J. Neurol. 7: 162-176.

[140] Pellionisz, A. (1987) Tensor geometry as the mathematical language of neuronal networks. Brain theory Ñ Foundation for neurobotics and neurocomputers. Neurosci. Suppl. to Vol.22(Proc. of IBRO II.World Congress): S101.

[141] Pellionisz, A. (1987) Tensor geometry: Mathematical brain theory for neurocomputers and neurobots. Proc. of the Ninth Annual Conference of IEEE Engineering in Medicine and Biology Society, Boston, eds., 1368-1369.

[142] Pellionisz, A. (1988) Intelligent decisions and dynamic coordination: Properties of geometrical representation by generalized frames intrinsic to neural and robotic systems. 2nd IEEE Internatl. Conf. on Neuronal Networks, San Diego, eds., 603-610.

[143] Pellionisz, A. (1989) Brain geometry: The need of researching association of covariant and contravariant coordinates that organizes a cognitive space by relating multisensory-multimotor representations. 3rd IEEE Internatl. Conf. on Neuronal Networks, Washington, D.C., eds., 711-715.

[144] Eckmiller, R. and C. Malsburg von (1988) Neural Computers. Proc of the NATO Adv. Research Workshop, Springer, eds., Springer, DŸsseldorf,

[145] Massone, L. and E. Bizzi (1990) On the Role of Input Representations in Sensorimotor Mapping. International Joint Conference for Neural Networks, Washington, D.C., M. Caudill, eds., Lawrence Erlbaum, 173-176.

[146] Eckmiller, R. (1990) Internal Representation of Space in Neural Networks of Primates and Other Sensorimotor Machines. International Joint Conference for Neural Networks, Washington, D.C., M. Caudill, eds., Lawrence Erlbaum, 165-168.

[147] Fukushima, K. and T. Ito (1970) A Neural Network Model Extracting Features from Speech Signals. The Transactions of the !nstitute of Electronics, Information and Communication Engineers, Japan, Vol. J70.D, No. 2 : 451-462.

[148] Albus, J. R. (1981) Brains, behavior and robotics. McGraw-Hill, New York,

[149] Eccles, J. C., M. Ito and J. Szentágothai (1967) The cerebellum as a neuronal machine. Springer-Verlag, Berlin, New York, Heidelberg,

[150] Llinás, R. (. ). (1969) Neurobiology of cerebellar evolution and development. Am. Med. Assoc., Chicago,

[151] Berthoz, A. and G. Melvill-Jones (1985) Adaptive Mechanisms in Gaze Control. Reviews in Oculomotor Research. Elsevier, Amsterdam,

[152] Cohen, B. (1981) Vestibular and oculomotor physiology. Int. Mtg. Barany Soc. Ann. N.Y.Acad. Sci., eds., 892.

[153] Robinson, D. (1981) The use of control systems analysis in the neurophysiology of eye movements. Ann Rev Neurosci. 4: 463- 503.

[154] Pellionisz, A. and R. Llinás (1979) Brain modeling by tensor network theory and computer simulation. The cerebellum: Distributed processor for predictive coordination. Neurosci 4: 323- 348.

[155] Pellionisz, A. (1984) Coordination: A vector-matrix description of transformations of overcomplete CNS coordinates and a tensorial solution using the Moore-Penrose generalized inverse. J. Theoret. Biol 110: 353-375.

[156] Simpson, J. and A. Pellionisz (1984) The vestibulo-ocular reflex in rabbit as interpreted using the Moore-Penrose generalized inverse transformation of intrinsic coordinates. Soc. Neurosci. Absts. 10: 909.

[157] Pellionisz, A. (1985) Tensorial aspects of the multidimensional approach to the vestibulo-oculomotor reflex and gaze. Adaptive Mechanisms in Gaze Control, A. Berthoz and G. Melvill-Jones, eds., Elsevier, Amsterdam, 281-296.

[158] Daunicht, W. and A. Pellionisz (1987) Spatial arrangement of the vestibular and the oculomotor system in the rat. Brain Res. 435: 48-56.

[159] Liverneaux, P., A. J. Pellionisz and F. G. Lestienne (1987) Morpho-anatomy and muscular synergy of sub-occipital muscles in Macaca Mulatta: Study of headtrunk coordination. Neurosci. Suppl. to Vol. 22: S847.

[160] Bloedel, J. R., S. Tillery and A. J. Pellionisz (1988) Experimentaltheoretical analysis of the intrinsic geometry of limb movements. Soc. Neurosci. Absts 14/2: 953.

[161] Berthoz, A., M. Benhamou and A. J. Pellionisz (1988) Postural positions yielding aligned Eigenvectors of coordinate frames intrinsic to vestibular and head-neck muscle systems in human. Soc. Neurosci. Absts 14/2: 1235.

[162] Laczkó, J., A. Pellionisz, H. Jongen and C. C. A. M. Gielen (1988) Computer modeling of human forelimb muscle activation in multidimensional intrinsic coordinate frames. Soc. Neurosci. Absts. 14/2: 955.

[163] Laczkó, J., A. J. Pellionisz, B. W. Peterson and T. S. Buchanan (1987) Multidimensional sensorimotor ÒpatternsÓ arising from a graphics-based tensorial model of the neck-motor system. Soc. Neurosci. Absts 13: 372.

[164] Georgopoulos, A. (1990) Coding of the Direction of Reaching by Neural Populations. International Joint Conference for Neural Networks, Washington, D.C., M. Caudill, eds., Lawrence Erlbaum, 169.

[165] Collins, T. J. (1988) Cerebral cortical parallel processing using a metric tensor. Proc. Internatl. Neur. Network Society 2nd Annual Meeting, Boston, eds., 85.

[166] Peterson, B. W., A. J. Pellionisz, J. A. Baker and E. A. Keshner (1989) Functional morphology and neural control of neck muscles in mammals. Am. Zoology 29: 139-149.

[167] Pellionisz, A. and B. W. Peterson (1988) A tensorial model of neck motor activation. Control of Head Movement, B. Peterson and F. Richmond, eds., Oxford University Press, Oxford, 178-186.

[168] Pellionisz, A. (1988) Tensorial aspects of the multidimensional massively parallel sensorimotor function of neuronal networks. Progress in Brain Research, O. Pompeiano and J. H. J. Allum, eds., Elsevier, Amsterdam, 341-354.

[169] Lestienne, F., P. Liverneaux and A. Pellionisz (1988) Morpho- anatomy of sub-occipital muscles in monkey: a tensor model of the musculo-skeletal head-neck system. Reunion Commune de la Physiological Society et de lÕAssociation des Physiologistes. J. Physiol. (Lond) : 2P.

[170] Wiener, N. (1948) Cybernetics, or control and communication in the animal and the machine. MIT Press, Cambridge, MA,

[171] Caianiello, E. R. (1961) Outline of a theory of thought- processes and thinking machines. J. Theor. Biol. : 204-235.

[172] Minsky, M. and S. Papert (1969) Perceptrons. MIT Press, Cambridge, MA,

[173] Adelman, W. (1971) Biophysics and physiology of excitable membranes. Van Nostrand Reinhold, NY,

[174] Werbos (1974)

[175] Edelman, G. M. and V. B. Mountcastle (1978) The Mindful Brain: Control Organization and the Group-Selective Theory of Higher Brain Function. MIT Press, Cambridge, MA,

[176] Hopfield, J. (1982) Neuronal networks and physical systems with emergent collective computational abilities. PNAS 79: 2554- 2558.

[177] Anderson, J. and E. Rosenfeld (1987) Neurocomputing: I. Foundations of Research. MIT Press, Cambridge, MA,

[178] Anderson, J., A. Pellionisz and E. Rosenfeld (1990) Neurocomputing. Vol.II. Directions of Research. MIT Press (in Press), Cambridge, MA,

[179] Rosenfeld, E. (1990) Neuralware is profitable. Intelligence 6(9): 1-8.

[180] Lupo, J. (1988) DARPA Neural Network Study. Armed Forces Communications & Electronics Association International, Fairfax, VA, 1-636.

[181] Rosenfeld, E. (1990) Government Funding Panel. Intelligence 6(10): 2-3.

---

EXECUTIVE SUMMARY

1.0 Introduction

On July 25, 1989, the President of the United States signed into law House Joint Resolution #174 declaring the 1990s the "Decade of the Brain". The joint resolution signals a new federal government commitment to neuroscience research. In the spirit of this law, two federal Agencies (NIH-NIMH and NASA) jointly sponsored the present report of whether and how NASA and the NIH might benefit from a cooperative program related to both brain sciences and technological research and development.

It is concluded and evidence presented in this report to document that neurocomputing is an area in which interagency cooperation is both most promising and needed. Developing brain-like computers is a new opportunity to basic science as well as to technological research and development. However, it is also a major challenge to scientific management, as such technology cannot be safely and economically developed without understanding of brain function, yet programmatic mechanisms to integrate neural science and neurocomputer technology are yet to be created. Thus, initiatives for research on neurocomputing often fall either entirely into the realm of technology or solely of neuroscience leaving a potentially fatal gap unless strategies and measures are implemented for systemic integration. This report gives an analysis of the problem and provides recommendations for its solution.

1.1 Goals and Objectives of this Report

Describe existing program elements within NASA-ARC and the NIH-NIMH which are relevant to a potential Neurocomputing Program co-sponsored by these organizations (Sect.2.)

# Discuss potential advantages and disadvantages of a joint venture (Sect. 3)

# Delineate major issues to be resolved in planning a neurocomputing program (Sect. 4)

# Provide conclusions and recommendations to the two organizations on how to create a civilian neurocomputing initiative by the government (Sect. 5).

1.2 The "Decade of the Brain": Neurocomputing is a Pivotal Challenge for

Neuroscience and Life Science Organization

Neurocomputing may be profitably used as a central theme for Neuroscience within Life Science programs. It is suggested, therefore, that:

* NASA and NIH, together with NSF, each with specific areas of interest and expertise, coordinate their activities in the field of neurocomputing, and cooperate to launch a future initiative. Joint action would yield a long term advantage for these agencies, as neurocomputing is a linchpin between research of the "wet brain" and corresponding electronic implementations.

* The cooperation should create a platform for a civilian neurocomputing initiative. Such a program is necessary to integrate neuroscience into neurocomputing technology and to create an organizational structure in the US that is most competitive with neurocomputer research and development worldwide.
2.0Existing Components of a Neurocomputing Program within NIH- NIHM

and NASA-ARC

Investment in experimental neuroscience, predominantly by NIH, has lead to development of an immense body of descriptive data on the brain. Neuroscience must now turn towards integration of these data. Theory should provide both testable hypotheses as well as mathematical concepts and formalisms for scientists and engineers who make artificial electronic neurocomputers guided by principles of brain organization. Neither pure experimentation nor pure theory is alone sufficient to accomplish this complex task. In the neurosciences, mathematical/computational/theoretical work is increasingly a partner with experimentation. This assures a more rounded and distinguished research environment which is consistent with the funding goals of more than one agency or program. In turn, implementation of "brain-like machines" reached the point at NASA when integration is much desired, for example to transform knowledge of biological coordination into its mathematical understanding and then utilization for flight control by neural nets. This trend towards mathematical theory is illustrated by the fact that:

# at an institutional level, at NIH-NIMH, a special program was created for this purpose. This program needs substantial further improvement as experimental neuroscience progresses towards theory

# at NASA, both life science and technology based neurocomputing projects exist or are planned which need to be integrated within a NASA research center, among all NASA centers, and with other research & development agencies

2.1 Mathematical/Theoretical/Computational (MTC) Neuroscience Program of NIH-NIMH

In response to the trend towards theory in neuroscience, a new program "Theoretical/ Mathematical/ Computational Neuroscience" opened for applications by February 1st 1989, jointly administered by NIH-NIMH and NINDS. The scientific program:

invites applications for studies using mathematical, computational, or theoretical approaches to understanding the fundamental mechanisms underlying behavior

declares the purpose as placing additional emphasis on the use of quantitative tools in solving basic problems in the neurosciences

is headed by Dr. Richard Nakamura at NIHM and Dr. Herbert Lansdell at NINDS

uses an ad hoc Study Section to evaluate proposals

operates without a set budget; in theory having an open access on a competitive basis with all other grants to all programs by NIH. To date, 6 proposals were funded amounting to about $1M(illion)

It is expected that neuroscience, just as other branches of natural sciences, will develop its own theoretical basis. The MTC program is seen as an instrument towards this goal, and thus needing further improvements. Therefore, the following recommendation are made:

2.1.1 Recommended Frontline Research for MTC Program: Evolution from Phenomenological Modelling to Rigorous and Tested Theory

The principle of promoting specific areas is that they should facilitate the conglomeration of knowledge into quantitative use of data. Ultimately, such integration should lead to broad concepts and mathematical theory.This is of interest and relevance to the research objectives and goals of the sponsoring agencies. Areas to be recommended for funding are:

Single Cell Modeling and Theory of Neural Coding

Single cell modeling is vital if electrophysiology and neurocomputation are to be connected

Sensorimotor Coordinate System Transformations

This approach provides the needed link between biological motor control and neurocomputation

Multielectrode Recording Technology and Theory

New methods needed to be invented to access a multitude of neurons simultaneously. However, theoretical modeling should generate well posed questions before expensive devices are built

# Brain and Body Imaging

Imaging has proven to be effective for detecting highly localized anatomical formations. Because neural function is highly dynamic, imaging may deny the ancient belief of localization of function, stimulating theory. Body imaging will provide a quantitative basis for analysis of sensorimotor function

# Functional Neuromuscular Stimulation

Replacement of sensorimotor function by neurocomputer prosthesis is a prime potential civilian application of neurobiology-based neurocomputer research and development

Mathematical Theory of CNS Function

Only mathematical theory can discern from data the mathematics intrinsic to brain functionl Theory of Unified CNS Spacetime Geometry

Sinusoidal test-signals are generally used, though time-derivatives, delays and inhibitory effects yield non- discriminative response. Theory could save animals and costs by proposing more effective methods of spacetime analysis

Theory on the Mind-Body Problem

The question whether we can understand details without a more general framework leads to the problem whether the broadest (philosophical) approach of neuroscience is appropriate

Theory and Modeling of CNS Systems

Because of the analytical nature of experimentation there is a shortage in integrative theories and models of CNS systems

Research on Emergent Properties

Theoretical weakness of neuroscience leads to creation of umbrellaterms, covering conceptual problems

Research on Neural Networks (Neurocomputing)
This field could effectively coalesce many of the above theoretical approaches

2.2 Existing Programs Relevant to Neurocomputing at NASA- ARC

Scattered projects throughout NASA address neurocomputation basically from a technological viewpoint. A major neurocomputing plan by JPl does not systematically integrate Life Science Program (neuroscience research) with Technology Research and Development at various NASA research centers. At NASA-ARC both technological and life-science aspects are pursued, but as yet with weak inter- connections.

2.2.1 Technological Approaches at NASA-ARC

# Sparse Distributed Memory Project (software approach) at Research Institute for Advanced Computer Science (RIACS)

lMassively parallel computations in aircraft design (hardware approach) at Aircraft Technology

Division of Aerospace System Directorate

# Parallel computing for high-speed computer architectures for space based intelligent systems (neurocomputing approach from technology viewpoint) at Information Sciences Division of Aerophysics Directorate

2.2.2 Life Science Approaches at NASA-ARC

Physiology of the vestibular system at Vestibular Research Laboratory at Life Science Division

Morphology of the vestibular system at Biocomputation Center at Life Science Division

3.0 Advantages and Disadvantages of a Joint NASA/NIH Neurocomputing Program

+ Advantages for NIH:

Neurocomputer applications will utilize massive investment in neuroscience research

Theory,validated or falsified by implementation, will rejuvenate experimental neuroscience

Mathematical and computational techniques will decrease dependency on animal experimentation

Disadvantages for NIH:

# Introducing and establishing theory and modeling in traditionally experimental neuroscience is difficult

+ Advantages for NASA:

Neurocomputers are computers for future aerospace applications. By concentrating on neurocomputing, NASA can serve its own longterm interests, as well as play a central role in a future US neurocomputer program

Neuroscience part of Neurocomputer research strengthens NASA's Life Science Program

Disadvantages for NASA

# Integration technology development with life science research is structurally difficult

+ Advantages for both Agencies

Neurocomputing is subcritical at both agencies; a cooperation is reinforcing

# The two agencies are intrinsically complementary in their expertise and wherewithals

# A joint program stands a chance for obtaining new funds. Shortterm investment in separate neurocomputing programs serves as seed for generating a Civilian Neurocomputing Program

Disadvantages for both Agencies

# Interagency cooperation is logistically difficult for bureaucratic reasons.

3.1 Conclusion: Consolidation of Compartmentalized Organization is Difficult

but Permits Launching Integrated Projects Necessary for Sustained

Program

* To minimize the general structural problems of interagency cooperation, create a US Interagency Neurocomputing Advisory Committee

4.0 A Neurocomputing Program: Major Issues to be Resolved

4.1 Secure Funding for Theoretical Neuroscience

Theoretical proposals compete with all (experimental) proposals. It is important to attain equal status of MTC program with other (experimental) programs

*Convert the presently ad hoc Study Section into a permanent committee

* Establish a definite budget for MTC. Even with an equal status, balance of funds for experimentation and theory will remain lopsided, although it is evident that a vocal portion of taxpayers prefer substituting animal experimentation with suitable alternatives

4.2 Establish Accountability of Theoretical Research

Theoretical work is not accountable. Funding theory would establish its accountability.

*Establish an NIH policy of fund theory first, and then evaluate performance rather than using upfront criticism to deny funding

4.3Define and Fund Integrative Projects

Launching integrative projects requires minimal organizational effort allowing creativity and resources to be focused on research

* Identify integrative research projects that are co- fundable. Neurocomputing efforts should be centered on specific CNS systems, aiming at transferring the body of knowledge into a theoretical understanding concluding in utilization by engineering. Projects such as Artificial Vestibulo-cerebellum, Retina, Cohlea, Olfactory bulb, Pyriform Cortex, Hippocampus, Colliculus are examples

4.4 Improve Evaluation Mechanisms for Theoretical Proposals

Theoretical proposals rarely get a true peer review and solution will only be possible when theoretical neuroscience reaches critical momentum.

* NIH-NIMH should create a Universities Neurocomputing Research Association, either fashioned after the NASA-affiliated Universities Space Research Association (USRA) or as a new branch of USRA, co-sponsored by NASA and NIH. Such a body could help make the Study Section a permanent and much more truly peer-review mechanism

4.5 Alleviate Dependency of Math Modelers and Theoreticians on Experimentalists

Mathematical modelers/theoreticians are existentially dependent on experimentalists

* Establish a pool of "seed-funds" for developing creative theoretical ideas to a stage where they can produce experimentally testable hypotheses. Such a two-sided arrangement would be similar to NASA's National Research Council Associateship, which provides a pool of money for bringing in scientists with creative ideas. Approval presupposes mutual agreement by the recipient and contributor, and is administered by the impartial body of NRC. It is recommended that the MTC program establish (or share) such a system

4.6 Refine Balance of Experimental and Theoretical Research

The healthy triad of experimental-, theoretical and joint proposals is missing.To resolve this problem, purposefully use the MTC neuroscience program to establish the triad of experimental, theoretical and joint proposals

* Insist on rewarding attempts with initial success in integrating theory and experimentation. Even with only a few "success stories" created the trend will spread

4.7 Intensify Interaction of Experimental and Theoretical Research

Theory and experimentation does not have enough cross-fertilization. To resolve this problem, create mechanisms for interdisciplinary access.

* Create a sabbatical program, co- sponsored and administered by NIH and NASA

4.8 Link Basic Research with Technological Development

Basic research and technological development aspects of neurocomputing are isolated from one another. Seek and fund (preferably jointly by a complementary agency) projects that involve both aspects of neurocomputing.

*Launch "Artificial CNS System" projects that encourage or even require cooperation of basic scientists and technologists



5.0 Conclusion and Recommendations

5.1 A US.Civilian Neurocomputer Initiative from the Government is Needed to Establish a Basic Research Foundation for Neurocomputing

Neurocomputing research in the US needs a government supported and coordinated civilian program complementary to the DoD (DARPA) neurocomputing initiative. Such a program would establish the vital basic research foundation necessary for neurocomputing. This leg is presently missing, and thus the current twotiered structure of neurocomputing, based on defense and business, is inherently faulty. The problem has historical roots and implications on worldwide competition.

5.1.1Slow Evolution of brain theory and modeling in Neuroscience

The above structural distortions emerged since neuroscience has not been ready with mathematical/ theoretical/ computational approaches

5.1.2 Historical Precedents: Cybernetics and Artificial Intelligence

Have not Incorporated Neuroscience

Cybernetics and Artificial Intelligence could not and would not rely on "wet" brain research. Neuro-computer research is presently coping with the same historical dilemma. Without an anchorage in neuro-sciences, "Neurocomputing" would be without proper basis. Opting for following evolution requires
l Settling down to a long haul and thus establishing appropriate scientific management

Carefully selecting systems in the biological brain that provide the best naturetested paradigms of neurocomputing: such as the vestibulo-cerebellum

Neurocomputing is in contrast in its organizational structure with two earlier stages of computer industry:

Classical computers were developed for strategic ballistic trajectory calculations, sponsored by defense. Mathematical basis was ready, development was a matter of technology required no basic research

Recent stage of computer industry is the development of home computers. This could be driven by commercial entrepreneurship as development did not require basic research

# Neurocomputer development follows these trends. This two-pillared, defense and entrepreneurial, structure of neurocomputing is incomplete since neither will support the neuroscience-oriented basic research

* Establish a third, civilian governmental pillar of neurocomputing based on an NIH-NASA cooperation (extended by NSF) that should take care of basic research

5.1.3 Implication for Worldwide Competition: Europe and Japan Organize

Civilian Neurocomputing Programs

European and Japanese neurocomputer developments are fostered entirely by civilian government organizations

Geopolitical changes cause a shift from defense expenditure to industrial competition based on R&D. A U.S. Civilian Governmental Program would complement DARPA's defense-based neurocomputing initiative. Such a program would:

Provide a balanced US neurocomputing program

Present a strong case for Congress to initiate a R&D program in the style of the Japanese MITI or German Ministry of Research & Development, in order to improve the US competitive posture in World

Such a program would be considered by ACTA, (Advanced Civilian Trade Agency; a "civilian DARPA", if the bill sponsored by Senator J.Glenn is approved and such new agency is created)

* 5.2 Establish an NIH-NASA-(NSF) US Civilian Neuro-computing Advisory Committee for Longterm Neurocomputer Research Initiative and Coordination

The civilian governmental agency that could play the spearheading role in such an initiative should:

# have a clear civilian profile yet strong connections with the defense

lconjoin the technological and basic life science aspects of neurocomputing

# include neurocomputers as a central component of its mission

NASA obviously fulfils the first two requirements. It also fulfils the third as neurocomputers are the computers for future aerospace activity because:

Parallel organization of neurocomputers yields faster performance given identical weight and dimensional constraints

Parallel organization makes neurocomputers error- tolerant, vital for systems exposed to harmful in-flight effects

Neurocomputers rely on "self-organizing" software that is simpler than conventional software, and is also the basis of autonomous intelligence

Concerning a chief mission of NASA-ARC(aerospace flight), atmospheric and space flight-control could benefit from an understanding of how natural evolution worked out biological "neuro-computers" for fast, precise and environment-adaptive control

* The Interagency Neurocomputing Advisory Committee for U.S. Civilian Governmental
Neurocomputing should be established with NASA's initiative

A mechanism is to use the standing "Interagency Working Group of NASA-NIH". This proposal should be put on their agenda for discussion and the proposed interagency Neurocomputing Advisory Committee be established. The Committee would prepare for consideration by ACTA and coordinate with DARPA a "Proposal for US. Civilian Governmental Neurocomputing Program". If NASA declines the role of initiator, ACTA (if established) could be the sponsor and administrative agency of such program, or as a contingency one of the National Laboratories could be assigned such role.

* 5.3 Broaden MTC Study Section to an Overall NIH Review Board in Order to

Allocate Centrally Created Seed-Budget for MTC Research

Beyond Institutes that jointly launched the MTC program (NIMH and NINDS), other Institutes, (e.g. NIDCD) are also interested. Therefore, this report recommends that the MTC Study Section be broadened; to be open to all neurocomputing-related proposals from all NIH Institutes.

* To cope with the increased volume of proposals it is recommended in concert with Sect.4.1. and 4.4. that the Study Section be made permanent.

Financial incentive should be created to encourage the use of this MTC Study Section by all Institutes.

* The MTC Study Section should be appropriated a separate budget. Once a proposal gets funded, its budget and administration should go to the sender Institute from which the proposal originated. Thus, neurocomputer research would be financed by new money channeled by the MTC Study Section to existing Institutes. The program is a reason for NIH to request new funds for neurocomputing, as coordinated by the Interagency Committee. The program will be NIH's "seed" for its share in a later US Civilian Neurocomputer Program.

This "nursery" program will

lensure adequate review of proposals that are notoriously difficult for individual institutes to evaluate

# establish research lines that are of potential importance to an individual Institute but which could not be established by that Institute alone

# procure additional new funds for the sender institutes

* Reviewers of MTC proposals should be named by different Institutes. For coordination and balance purposes, appointments should be confirmed by the Interagency Committee. This structure will further facilitate that initially separate neurocomputing review boards of these agencies would be merged once the US Civilian Neurocomputer Program is established.

* It is estimated that $12 M(illion) per year for the purposes of MTC is comparable to the dimensions of neuroscience-based neurocomputing spending in Europe and Japan (see Appendix) and is about what a single contractor to NASA proposes for neurocomputer technology development alone. Even with an MTC budget as recommended, spending on theory will continue to be a tiny fraction of the NIH budget.

* 5.4 Establish NASA Organization and Seed-Budget for Neurocomputing to Parallel that of NIH

It is recommended that NASA establish an organization for neurocomputing with the following features:

the structure parallels that of NIH such that an interaction and possible merger is facilitated

the program be guided and coordinated by the Interagency Committee

be aimed as a seed program used beyond a 2-year preparatory stage as leverage for taking part in a US Civilian Neurocomputer Program.

* 5.4.1 Create NASA Neurocomputer Advisory Committee and Program to Integrate Neurobiological Life Science Research with Neurocomputer Technology Development

The Life Science program does not actively participate in planned technology-oriented neurocomputer research and development (e.g. that of JPL). The "Strategy for Space Life Sciences" Report makes no specific mentioning of neurocomputing. A coordinating and organizing body should

participate in and be guided by the Interagency Committee

plan and organize neurocomputer research at NASA and coordinate it with similar activities with other agencies

ensure that Neurobiological Life Science Research component and Neurocomputer Technology Development components are balanced

structure neurocomputer research at NASA in a two-staged manner: creating "seed"-programs from its own funds for securing a share in a later US Civilian Neurocomputer Program, to be financed directly from Congress.

* 5.4.2 Organize Short Term Neurocomputer Technology Development at NASA by Allocating Neurocomputer Component to Specific Missions

Shortterm goals of neurocomputer technology development are achieved by infiltrating existing specific missions with the projected use/development of this new technology.

* 5.4.3 Organize Long Term Neurocomputer Basic Research around Integrative Artificial CNS System Projects: E.g. Establish a Neurocomputer Laboratory based on the Artificial Vestibulo- Cerebellum Project at NASA-ARC

Joint neurocomputer initiative permits launching integrated projects that require both neuroscience basic research and technological applications. Natural evolution developed the vestibulo- cerebellum for fast and coordinated movements. Development of an artificial equivalent serves the main purpose, flight control research, of NASA-ARC. The vestibulo-cerebellum is an integral part of gaze- stabilization systems. Because the overall function is so important and well-defined, and the underlying biological mechanism is relatively simple, the gaze system has been the target of intense research. As a result, the vestibulo-cerebellum is arguably the best known biological prototype of a "neurocomputer".

The advantages of an integrative Artificial Vestibulo-cerebellum project are:
# Combines neuroscience research with electronic implementation

# Provides organizational as well as scientific anchorage of neural network research

# Natural evolution guarantees scientific soundness and feasibility

# Experimental data-gathering is guided towards missing information

lUtilization is a practical check of the soundness of knowledge

# Models and theories different from the biological organism can also be utilized, these projects will be conglomerating rather than divisive

# Could be well justified as it is not pure expenditure but potentially generates return

# Brings researchers with different background and expertise together

lCan serve as a scientific platform when organizational structure cannot play such a role

# Common interest of funding such integrative project will increase the coherence among different Agencies, as they evolve from neuroscience-oriented approaches to become engineering- oriented

* 5.5 Converge NIH-NASA-(NSF) Parallel Neurocomputer Organizations by Cooperative Agreements between Civilian Governmental Agencies

During a 2-year preparatory stage NIH and NASA should establish parallel organization and finance "seed" programs or their own. The MTC Neuroscience Program Administration at NIH and NASA's Advisory Committee on Neurocomputing, both guided by the Interagency Advisory Committee, should coordinate actions during the preparatory stage and elaborate specific plans for joint initiative, and convergence or merger of their neurocomputing programs.

* 5.5.1 Use NIH-NASA-(NSF) Neurocomputer Seed-Budgets to Request from
Congress New Funds for a Joint US Civilian Neurocomputing Program

Time would be right now to come forward with a "civilian booster" of DARPA's momentarily not fully successful defense neurocomputer initiative. The report suggests proceeding with recommendations leading to the preparatory stage with the utmost expediency.

* 5.5.2 Facilitate Interactive Research Proposal Evaluation by Merged Review Board for Allocation of Merged or Separate Funds for Neurocomputer Basic Research

NIH and NASA should consider merging their review boards (and possibly their funds allocated for neurocomputer basic research) as soon as possible. It would be the responsibility of the overseeing Interagency Committee to maintain coordination and balance of the two counterparts. For further ensuring evenhandedness, it is suggested that NASA matches the yearly $12M fund recommended for NIH neurocomputing basic research, resulting in a size of the civilian neurocomputing initiative that is roughly equivalent to DARPA's defense based neurocomputing program.

* 5.5.3 Create Mechanisms to Use Manpower in an Interactive Joint Fashion: Establish Universities Neurocomputer Research Association to Administer Inter-change, Sabbatical and Conference Programs between NASA and NIH

A mechanism similar to the one recommended already exists at NASA:Universities Space Research Association (USRA). Establishing, either as its division with NIH participation, or as a similar organization set up separately, such a mechanism could serve the purposes of creating a pool of researchers suitable for peer-review system, manpower interchange for cooperative research, sabbatical and conference programs.

* 5.5.4 Utilize Intramural Facilities in an Integrative Joint Fashion by NIH-NASA

The interactive use of manpower, facilitated by the above recommendation, also serves this goal of integrative utilization of intramural NIH and NASA facilities by personnel of the other agency. Joint utilization and further development of equipment is also a distinct possibility.

* 5.5.4.1 Utilize Vestibular Research Facility at NASA-ARC to Link Neuro-physiology to Systems Modeling and Neurocomputer Analysis

* Suggested Project: 3D Skeletomuscular Systems' Modeling

* Suggested Project: Multidimensional Geometry of CNS Revealed by Multielectrode Electro-physiological Analysis

* 5.5.4.2 Utilize Biocomputing Center at NASA-ARC to Link Morphology
to Computer Modeling to Discern Neurocomputer Mathematics

* Suggested Project: Fractal Structural Geometry of Neurons and Biological Organelles Revealed by Computerized 3D Histology and Modelling