In: Proc. of the Ninth Annual Conf. of IEEE/Engineering in Medicine  
and Biology Society, Boston, 13-16 Nov. 1987

Tensor Geometry: Mathematical Brain Theory for Neurocomputers  
and Neurobots.  A Parallel Algorithm for Functional Neuromuscular 
Stimulation

A. J. Pellionisz

Department of Physiology and Biophysics, New York University  
Medical Center New York, N.Y. 10016, USA

Abstract

Conceptually and formally homogeneous quantitative brain  theory  
serves as a mathematical basis for our understanding of CNS  
function and as a medium for blueprints of constructing brain-like  
machines. Tensor Network Theory [9-10,19-22] is based on the  
axiom that neuronal networks trans-form generalized vectorial  
(tensorial) expressions of external invariants in intrinsic coordinates.  
How neural networks transform such expressions within and among  
representation-spaces (with not necessarily Euclidean or Riemannian  
geometries) can be described by tensor geometry, which thus  
becomes a unifying language of theoretical neuroscience, and  
neurobotics & neurocomputing. An example of coordinated motor  
control by neural networks provides with a parallel algorithm for  
functional neuromuscular stimulation.

Introduction

     Brain Theory, in a primary sense, enables us to arrive at  
structuro-functional principles of the operation of distinct  
subsystems of the CNS. The geometrical approach of Tensor Network  
Theory [reviews: 10,15,17] conceives brain function as a geometrical  
representation comprised by the complex structural interconnections  
of a neuronal network. Thus, neuronal nets implement functional  
geometries that match those of the external world [22]. Cerebellar  
and gaze systems and the  role they play in adaptive, coordinated  
sensorimotor  operations could already be functionally interpreted 

[11,12,14,18,20, 21,23].  Further, however, theory also leads from  
knowledge to its utilization. Mathematical neuronal network theories  
already started to yield new means, brain-like machines, both for  
production and control in the fields of neurobotics  (robots equipped  
with brain-like controllers [9,13,16]) and neuro-computers   (new  
bread of non von Neumann-type "computers" implementing brain- 
like functions; [14,16,17]).

Fig.2. Tensor network model of transformations of intrinsic  
coordinates: a model of the three-step vestibulo-collic head- 
stabilization sensorimotor reflex [23].  Transformation from  
coordinates intrinsic to a sensory system to an intrinsic motor frame  
(where the latter may be of higher dimensions) can be accomplished  
by a three-step tensorial scheme [20].  The vestibular sensory metric  
tensor, vestibulo-collic sensorimotor  tensor and contravariant neck  
motor metric tensor transformations can be expressed verbally or by  
abstract, reference frame-aspecific tensor-symbolism or by matrix-  
(patch-) and network-diagrams.  Here, these three matrices are  
shown, for the particular vestibular [3] and neck-motor frames [2] of  
the cat by patch-diagrams only. A quantitative visualization of  
corresponding neuronal networks that can accomplish such transfer  
is also presented.  Network diagrams display the massively parallel  
architecture of the CNS, where convergences & divergences are the  
rule and separated point-to-point connections rarely, if ever,  
characterize the structure. Such  existing neuronal networks that  
transform infromation from sensory frame to motor, can also be  
implemented by various man-made means, eg. parallely organized  
software and/or hardware. Neuroscience provides blueprints for  
neurocomputer algorithms.

 Fig.1. Schematic illustration of (joint-angle) coordinates intrinsic to  
the structure of a motor apparatus (A) and the required metric  
tensor transformation from a unique set of projection-type  
(covariant, B) to overcomplete parallelogram-type (contravariant, C)  
vectorial expression.

Tensor Model: Neuronal Networks Transform Intrinsic  Coordinates    
Tensorial approach necessitates three major classes of investments.   
Beyond the central task of development of the tensorial concepts,  
formalisms, software & parallel hardware (neurocomputer) systems,  
a knowledge of the neuronal net-works underlying CNS performance  
is required: abundant in neuroanatomy. Third, quantitative  
experimental establishment of the intrinsic coordinate systems is  
inevitable  to move from illustrative joint-coordinates (Fig.1) to  
anatomically correct biomechanical models of various musculo- 
skeletal systems that the CNS operates with. An exemple is given in  
Fig.2. [details in 23] showing neuronal networks that transform  
general intrinsic coordinates (established by quantitative  
computerized anatomy) from vestibular sensory- [3] to neck-motor  
frame [2].

 Metaorganization of Tensor Transformations in Co-ordinated 
Robotic-  and Living Sensorimotor Systems    Once theory is 
elaborated [9-10,19-22] and experimentally tested 

[5,23], tensor geometry can serve as a common language describing  
coordinated control both in natural (gaze, loco-motory, haptic) and  
man-made (robotic) sensorimotor systems.   Tensors are 
mathematical operators connecting general coordinates,  where one 
must distinguish between sensory- and motor-type  vectorial 
representations (covariant and contravariant tensors, [20] ).  Thus, 
the features [10] and the emergence, via the so-called  
Metaorganization-principle [23], can be specified for the metric  
tensor that comprises a multidimensional functional geometry  that  
transform the dual representations from one to another : 1) Sensory 
metric tensor (gpr), transforming a covariant reception  vector (sr) to 
contravariant perception (sp). Lower and upper indeces  denote co- 
and contravariants: sp=gpr.sf  where gpr = |gpr|-1=  |cos(½pr)|-1   
andÊ|cos(½pr)| is the table of cosines of angles among  sensory unit-
vectors. 2) Sensorimotor covariant embedding tensor (cip), 
transforming the  sensory vector (sp) into covariant motor  intention 
vector (mi).   Covariant embedding is unique, including the case of  
overcompleteness, but results in a non-executable covariant  
expression [20]:  mi = cip . sp  where  cip= ui. wp  and ui  , wp  are  
the i-th  sensory- and p-th motor unit-vector. 3) Motor metric tensor 
[20,10] that converts intention mi to  executable contravariants; me = 
gei .mi (where  gei  is computed as   gpr  was for sensory axes in 1). 
In case of overcompleteness of either  or both the sensory and motor 
coordinate systems, tensor network  theory suggests that the CNS 
uses the Moore-Penrose generalized  inverse  of the unique covariant 
metric tensor [10]:  gj,k=  Sm {1/lm+  . |Em > < Em|},  where Emand lm+ 
are the m-th Eigenvector of gj,k and  its generalized Eigenvalue	 
Fig.3.Tensor model of the coordination of a 4-joint 7-muscle  
overcomplete skeletomuscular apparatus using a Moore-Penrose  
inverse.Predicted muscle activation (right) is a basis for functional  
neuromuscular stimulation.    The progress of neurocomputation 
from Nature's networks towards  a new breed of brain-like hardware 
(neurochips or even biochips),  can be illustrated by Fig.3 (and its 
video), which projects the use of  tensorial modeling for functional 
neuro-muscular stimulation of  paraplegics Ð a research field in need 
of massively parallel control  algorithm [6,7,8].  For the neural 
paradigm of cerebellar-type motor  coordination [10,12], several 
steps have already been accomplished.  The first was to 
experimentally reveal the networks.  Second, to  arrive at mathe-
matical algorithms, eg. the metaorganization process,  that these 
networks implement [23].   Third, to build the software  (using von-
Neumann computers) for applications [13-17]. Fig.3.  shows the 
application of the tensorial "netware" utilized to gene-rate  a 
coordinated movement of thee overcomplete system of 4-joint, 7- 
muscle cat hindlimb-model. There are three further stages (steps  
4,5,6) in the realization of the metaorganization algorithm. The  
concept of "virtual instrument" (realized by graphic-based software  
on von-Neumann type computers) is a 

relatively small step to make (towards level 4, the stage of R&D)  
from the present exploratory state of art. Stage 5, future stand-alone  
traditional microprocessor implementations of such  
"neurocomputation" are likely to yield marketable products to help  
the vast community of handicapped paraplegics.  Ultimately, the  
understanding of how cerebellar neuronal networks bring about  
coordinated motor control will be matched by utilization of the  
principles by means of neurochips that will implement the  
metaorganization-algorithm by massively parallel, neurocomputer  
manner Ñ just like the living brain.  

Acknowledgement: Research supported by NS 13742 & 22999 from  
NINCDS 

 [1]  Albert,A. (1972) Regression and the Moore-Penrose  
Pseudoinverse.  NY: Academic Press

[2]  Baker,J, Goldberg,J, Peterson,B. (1985) Spatial and temporal 
response properties of the vestibulocollic reflex in decerebrate cats. J. 
Neurophysiol. 54, 735-756.

[3]  Blanks,R.H.I., Curthoys,I.S. Markham,C.H.(1972) Planar rela-
tionships of semicircular canals in the cat . Am.J.  Physiol.  223, 55-
62.

[4]  Bloedel,J.R, Dichgans,J, Precht,W. (1985)Cerebellar Functions. 
Berlin: Springer Verlag.

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

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

[7]  Kralj,A., Grobelnik,S. (1973) Functional electrical stimulation - A 
new hope for paraplegic patients? Bull. of Prosthesis Res. 10-20,75-
102.

[8]  Mauritz,K.H. (1986) Restoration of posture and gait by functional 
neuromuscular stimulation (FNS).  In Disorders of Posture and Gait,  
ed. W. Bles & T. Brandt, Amsterdam: Elsevier. pp. 367-385.

[9]  Pellionisz,A. (1983) Brain theory: connecting neurobiology to 
robotics. Tensor analysis: utilizing intrinsic coordinates to describe, 
understand and engineer functional geometries of intelligent 
organisms.  J. Theoret. Neurobiol. 2(3), 185-211.

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

[11]  Pellionisz,A. (1985) Tensorial aspects of the multidimensional 
approach to the vestibulo-oculomotor reflex and gaze.  In: Reviews of 
Oculomotor Research.  I. Adaptive Mechanisms  in Gaze Control,  ed. 
A. Berthoz & G. Melvill-Jones, Amsterdam:  Elsevier, pp. 281-296.

[12]  Pellionisz,A. (1985) Tensorial brain theory in cerebellar 
modeling.  In Cerebellar Functions,  ed.J. Bloedel, J.  Dichgans & W. 
Precht, Heidelberg: Springer,  pp. 201-229.

[13]  Pellionisz,A. (1985) Robotics connected to neurobiology by 
tensor theory of brain function. Proceedings of  IEEE Internatl. Conf. 
on Systems, Man & Cybernetics. pp. 411-414.

[14]  Pellionisz, A. (1986) Tensor network theory and its application 
in computer modeling of the metaorganization of sensorimotor 
hierarchies of  gaze.  In Proc. "Neuronal Networks for Computing"., 
AIP 151,  New York: American Institute of Physics. pp.339-344.           

[15]  Pellionisz,A. (1986) Tensor network theory of the central 
nervous system and sensorimotor modeling.  In Brain Theory: 
Proceedings of the First Trieste Meeting on Brain Theory, October 1-
	4, 1984.  ed. G. Palm, & A.Aertsen, Berlin: Springer  Verlag. pp. 
121-145.

[16]  Pellionisz,A. (1987) Sensorimotor operations: a ground for the 
co-evolution of brain theory with neurobotics and neurocomputers. 
Proc. IEEE 1st Ann. Internatl. Conf. on Neural Networks, San Diego, CA

[17]  Pellionisz,A. (1987) Vistas form Tensor Network Theory: A 
horizon from reductionalistic neurophilosophy to the geometry of 
multi-unit recordings. In: Computer Simulation in Brain Science (ed. 
by R. Cotterill) Cambridge University Press.

[18]  Pellionisz,A, Graf, W. (1987) Tensor network model of the 
"Three-neuron vestibulo-ocular reflex arc" in cat.  J. Theoretical 
Neurobiology 

[19]  Pellionisz,A, Llinás,R. (1979) Brain modeling by tensor 
network theory and computer simulation. The cerebellum: 
distributed processor for predictive coordination. Neuroscience, 
4,322-348.

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

[21]  Pellionisz,A, Llinás,R. (1982) Space-time representation 
in the brain.  The cerebellum as a predictive space-time metric 
tensor.  Neuro-science, 7,  2949-2970.

[22]  Pellionisz,A, Llinás,R. (1985) Tensor network theory of 
the meta-organization of functional geometries in CNS.Neuroscience, 
16, 245-274

[23]  Pellionisz,A, Peterson,B.W. (1987) A tensorial model of neck 
motor activation. In  Control of Head Movement, ed. B.W.Peterson & 
F. Richmond, Oxford: Oxford University Press