Claims
- 1. A neural architecture for solving a partial differential equation in a state vector consisting of at least one term in a spatial partial derivative of said state vector and a term in a partial time derivative of said state vector, comprising:
- means for storing an array of matrix elements of a matrix for each partial derivative term of said equation, said matrix relating said partial derivative to said state vector;
- means for multiplying all rows of said matrix elements by corresponding elements of said state vector simultaneously to produce terms of a product vector; and
- means for combining said product vector with a previous iteration said state vector to produce a next iteration of said state vector.
- 2. The neural architecture of claim 1 wherein said equation comprises plural spatial partial derivative terms and said neural architecture comprises plural arrays of matrix elements of matrices corresponding to said partial spatial derivative terms and plural means for multiplying, said neural architecture further comprising means for combining the product vectors produced by said means for multiplying.
- 3. The neural architecture of claim 1 wherein said partial spatial derivative terms definable over a spatial grid of points and said matrix elements are obtained in a pseudo-spectral analytical expression relating each matrix element to the distance between a corresponding pair of grid points.
- 4. The neural architecture of claim 3 wherein the matrix element of column m and row n of the matrix is: ##EQU25## where x.sub.n and x.sub.m are the corresponding pair of grid points, d is the order of the derivative and k is a one of a set of integers.
- 5. The neural architecture of claim 1 wherein said means for storing said array comprises:
- an array of N rows and M columns of CCD matrix cell groups corresponding to a matrix of N rows and M columns of matrix elements, each of said matrix elements representable with b binary bits of precision, each of said matrix cell groups comprising a column of b CCD cells storing b CCD charge packets representing the b binary bits of the corresponding matrix element, the amount of charge in each packet corresponding to one of two predetermined amounts of charge.
- 6. The neural architecture of claim 5 wherein said means for multiplying comprises:
- an array of c rows and M columns CCD vector cells corresponding to a vector of M elements representable with c binary bits of precision, each one of said M columns of CCD vector cells storing a plurality of c charge packets representing the c binary bits of the corresponding vector element, the amount of charge in each packet corresponding to one of two predetermined amounts of charge; and
- multiplier means operative for each one of said c rows of said CCD vector cells for generating, for each one of said rows of matrix CCD cells, a signal corresponding to the sum of all charge packets in both (a) said one row and (b) columns of said matrix CCD cells for which the corresponding vector CCD cell contains a charge packet of a predetermined amount.
- 7. A neural architecture for detecting moving targets in images sensed by an array of sensors, comprising:
- means for storing an array of matrix elements of a matrix for each partial spatial derivative term of a partial differential equation in a state vector consisting of at least one term in a spatial partial derivative of said state vector and a term in a partial time derivative of said state vector, each matrix relating the corresponding spatial partial derivative to said state vector;
- means for multiplying all rows of said matrix elements of each matrix by corresponding elements of said state vector simultaneously to produce a product vector for each matrix; and
- means for combining the product vectors with (a) a previous iteration said state vector and (b) a vector derived from the outputs of said sensors, to produce a next iteration of said state vector.
- 8. The neural architecture of claim 7 wherein said equation is a soliton equation having first and third order partial spatial derivatives of said state vector, whereby said neural architecture comprises means for storing respective arrays corresponding to said first and third order derivatives, whereby said means for multiplying produces two product vectors corresponding to said first and third order derivatives, respectively, said neural architecture further comprising means for multiplying the product vector of said first order derivative by said state vector.
- 9. The neural architecture of claim 8 wherein said partial spatial derivative terms are definable over a spatial grid of points corresponding to said sensor array and said matrix elements are obtained in a pseudo-spectral analytical expression relating each matrix element to the distance between a corresponding pair of grid points.
- 10. The neural architecture of claim 9 wherein the matrix element of column m and row n of the matrix is: ##EQU26## where x.sub.n and x.sub.m are the corresponding pair of grid points, d is the order of the derivative and k is a one of a set of integers.
- 11. The neural architecture of claim 8 wherein said means for storing said array comprises:
- an array of N rows and M columns of CCD matrix cell groups corresponding to a matrix of N rows and M columns of matrix elements, each of said matrix elements representable with b binary bits of precision, each of said matrix cell groups comprising a column of b CCD cells storing b CCD charge packets representing the b binary bits of the corresponding matrix element, the amount of charge in each packet corresponding to one of two predetermined amounts of charge.
- 12. The neural architecture of claim 11 wherein said means for multiplying comprises:
- an array of c rows and M columns CCD vector cells corresponding to a vector of M elements representable with c binary bits of precision, each one of said M columns of CCD vector cells storing a plurality of c charge packets representing the c binary bits of the corresponding vector element, the amount of charge in each packet corresponding to one of two predetermined amounts of charge; and
- multiplier means operative for each one of said c rows of said CCD vector cells for generating, for each one of said rows of matrix CCD cells, a signal corresponding to the sum of all charge packets in both (a) said one row and (b) columns of said matrix CCD cells for which the corresponding vector CCD cell contains a charge packet of a predetermined amount.
Parent Case Info
This is a division of application Ser. No. 08/049,829, filed Apr. 19, 1993 now U.S. Pat. No. 5,491,650.
ORIGIN OF THE INVENTION
The invention described herein was made in the performance of work under a NASA contract, and is subject to the provisions of Public Law 96-517 (35 USC 202) in which the contractor has elected to retain title.
US Referenced Citations (14)
Non-Patent Literature Citations (4)
Entry |
Barhen et al., New directions in massively parallel neurocomputing, Fifth international conference. Neural networks and their applications, pp. 543-554. Nov. 6, 1992. |
Sage et al., MNOS/CCD circuits for neural network implementations, IEEE international symposium on circuits and systems, pp. 1207-1209 May 11, 1989. |
Schwartz et al., A programmable analog neural network chip, IEEE journal of solid-state circuits, vol. 24, issue 2, pp. 313-319 Apr. 1989. |
Coolen et al., Evolution equations for neural networks with arbitrary spatial structure, First IEEE International conference on artificial neural networks, pp. 238-241 Oct. 16, 1989. |
Divisions (1)
|
Number |
Date |
Country |
Parent |
49829 |
Apr 1993 |
|