Title: Software engine for multiple, parallel processing with neural networks
Abstract: The hardware of the present invention must be structured with the Brownian motion equation, Bayes' equation and its matrices as integral components. All data are input to a common bus bar. All data are then sent to all nodes simultaneously. Each node will have coded gates to admit the proper data to the appropriate matrix and Bayes' equation. Then, as the data are processed, they will be sent to a central data processing unit that integrates the data in the Brownian motion equation. The output is displayed in linguistic terms or in digital form by means of fuzzy logic.
Patent Number: 7,010,513 Issued on 03/07/2006 to Tamura
| Inventors:
|
Tamura; Raymond M. (1414 Alexander St. Apartment 204, Honolulu, HI 96822)
|
| Appl. No.:
|
602627 |
| Filed:
|
June 25, 2003 |
| Current U.S. Class: |
706/15; 706/26; 706/27; 700/4 |
| Current Intern'l Class: |
G06F 15/18 (20060101); G05B 19/18 (20060101) |
| Field of Search: |
706/15,26-27,4
700/4
|
References Cited [Referenced By]
U.S. Patent Documents
| 5454064 | Sep., 1995 | Castelaz.
| |
| 5768476 | Jun., 1998 | Sugaya et al.
| |
| 5826249 | Oct., 1998 | Skeirik.
| |
| 2003/0055798 | Mar., 2003 | Hittle et al.
| |
| 2003/0074338 | Apr., 2003 | Young et al.
| |
Other References
Neural network techniques for object orientation detection. Solution by optimal
feedforward network and learning vector quantization approaches Morris, R.J.T.;
Rubin, L.D.; Tirri, H.; Pattern Analysis and Machine Intelligence, IEEE, 1990,
pps. 1107-1115.
Joint solution of low, intermediate, and high-level vision tasks by evolutionary
optimization: Application to computer vision at low SNR Bhattacharjya, A.K.; Roysam,
B.; Neural Networks, IEEE, 1994, Digital Object Identifier 10.1109/34.61712, pps. 83-95.
A parallel hybrid genetic algorithm simulated annealing approach to finding most
probable explanations on Bayesian belief networks Abdelbar, A.M.; Hedetniemi, S.M.;
Neural Networks, 1997, Digital Object Identifier 10.1109/ICNN, pps. 450-455.
Bayesian case-based reasoning with neural networks Myllymaki, P.; Tirri, H.;
Neural Networks, 1993., IEEE International Conference, 1993 pp.: 422-427 vol. 1
Digital Object Identifier 10.1109/ICNN. 1993.298594.
Tamura, R.M.; Innovative Approach to Predictive Utilization of Medical Services;
Proceedings of 6th Hawaii International Conference on Systems Sciences; Jan. 1973.
Tamura, R.M.; Use of Veitch-Karnaugh Mapping in Management; Proceedings
of 11th Hawaii International Conference on Systems Sciences: vol. II, pp 168-178;
Jan. 5-6, 1978.
Tamura, R.M.; User's View of Linear and Non-linear Medical Models; Proceedings
of Nueroequilobiometric Society Conference, Bad Kissengen, FRG; 1982.
|
Primary Examiner: Starks, Jr.; Wilbert L.
Assistant Examiner: Holmes; Michael B.
Attorney, Agent or Firm: Wray; James Creighton, Narasimhan; Meera P.
Parent Case Text
This application claims the benefit of U.S. Provisional Application No. 60/462,323,
filed Apr. 14, 2003.
Claims
I claim:
1. A method of parallel processing in a neural network system with a software
engine comprising:
providing multiple inputs of data in parallel paths,
providing converters, supplying the multiple inputs of data in the parallel paths
to the converters and converting data inputs in languages other than languages
of the software engine to languages of the software engine,
providing a primary bus and supplying the multiple inputs of data and converted
inputs of data into languages of the software engine from the converters to the
primary bus,
providing multiple gigabit primary chips connected to the primary bus and providing
the inputs of data and the converted inputs of data to the primary chips which
have multiple loci for multiple classes,
manipulating data through groups or sets of entities from the inputs and the
converted inputs in the primary chips,
manipulating data in the primary chips by means of Brownian motion equations
divided into two vectors sized for solving specific problems and predicting outcomes
relating to entities,
providing entities that are open ended for coping with unanticipated or random
factors affecting the entities,
representing element of a matrix as Bayes' equation and representing the denominator
of Bayes' equation as a summation of data from a population of entities in a similar
class or group,
connecting primary chips together for continuously updating the denominator of
Bayes' equation,
providing multiple secondary chips with multiple loci for specific entities connected
to the primary chips,
transferring the manipulated data from the primary chips to the secondary chips
connected to each of the primary chips,
placing the manipulated data in secondary chips which have Brownian motion equations
coded for a specific entity and vectors that are a series of matrices with elements
of Bayes' equation for a specific entity,
providing connections between the secondary chips and a secondary bus bar for
continuously updating denominators of a Bayes' equation,
providing multiple loci and coding each loci for a specific entity,
providing a central processor,
transferring the data in matrices or Bayes' equations from the secondary chips
to the central processor,
providing Brownian motion equations, internal and external vectors to the central
processor for evaluation of a specific entity,
processing the data in the central processor by manipulating the data in the
Brownian motion equations and matrices or Bayes' equations, and
producing results of the manipulating from the central computer.
2. The method of claim 1, wherein all inputs pass through the converters and
the bus and go to each of the primary chips simultaneously.
3. The method of claim 1, wherein the primary chips number from about several
hundred to several thousand.
4. The method of claim 1, wherein the primary chips are organized into related
groups of primary chips and wherein the secondary chips are connected to more than
one primary chip within a group.
5. The method of claim 1, wherein the primary chips are organized into related
levels of primary chips and wherein the secondary chips are connected to more than
one primary chip within a level.
6. The method of claim 1, wherein the primary chips are data banks and collectively
are a data bank of the system.
7. The method of claim 1, wherein the primary chips incorporate a Brownian motion equation.
8. The method of claim 7, wherein each Brownian motion equation is split into
first and second vectors, the first vector embodying internal factors that are
internal to an entity, event or function and the second vector embodying external
factors that are external to the entity, event or function.
9. The method of claim 8, wherein each vector is a series of matrices and each
matrix has elements that are Bayes' equations.
10. The method of claim 1, wherein each primary chip is distinct.
11. A method of parallel processing in a neural network system with a software
engine comprising:
providing multiple inputs of data in parallel paths,
supplying the multiple inputs of data in the parallel paths to converters,
converting data inputs in languages other than languages of the software engine
to languages of the software engine,
supplying the multiple inputs of data and converted inputs of data from the converters
to the bus,
providing the inputs of data and the converted inputs of data to the primary chips,
manipulating data from the inputs and the converted inputs in the primary chips,
which communicate with each other,
transferring the manipulated data from the primary chips to secondary chips,
placing the manipulated data in matrices or Bayes' equations in the secondary chips,
transferring the data in matrices or Bayes' equations from the secondary chips
to a central processor,
processing the data in the central processor by manipulating the data in Brownian
motion equations and its vectors and in the matrices or Bayes' equations, and
producing results of the manipulating from the central computer.
12. The method of claim 11, wherein all inputs pass through the converters and
the bus and go to each of the primary chips simultaneously.
13. The method of claim 11, wherein the primary chips are gigabyte chips and
number from about several hundred to several thousand.
14. The method of claim 11, wherein the primary chips are organized into related
groups of primary chips, which communicate with each other, and wherein the secondary
chips are connected to more than one primary chip within a group.
15. The method of claim 11, wherein the primary chips are organized into related
levels of primary chips and wherein the secondary chips are connected to more than
one primary chip within a level.
16. The method of claim 11, wherein the primary chips are data banks, which communicate
with each other, and collectively are a data bank of the system.
17. The method of claim 11, wherein the primary chips incorporate Brownian motion
equations and wherein the secondary chips organize data from the primary chips
in matrices of elements and wherein each element in a matrix is a Bayes' equation.
18. The method of claim 17, wherein each Brownian motion equation is split into
first and second vectors, the first vector embodying internal factors that are
internal to an entity, event or function and the second vector embodying external
factors that are external to the entity, event or function.
19. The method of claim 18, wherein each vector comprises a series of matrices
and each matrix further comprises elements that are Bayes' equations.
20. The method of claim 1, wherein each primary chip is distinct, with multiple
loci for distinct classes or groups or sets of entities.
21. A parallel processing neural network system comprising:
multiple inputs,
converters connected to the multiple inputs through parallel paths,
a bus connected to the converters through parallel paths,
related multiple gigabyte primary chips connected to the bus and arranged in
multiple groups,
multiple secondary chips connected to the primary chips with each of the secondary
chips connected to more than one of the primary chips,
a processor having processor units connected to the secondary chips,
the multiple inputs passing through the converters and the bus into the primary
chips simultaneously,
outputs from the secondary chips passing to the processor simultaneously, and
the processor processing the outputs from the secondary chips simultaneously
and producing results.
Description
BACKGROUND OF THE INVENTION
Many fields dealing with research or other endeavors require the use and manipulation
of huge data arrays. All fields that need to solve and predict problems and solutions
need methods of handling large data arrays. Current methods of dealing with large
data arrays are inadequate to properly consider and integrate data from the real world.
Past and present machines are deficient because the software for these processes
is sequential. Sequential processes are not able to handle the huge data arrays
of randomness found in the real world. Additionally, current machines are not able
to communicate with one another in order to upgrade information.
Current machines are also not able to solve problems or predict problems
and solutions in tiered structures.
Needs exist for improved methods for integrating data into large data arrays
and predicting and solving problems using these data arrays.
SUMMARY OF THE INVENTION
The present invention is a method for handling large data arrays that incorporates
data from the past and present in many fields and integrates this data for a solution.
The present invention can handle most software. It is possible to incorporate the
ubiquitous randomness in the real world and integrate data from the laboratory
and academic worlds. Formulae for accomplishing this integration are located in
the chips as integral parts. The chips communicate with each other in a neural network.
The present software engine is unique in that it is designed to deal with real
world events in any time interval, and predict events in succeeding time intervals.
Another unique feature is that the software and hardware elements are designed
as integral entities. This is different from all other computer systems where software
and hardware are designed as separate entities, generally by separate companies.
The Brownian motion equation is used to deal with real world, random events and
to predict succeeding events. This equation is divided into two vectors. Each vector
is a sum of a series of matrices that approaches infinity to avoid the problem
of coping with random events by using deterministic functions and methods. Elements
of each matrix are a Bayes' equation.
A vector is equal to the sum of all the matrices that represent the various components
of the entity. One of the two vectors can be designated by the user to represent
the vector on which external forces are acting. The other vector represents forces
acting on the entity internally.
The hardware of the present invention must be structured with the Brownian motion
equation, Bayes' equation and its matrices as integral components. All data are
input to a common bus bar. All data are then sent to all nodes simultaneously.
Each node will have coded gates to admit the proper data to the appropriate matrix
and Bayes' equation. Then, as the data are processed, they will be sent to a central
data processing unit that integrates the data in the Brownian motion equation.
The output is displayed in linguistic terms or in digital form by means of fuzzy logic.
These and further and other objects and features of the invention are apparent
in the disclosure, which includes the above and ongoing written specification,
with the drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a schematic diagram of the software engine from inputs to outputs.
FIG. 2 is a top view schematic diagram of the software engine.
FIG. 3 is an example of vector inputs for internal and external parameters.
FIG. 4 is an exploded view of a matrix system.
FIG. 5 is a diagram of a matrix system superimposed on a human to show how the
vector information is applied to a diagnosis situation.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
The present software engine is unique in that it is designed to deal with real
world events in any time interval, and predict events in succeeding time intervals.
Another unique feature is that the software and hardware elements are designed
as integral entities. This is different than all other computer systems where,
generally, separate companies design software and hardware as separate entities.
The hardware of the present invention is structured with the Brownian motion
equation, Bayes' equation and its matrices as integral components. All data are
input to a common bus bar. All data are then sent to all nodes simultaneously.
Each node will have coded gates to admit the proper data to the appropriate matrix
and Bayes' equation. Then, as the data are processed, they will be sent to a central
data processing unit that integrates the data in the Brownian motion equation.
The output is displayed in linguistic terms or in digital form by means of fuzzy logic.
Although randomness is ubiquitous in the real world, there are events and
entities that are constant. The diagonal of each matrix can be designated to manage
these events and entities. A user can code the other positions in a matrix so that
these positions can be used for events and entities that occur randomly.
All elements and entities are linked by software. This software must be able
to handle all dataflow in parallel paths simultaneously. Matrices must communicate
with each other in massively parallel manner. This communication requires a neural
architecture. Since each element is a Bayes' equation, processing requires parallel,
simultaneous communication of the elements with data stored from population data
and/or data stored from past events.
Past and present software are almost exclusively sequential software. Multiple,
parallel processing software in a neural network is a departure from these sequential
entities. Therefore, software for this engine is written anew, or the best of sequential
software is grafted/spliced together in multiple, parallel processing neural networks
to accomplish this task.
The present invention allows users with their own software to utilize the software
engine. There are commercially available systems that convert data in one software
system to another. Fuzzy logic also allows the conversion of linguistic data to
digital data and vice versa.
Another unique feature of this software engine is the integration of multiple,
parallel processing in a neural network with Veitch-Karnaugh mapping. In the real
world, there are numerous circumstances where events and entities are linked at
varying levels. The Brownian motion equation permits prediction of changes at varying
levels as well as changes in different areas within the same level. Bayes' equation
permits evaluation and prediction based on past and present data. Fuzzy logic will
make manipulation of linguistic and digital data possible.
FIG. 1 illustrates the general structure of the present software engine
1.
FIG. 2 shows the input part
2. Multiple inputs
3 can be fed
5
to a software converter
7 if software language is different from the specific
language designed for the software engine
1. Inputs
3 in the user's
language will pass through the software converter
7 without manipulation
except for converting from any user's language to the language or languages chosen
for the software engine
1.
A bus bar
11 is circular, around which hundreds or thousands of chips
15
are arranged. All inputs
9 go to every chip simultaneously. Each chip
15
handles about four gigabits or more of information. The heavy lines in FIG. 1 indicate
the paths
17 to these primary chips
21. The primary chips
21
are designated A
1, 2, . . . , B
1, 2, . . . , C
1, 2,
. . . , D
1, 2, . . . ad infinitum. Pairing of uppercase letters
indicates different levels on the circular bus bar
11. At level A, B, for
example, hundreds or thousands of chips
15 can exist. However, each chip
15 is distinct and indicated by subscript numbers A
1, A
2,
A
3, A
i, + . . . , and B
1, B
2, B
3,
B
i, + . . . These chips
15 are unique because they are the data
bank of the system. The chips
15 are the foundation of a system that can
cope with large randomness, expressed as:
P(
m, n, s+1)=1/4
[P(
m+1
, n, s)+
P(
m-1
,
n, s)+
P(
m, n+1
, s)+
P(
m, n-1
, s)] (1)
The primary chips
21 incorporate the equation 31 as though it were in
virtual reality. The Brownian motion equation is split into two vectors. These
vectors are m and n, as shown in Table 1 and FIG. 3. The m vector
33 embodies
all factors that are internal
34 to the entity, event or situation. The
n vector
35 embodies all factors that are external
36 to and have
an effect on an entity, event or situation. The size of the n
35 and m
33
vectors varies according to the size and complexity of the existing problem. The
size of the vectors corresponds to the size necessary to solve the problem and
predict outcomes related to a specific entity or entities. The entities are open-ended
to cope with unanticipated or random factors that affect the entity or entities.
Each vector is a series
41 of matrices
43 and each matrix
43
has elements
45 that are Bayes' equations. In the primary chip, the denominator
of a Bayes' equation is a summation of data from a population of entities in a
similar class or group, and primary chips are connected to each other by a secondary
bus bar so that the denominator of a Bayes' equation can be continuously updated.
Graphically, these m
33 and n
35 vectors are represented in Table
1 as follows:
| |
TABLE 1 |
| |
|
| |
##EQU1##
|
| |
|
| |
##EQU2##
|
| |
|
Each element of a matrix is a Bayes' equation that is represented in the form:
##EQU3##
As an example, the denominator,
##EQU4##
is a data storage bank or similar to long-term memory in a human hippocampus of
the brain. The numerator, P(c
i)P(E|c
i), is recent events
or the equivalent of short-term memory in the human prefrontal cortex of the brain.
This makes Bayes' equation a key to the Turing Machine.
The entities m, n, and Bayes' equation are located in the primary chips
21,
A, B, C, . . . . Vector m
33, internal factors
34, is located in
primary chips
21 A, C, E, . . . . Vector n
35, external factors
36,
is located in primary chips
21 B, D, F . . . . Vectors in and n are a series
41 of matrices
43. Each matrix can be designated by the user to include
factors integral with the entity and those factors that influence the entity from
external sources. Data in the primary chips
21 are obtained from historical
sources, data banks and population studies treated by meta analysis.
As multiple inputs
9 stream into the primary chips
21, they are
matched with data already present in the chips. If they match these data, then
they are manipulated in the same manner as the stored data. The stored data is
located in the diagonals of each matrix. Inputs
9 that are not matched are
placed in other elements of the matrix.
Once data inputs are manipulated in the primary chips
21, data are sent
to secondary chips
47. The secondary chips
47 are each associated
with a primary chip, such as A
1, 2, . . . and B
1, 2, . . . ,
and are associated with p
1, p
2, p
3, p
i,
+ . . . . Other primary chips, such as C
1, 2, . . . and D
1, 2,
. . . , are associated with t
1, t
2, t
3,
t
i, + . . . , and E
1, 2, . . . and F
1, 2, . . . are
associated with h
1, h
2, h
3, h
i, + .
. . . In the secondary chips
47, data are placed in the appropriate matrices
and Bayes' equations. The secondary chips are connected to the secondary bus bar
so that the denominator of a Bayes' equation for a specific entity can be updated
continuously. A secondary chip
47 has multiple loci and each loci is coded
for a specific entity. After data is placed in the appropriate matrices and Bayes'
equations, the stream of matrices and Bayes' equations are sent to the central
processing unit
49 where they are manipulated in the Brownian motion equation.
The result of this manipulation is P(m, n, s+1) displayed
51 in linguistic
or digital form or a combination of these forms. The central processing unit
49
uses the Brownian motion equation and internal and external vectors that are divided
into series of matrices with elements of Bayes' equation for a specific entity.
There is a secondary bus bar
53, shown in FIGS. 1 and 2, that connects
the primary chips
21 at all levels. All primary chips
21, that deal
with information from data banks and population studies with meta analysis or recently
acquired data, have automatic updating of each Bayes' equation as determined by
the coding of information bits. Loop circuits and learning algorithms make this possible.
Diagnosis of human conditions, disease or dysfunctions is an example that
illustrates the architecture and operations of the present software engine. In
particular, the diagnosis of hypertension shows the operation of the system. When
a patient is first seen in an office, clinic or hospital, his or her vital signs
or demographics are coded into the software to define the patient as a specific
entity or object. Blood pressure, pulse, heart rate, physical examination findings
and laboratory findings, as well as negative findings are entered into the packet
of information defining the patient.
During the data generation, each bit is designated according to whether it
belongs to the vector m, internal factors, or the vector n, external factors. Internal
factors can include such factors as type A personality, genetics/ethnicity, diseases,
such as diabetes, nephritis, hypercholesterolemia/hyperlipidemia, etc. External
factors can include such factors as job stress, divorce, errant children, socio-economic
problems, difficulties with police, alcohol or drug addictions, etc.
FIG. 3 shows inputting of data into the m
33 and n
35 vectors.
In the primary chips
21, for vector m
33, internal factors
34
are A, C, E, . . . , and in the primary chips
21 for vector n
35,
external factors
36 are B, D, F, . . . . The first data packet is processed,
each component
45 of each element of a matrix is a Bayes' equation
31
and it is compared to population studies defined by meta analysis. Once these calculations
are done, the packet is sent to a chip or locus, p
1, p
1,
. . . , p
n, in a large array that is specific for that patient. P
1,
p
1, . . . , P
n receive data packets from primary chips A
and B. Then the data packet is sent to the central processing unit where data are
manipulated by the Brownian motion equation:
P(
m, n, s+1)=1/4
[P(
m+1
, n, s)+
P(
m-1
,
n, s)+
P(
m, n+1
, s)+
P(
m, n-1
, s)] (3)
Where P(m, n, s+1) probability of hypertension.
This probability is treated by fuzzy logic so that it can be presented as an
output in linguistic terms or digital form or a combination of both.
FIG. 4 shows the hypertension example illustrating the software engine linkage
to Veitch-Karnaugh mapping
55. Squares are either 0 or 1. In the illustration,
the squares that do not have a one,
57, contain a value of zero,
59.
The equation that expresses this particular configuration is:
f={overscore (x)}1{overscore (x)}
3{overscore (x)}
5Vx1x
3x
5V{overscore
(x)}
3x
4{overscore (x)}
5x
6 (4)
FIG. 5 shows the Veitch-Karnaugh
55 mapping overlaid on a human
61
to show the application of the matrix values to a real world system.
Each patient's data packet can be sent to the Veitch-Karnaugh mapping system
55 as well as the output screen
51. Here, the Brownian motion equation
is used in a slightly different form from that used to diagnose hypertension. However,
the patient's data packet is still utilized. Other factors, such as cardiac ejection,
arterial and venous vectors, pulmonary circulation vectors, pulse pressure, etc.
are also included to determine cardiovascular vectors in the system as compartmentalized
by Veitch-Karnaugh mapping. The probability, P(m, n, s+1), being one in any square
can be determined by fuzzy logic since the probability can be measured from 0,
0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0. Squares with zero can be
treated similarly. Output from the Veitch-Karnaugh mapping
55 can be displayed
on a Computer Assisted Design (CAD) monitor. A windows-type of display can be part
of this screen that displays patient parameters.
Shortening or lengthening the bus bar
1, as shown in FIG. 1, creates
modifications of the architecture. The bus bar
11 can be modularized to
fit the purpose of the user and applications. The secondary bus bar
53 can
also be modularized.
To build architecture for a massively parallel supercomputer, the bus bar
11
can be enlarged with a central core of liquid nitrogen. The material for the bus
bar
11 will be made of a high temperature superconductor. The bus bar
11,
primary
21 and secondary
47 chips, as well as central processing
units
49 will be surrounded by liquid nitrogen cooling coils. The secondary
bus bar
53 will also be made of high temperature super conducting materials.
Massive cooling permits the use of Josephson junctions and circuits. As in the
smaller computer, the bus bars can be modularized so that their lengths can be
increased. In addition, many bus bars can be linked in parallel and in a neural
network for a massive architecture.
The example of using this system in hypertension diagnosis, treatment and outcome
is only one of many applications in many diverse fields. The following is a partial
list of applications:
1. Medicine
2. Biomedical technology
3. Chemistry.
4. Drug design
5. Toxicology
6. Business management
7. Earthquake prediction
8. Biotechnology
9. Personnel selection, training and upgrading
10. Air traffic control and management
11. Global decision making
12. Domestic decision making
13. Crime prevention and detection
14. Food and water supply and demand
15. Weather prediction
16. Oil exploration
17. Environmental pollution
18. Plasma physics
19. Stock market prediction
20. Computer security
21. Aerodynamics
While the invention has been described with reference to specific embodiments,
modifications and variations of the invention may be constructed without departing
from the scope of the invention.
*
