(19)
(11)EP 1 212 727 B1

(12)EUROPEAN PATENT SPECIFICATION

(45)Mention of the grant of the patent:
17.12.2003 Bulletin 2003/51

(21)Application number: 00952231.9

(22)Date of filing:  25.07.2000
(51)International Patent Classification (IPC)7G06K 9/66
(86)International application number:
PCT/US0020/479
(87)International publication number:
WO 0100/8094 (01.02.2001 Gazette  2001/05)

(54)

TRAINABLE ADAPTIVE FOCUSED REPLICATOR NETWORK FOR ANALYZING DATA

TRAINIERBARES, ANPASSUNGFÄHIGES FOKUSSIERTES REPLIKATORNETZWERK ZUR DATENANALYSE

RESEAU REPLICATEUR FOCALISE ADAPTATIF ENTRAINABLE DESTINE A L'ANALYSE DE DONNEES


(84)Designated Contracting States:
AT BE CH CY DE DK ES FI FR GB GR IE IT LI LU MC NL PT SE

(30)Priority: 26.07.1999 US 145593 P

(43)Date of publication of application:
12.06.2002 Bulletin 2002/24

(73)Proprietor: Marti Nelson Cancer Research Foundation
Vacaville, CA 95687 (US)

(72)Inventor:
  • MALYJ, Wasyl
    Davis, CA 95616 (US)

(74)Representative: Schüssler, Andrea, Dr. 
Kanzlei Huber & Schüssler Truderinger Strasse 246
81825 München
81825 München (DE)


(56)References cited: : 
  
  • ALDRICH C: "Visualization of transformed multivariate data sets with autoassociative neural networks" PATTERN RECOGNITION LETTERS, JUNE 1998, ELSEVIER, NETHERLANDS, vol. 19, no. 8, pages 749-764, XP002157124 ISSN: 0167-8655
  • SCHWENK H ET AL: "Transformation invariant autoassociation with application to handwritten character recognition" ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS 7, PROCEEDINGS OF NIPS94 - NEURAL INFORMATION PROCESSING SYSTEMS: NATURAL AND SYNTHETIC, DENVER, CO, USA, 28 NOV.-3 DEC. 1994, pages 991-998, XP000978424 1995, Cambridge, MA, USA, MIT Press, USA ISBN: 0-262-20104-6
  • LON-CHAN CHU ET AL: "OPTIMAL MAPPING OF NEURAL-NETWORK LEARNING ON MESSAGE-PASSING MULTICOMPUTERS*" JOURNAL OF PARALLEL AND DISTRIBUTED COMPUTING,US,ACADEMIC PRESS, DULUTH, MN, vol. 14, no. 3, 1 March 1992 (1992-03-01), pages 319-339, XP000261259 ISSN: 0743-7315
  • NOBUO UEKI ET AL: "EXPRESSION ANALYSIS/SYNTHESIS SYSTEM BASED ON EMOTION SPACE CONSTRUCTED BY MULTILAYERED NEURAL NETWORK" SYSTEMS & COMPUTERS IN JAPAN,US,SCRIPTA TECHNICA JOURNALS. NEW YORK, vol. 25, no. 13, 1 November 1994 (1994-11-01), pages 95-107, XP000526366 ISSN: 0882-1666
  • C.M. BISHOP: "Neural Networks for Pattern Recognition Chapter 8.6.2", 1997, OXFORD UNIVERSITY PRESS INC., NEW YORK
  
Note: Within nine months from the publication of the mention of the grant of the European patent, any person may give notice to the European Patent Office of opposition to the European patent granted. Notice of opposition shall be filed in a written reasoned statement. It shall not be deemed to have been filed until the opposition fee has been paid. (Art. 99(1) European Patent Convention).


Description

BACKGROUND OF THE INVENTION


A. Field Of The Invention



[0001] The invention relates to replicator networks trainable to create a plurality of basis sets of basis vectors used to reproduce data for confirming identification of the data.

B. Description Of the Related Art



[0002] Computers have long been programmed to perform specific functions and operations by means of sophisticated computer programming. However, in order to distinguish between data having similar features, human intervention is often required to make decisions about identification, categorization and/or separation of such data. There are no automated analysis systems that can perform sophisticated classification and analysis tasks at levels comparable to those of skilled humans.

[0003] A computer is, in essence, a number processing device. In other words, the basic vocabulary computers use to communicate and perform operations consists of numbers, mathematical relationships and mathematically expressed concepts. One portion of a computer's vocabulary includes basis vectors.

[0004] Basis vectors play an important role in multimedia and telecommunications. For instance, the images transmitted across the Internet and digital satellite television use powerful data compression technologies that encode images and sounds using predetermined sets of basis vectors to reduce the size of the transmitted data. After transmission, the encoded images and sounds are decoded by a receiver using the predetermined basis vector sets. By using pre-determined basis vectors in the transmission and reception of images and sounds, the data can be stored and transmitted in a much more compact form. Typical data compression techniques using codecs (coders and decoders) using basis vector sets include:

JPEG & NWEG codecs - - cosine waves form the basis vector sets,

Wavelet codecs - - wavelets form the basis vector sets, and

Fractal codecs - - fractals form the basis vector sets.



[0005] FIG. 1 is a grey-scale rendering of the basis vector set used in the JPEG compression technique. FIG. 1 shows an 8x8 array of basis vectors, each basis vector being a two-dimensional cosine wave having a different frequency and orientation. When an object image is to be transmitted over the Internet, the JPEG coder identifies a combination of these basis vectors that, when put together, define each section of the object image. Identification of the combination of basis vectors are transmitted over the Internet to a receiving computer. The receiving computer reconstructs the image using the basis vectors. In any given image, only a relatively small subset of basis vectors are needed in order to define the object image. The amount of data transmitted over the Internet is greatly reduced by transmitting identification of the basis vectors compared to transmitting a pixel by pixel rendering of the object image. The basis vectors in the JPEG technique are the limited vocabulary used by the computer to code and decode information. Similar basis vector sets are used in other types of data transmission, such as NV3 audio files. The smaller the vocabulary is, the more rapid the data transmission. In data compression, each data compression technique has its own pre-determined, fixed set of basis vectors. These fixed sets of basis vectors are the vocabulary used by the compression technique. One of the primary purposes of the basis vector sets in data compression is to minimize the amount of data transmitted, and thereby speeding up data transmission. For instance, the JPEG data compression technique employs a predetermined and fixed set of basis vectors. Cellular telephone data compression techniques have their own unique basis vectors suitable for compressing audio signals.

[0006] Traditionally basis vectors have been, in essence, a vocabulary used by computers to more efficiently compress and transmit data. Basis vectors may also be useful for other purposes, such as identification by computers of information. However, if an identification system is provided with a limited vocabulary, then only limited types of data are recognizable. For instance, identification systems have been developed to scan and recognize information in order to sort that information into categories. Such systems are preprogrammed to recognize a limited and usually very specific type of data. Bar code readers are a good example of such systems. The bar code reader is provided with a vocabulary that enables it to distinguish between the various width and spaces between bars correlating to a numeric value. However, such systems are fixed in that they can only recognize data pre-programmed into their computer systems. Once programmed, their function and vocabulary are fixed.

[0007] One system which is not limited to specific type of data is known from Schwenk et al "Transformation Invariant Autoassociation with Application to Handwritten Character Recognition", Advances in Neural Information Processing System 7, Proceedings of NIPS94-Neural Information Processing Systems: Natural and Synthetic, Denver, Colorado, US, 28 Nov- 3 Dec, 1994. Cambridge, MA, US: MIT Press, 1995, pages 991-998. This publication teaches a method for using a plurality of neural network to learn a mapping between an input and a corresponding output space. Each one of the plurality of neural networks is trained for only one class and therefore preserves optimally the information of the example of the one class alone. The learned networks can be used like discrimnant functions: the reconstruction error for the network is in general much lower for examples of the learned class than for other ones of the classes. Using the networks taught in this article, unknown input data can be classified as falling into one of the classes. The classification is based on the idea that the unknown input data must fall into the class in which the corresponding network gives the smallest error (or smallest distance)

[0008] The Schwenk at al. article always attempts to classify unknown input data into one of the classes for which the networks have been trained. There is no provision for the possibility that the unknown input data might fall into a class for which the networks have not been trained. Furthermore there is no provision for the case in which the replication errors between two or more networks are so similar that it is unclear into which of the classes the unknown input data might fall.

[0009] Another type of pre-programmed recognition system is in genome-based research and diagnostics. Specifically, sequencers have been developed for analyzing nucleic acid fragments, and for determining the nucleotide sequence of each fragment or the length of each fragment. Both Perkin-Ehner Corporation and Pharmacia Corporation currently manufacture and market such sequencer devices. In order to utilize such devices, a variety of different procedures are used to break the nucleic acid fragment into a'variety of smaller 20 portions. These procedures include use of various dyes that label predetermined nucleic acids within the fragment at specific locations in the fragment. Next, the fragments are subjected to gel electrophoresis, subjected to laser light by one of the above mentioned devices and the color and intensity of light emitted by the dyes is measured. The color and intensity of light is then used to construct an electropherograxn of the fragment under analysis.

[0010] The color and intensity of light measured by a device indicates the presence of a dye further indicating the location of the corresponding nucleic acid within the sequence. Such sequencers include scanners that detect fluorescence from several illuminated dyes. For instance, there are dyes that are used to identify the A, G, C and T nucleotide extension reactions. Each dye emits light at a different wavelength when excited by laser light. Thus, all four colors and therefore all four reactions can be detected and distinguished in a single gel lane.

[0011] Specific software is currently used with the above mentioned sequencer devices to process the scanned electropherograms. The software is pre-programmed to recognize the light pattern emitted by the pre-designated dyes. The vocabulary of the software is limited to enable the system to recognize the specific patterns. Even with pre-designated pattems and logical results, such systems still require human intervention for proper identification of the nucleic acid sequences under study. Such systems yield significant productivity enhancements over manual methods, but further improvements are desirable.

[0012] There exists a need for a reliable, expandable and flexible means for identifying and classifying data. In particular, there is a need for more flexible identification systems that can be easily enhanced for identification of new and differing types of data.

SUMMARY OF THE INVENTION



[0013] One object of the invention is to provide a simple and reliable system for identifying data.

[0014] Another object of the present invention is to provide a data classification system with more flexible means for identifying and classifying data.

[0015] The invention as defined by the appended claims relates to a method and apparatus that is trainable to classify and identify data. The invention includes a training step inputting several identified data sets into a computer and creating within the computer a plurality of replicators having unique basis vector sets. Each basis vector set includes a plurality of basis vectors in one to one correspondence with each of the identified data sets.

[0016] After completion of the training step, new data is input into the computer which replicates the new data by each of the replicators. The replicated data from a first one of the replicators is compared with the new data to determine accuracy of the replicated data replicated by the first one of the replicators to see whether it is within a predetermined error threshold. This replication step is repeated for each ones of the replicators. A determination is made to see if one or more of the comparisons yielded confirmation of accurate replication for one or more of the replicators. Finally the new data is classified in response to the determination of accurate replication by only one replicator. Should none or more than one replicator accurately replicate the new data, then this new data is tagged for manual processing.

[0017] In one embodiment of the invention, the training step involves the creation of a comparison data set using only the basis vector sets. A comparison is made between each of the identified data sets and the corresponding comparison data set, thereby generating error signals for each comparison. A determination is made to determine the acceptability of the error. Once the error is determined to be acceptable, the training step is completed.

[0018] The foregoing and other objects, features, aspects and advantages of the present invention will become more apparent from the following detailed description, when taken in conjunction with the accompanying drawings.

BRIEF DESCRIEPTION OF THE DRAWINGS



[0019] 

FIG. 1 is a graphical representation of a gray scale JEPG palette used by computers to compress image data;

FIG. 2 is a schematic representation of a plurality of computers in a network in accordance with one embodiment of the present invention;

FIG. 3 is a graphical representation of an adaptive focused replicator network in a training phase;

FIG. 4 is a graphical representation of a plurality of adaptive focused replicator networks in the training phase;

FIG. 5 is a graphical representation of an adaptive focused replicator network in a replicating phase;

FIG. 6 is a flowchart showing various steps of the training phase and replicating phase of the adaptive focused replicator network;

FIG. 7 is flowchart showing various steps of an identification phase of the adaptive focused replicator network;

FIGS. 8A, 8B and 8C show chirps input references sets 1, II and III, demonstrating first example of the present invention;

FIGS. 9A, 9B and 9C show input of reference set I, replicas of reference set I created by AFRN I during a training phase and an output error generated by comparing reference set I and the replicas generated by AFRN I;

FIGS. 10A, 10B, and 10C show a novel (unknown) input having features similar to reference set I, a replica of the novel input generated by trained AFRN I and output error;

FIGS. 11A, 11B and 11C showing input reference vectors II (input into AFRN I), the output from AFRN I attempting to replicate input reference vectors II and the output error demonstrating the inability of AFRN I to replicate reference vectors II;

FIGS. 11D, 11E and 11F showing input reference vectors III (input into AFRN I), the output from AFRN I attempting to replicate input reference vectors III and the output error demonstrating the inability of AFRN I to replicate reference vectors III;

FIGS. 12A, 12B and 12C show input of reference set II, replicas of reference set II generated by AFRN II during the training phase and an output error generated by comparing reference set II and the replicas created by AFRN II;

FIGS. 13A, 13B and 13C show a novel (unknown) input having features similar to reference set II, a replica of the novel input generated by trained AFRN II and output error;

FIGS. 14A, 14B and 14C showing input reference vectors I (input into trained AFRN II), the output from AFRN II attempting to replicate input reference vectors I and the output error demonstrating the inability of AFRN II to replicate reference vectors I;

FIGS. 14D, 14E and 14F showing input reference vectors III (input into AFRN II), the output from AFRN II attempting to replicate input reference vectors III and the output error demonstrating the inability of AFRN II to replicate reference vectors III;

FIGS. 15A, 15B and 15C show input of reference set III, replicas of reference set III generated by AFRN III during the training phase and an output error generated by comparing reference set III and the replicas created by AFRN III;

FIGS. 16A, 16B and 16C show a novel (unknown) input having features similar to reference set III, a replica of the novel input generated by trained AFRN III and output error;

FIGS. 17A, 17B and 17C showing input reference vectors I (input into trained AFRN III), the output from AFRN III attempting to replicate input reference vectors I and the output error demonstrating the inability of AFRN III to replicate reference vectors I;

FIGS. 17D, 17E and 17F showing input reference vectors II (input into AFRN III), the output from AFRN III attempting to replicate input reference vectors II and the output error demonstrating the inability of AFRN III to replicate reference vectors II;

FIGS. 18A, 18B, 18C, 18D and 18E show basis vectors generated by AFRN I during the training phase and used to replicate inputted data;

FIG. 18F shows the basis vectors depicted in FIGS. 18A, 18B, 18C, 18D and 18E combined in a single graph,

FIGS. 19A, 19B, 19C, 19D and 19E show basis vectors generated by AFRN II during the training phase and used to replicate inputted data;

FIG. 19F shows the basis vectors depicted in FIGS. 19A, 19B, 19C, 19D and 19E combined in a single graph;

FIGS. 20A, 20B, 20C, 20D and 20E show basis vectors generated by AFRN III during the training phase and used to replicate inputted data;

FIG. 21 is a graph showing an output from a nucleic acid sequencing device showing representation of indications of four dyes labeling nucleic acids;

FIGS. 22A, 22B and 22C are classification reference vectors I, II and III produced from scanning the known length of nucleotide sequences;

FIGS. 23A, 23B and 23C are basis vectors generated by genotyping AFRNs I, II and III, respectively;

FIGS. 24A, 24B and 24C are classification reference vectors I input into trained genotyping AFRN I, replications of classification reference vectors I generated by AFRN I, and replication error produced by comparing classification reference vectors I and the replications of classification reference vectors I;

FIGS. 25A, 25B and 25C are classification reference vectors II input into trained genotyping AFRN I, replications of classification reference vectors II generated by AFRN I, and replication error produced by comparing classification reference vectors II and the replications of classification reference vectors II (revealing that one vector is actually a class I vector);

FIGS. 26A, 26B and 26C are classification reference vectors III input into trained genotyping AFRN I, replications of classification reference vectors III generated by AFRN I, and replication error produced by comparing classification reference vectors III and the replications of classification reference vectors III; and

FIG. 27 is a flowchart showing another embodiment of the invention, showing steps for automatically refining and expanding the data analysis system depicted in FIGS. 6 and 7.


DETAILED DESCRIPTION OF THE INVENTION



[0020] The present invention relates to adaptive focused replicator networks (AFRNS) that are trainable in order to distinguish between a plurality of differing data sets, thereby providing a means for separating and classifying such data. On a small scale, AFRNs may be utilized in a single computer performing the techniques described below, or may be utilized in a system having a plurality of computers connected to one another, as shown schematically in FIG. 2 where a plurality of microprocessors MP1, MP2 through MPN are linked together via a Negotiator. Such connected computers may include a cluster network, a SMP system (Symmetric Multi-Processor system), Asymmetric MultiProcessor systems or a dedicated firmware and hardware system such as GRAPE (Gravity Pipe), which is used to greatly accelerate calculations in computational astrophysics. It should be understood that the plurality of microprocessors may be linked to one another via the negotiator or at the chip level or in clusters.

[0021] AFRNs include a plurality of array elements, each array element being trainable to recognize and classify one specific type of data, as is described in greater detail below. One group of array elements may be effected in single computer or may be divided in the group of microprocessors in FIG. 2, depending upon the size and requisite speed of the data classification system of the present invention.

[0022] The present invention works as follows. A custom basis vector set made up of one or more basis vectors is generated from a reference set of input data. Then, new data is inputted and the custom basis vectors are used to reconstruct (replicate) the input data. The custom set of AFRN basis vectors permits accurate reconstruction of closely similar data but prevents faithful reconstruction of data that differ from the reference set in one or more important details.

[0023] AFRNs use basis vector sets but not in the manner typically associated with data compression. In the present invention, each array element in an AFRN creates its own unique combination of basis vectors creating a basis vector set that is later used to identify and thereby classify unknown inputted data. In the present invention, basis vector sets are constructed in order to maximize a computer system's ability to recognize and classify the type of data inputted. There is no limit in the present invention to the number of basis vectors that may be selected in order to define the various basis vector sets. However, as discussed further below, each array element of the present invention is trained to have its own unique individual basis vector set having a limited group of basis vectors used for replicate a specific type or group of data.

[0024] In accordance with one embodiment of the present invention, a method for classifying types of data in a data set includes two fundamental phases: a learning phase and a replicating phase.

[0025] In the learning phase, previously categorized sets of data C1 through CN are inputted into a computer or network of computers. In order to explain the training phase, only one data set is considered in FIG. 3. One of the sets of data, for instance, data set C1, is inputted into the AFRN array element E1, as shown in FIG. 3. The array element E1 then generates a basis vector set having a plurality of basis vectors based upon data set Cl. The basis vectors are determined in any of a variety of analytical or empirical ways. For instance, the data elements in the data set C, are solvable as linear equations to determine the basis vectors.

[0026] The basis vectors in the basis vector set may be any of a combination of basis vector types such as: chirplets, wavelet, Fourier based functions, Fractal and/or radial basis functions (BFs), basis functions generated using MLP (multi-layer perceptrons) or other basis vectors determined by the computer and assigned to the array element E1. For instance, if the data set C1 includes audio signals, then the basis vectors will likely be solved by the computer and made part of element E1 as a plurality of Fourier based functions. If the data set C1, includes images, then the basis vectors determined by the element E1 will likely, but not necessarily, be two dimensional cosine related basis vectors similar to the above described basis vectors used in the JPEG compression technique (FIG. 1).

[0027] As indicated in FIG. 3, as a part of the training phase, the array element E1 constructs a replicate data set C1, The replicate data set C1r, is then compared with the elements of original data set C1, and an error value is generated, at evaluation point L1. The error is evaluated to determine whether the error is acceptable or unacceptable. If acceptable, then the array element E1 is trained. If unacceptable, then the generation or refinement of basis vectors continues.
Similarly, although not shown in FIG. 3, an array element E2 generates a basis vector set based upon a data set C2, and so on until array element EN has generated a basis vector set for data set CN.·A plurality of AFRN arrays, each having a plurality of array elements are depicted in FIG. 4, each array element trained in a manner described above with respect to FIG. 3, but with each array element trained to replicate a differing data set. As is indicated in FIG. 4, the plurality of AFRN arrays may be grouped by data types. For instance, one AFRN array may have array elements that are trained to replicate and classify one data class having multiple sub-classes of data, another AFRN array may have array elements trained to replicate and classify another data class having a only a few subclasses of data. Specifically, in an optical recognition application of the AFRN arrays of the present invention, one AFRN array in FIG. 4 may be for replicating and classifying different hues of blue, another AFRN array may be trained for replicating and classifying different hues of red, and so on.

[0028] Once the complete array of elements has been trained, it is now possible to classify new data by passing the data through all of the array elements. In other words, each data element of an unknown data set is considered by each and every array element in one AFRN array. As indicated in FIG. 5, all the data in unknown data set U is inputted into the AFRN array elements E1 through EN. The array element E1 then attempts to generate a replica U1r of each and every data element of unknown data set U. At the evaluation point L1, a comparison is made between the replica Ur of data element U and the data element U. Since array element E1 includes basis vectors capable of effectively replicating data elements of only data set C1, replication of any data not of data set C1, will not be successful. Therefore, the comparison between the replicated data element and the original data element yield either acceptable or unacceptable results. If the results are acceptable, then the unknown data element is categorized as being of the data set C1. If the results are not acceptable, then the unknown data element is not of the data set C, and is not categorized by array element E1.

[0029] Referring now to FIG. 6, operational steps are depicted as performed by a computer or network of computers in accordance with the present invention. Specifically, at step S1, pre-sorted and categorized data sets C1 through CN are inputted into the computer. However, at step S2, data set C1, is sent only to a first array element E1 (not shown in FIG. 6), at step S8 data set C2 is sent only to a second array element E2 (not shown in FIG. 6), and so on until a final iteration begins at step S14 where a final data set CN is sent to a final array element EN(not shown in FIG. 6).

[0030] It should be appreciated that the counter N used throughout the description of the invention, for instance in and EN, is any positive integer ranging from 1 to about 106. N could be greater than 106, depending upon computer capability and complexity, and the requirements of the data classification.

[0031] At step S3 in FIG. 6, the first array element E1 generates a first basis vector set AFRNC1, based upon the data elements of categorized data set C1. The first basis vector set AFRNC1, consists of basis vectors generated by the computer's mathematical analysis of each of the data elements of categorized data set C1. In other words, the first basis vector set AFRNC1, defines the replication vocabulary of first array element E1. At step S4, data replicas C1r, of each data element of the data set C1, are constructed using the first basis vector set AFRNC1. At step S5 a comparison is made between each of the replicas C1, and corresponding data elements of the data set C1, thereby producing an error value for each comparison. At step S6 a determination is made whether or not the error values are within an acceptable range. If the error values are within an acceptable range, then the vocabulary of basis vectors in the first basis vector set AFRNC1, is acceptable and the first array element E1 is trained. If the error values are not within an acceptable range, then the process repeats from step S3 until an appropriate basis vector set can be determined.

[0032] Similarly, at step S9, the second array element E1 generates a second basis vector set AFRNC2 based upon the data elements of categorized data set C2. The second basis vector set AFRNC2 consists of basis vectors generated by the computer's mathematical analysis of each of the data elements of categorized data set C2. At step S10, data replicas C2r, of each data element of the data set C2 are constructed using the second basis vector set AFRNC2. At step S11 a comparison is made between each of the replicas C2, and corresponding data elements of the data set C2 thereby producing an error value for each comparison. At step S12 a determination is made whether or not the error values are within an acceptable range. If the error values are within an acceptable range, then the second basis vector set AFRNC2 is acceptable and the second array element E2 is trained. If the error values are not within an acceptable range, then the process repeats from step S9 until an appropriate basis vector set can be generated.

[0033] The training process continues through the Nth iteration. Specifically, the Nth iterations is depicted beginning at step S15, where the Nth array element EN generates an Nth basis vector set AFRNCN based upon the data elements of categorized data set CN. The Nth basis vector set AFRNCN consists of basis vectors generated by the computer's mathematical analysis of each of the data elements of categorized data set CN. At step S16, data replicas CNr of each data element of the data set CN are constructed using the Nth basis vector set AFRNCN. At step S17 a comparison is made between each of the replicas CN, and corresponding data elements of the data set CN thereby producing an error value for each comparison. At step S 18 a determination is made whether or not the error values are within an acceptable range. If the error values are within an acceptable range, then the Nth basis vector set AFRNCN is acceptable and the Nth array element EN is trained. If the error values are not within an acceptable range, then the process repeats from step S15 until an appropriate basis vector set can be generated.

[0034] Once the AFRNs are trained, then the system is ready to identify new data by categorizing it in the previously identified categories.

[0035] Specifically, as shown in FIG. 7, an unknown data element U is inputted into the computer or computer network for categorization. N number of data paths are depicted in FIG. 7 with the data U being inputted into each of the data paths for analysis by AFRNC1, AFRNC2, and so on until the Nth iteration of analysis by AFRNCN. the following description goes through operations in one data path at a time. However, it should be understood that computers operate at such an accelerated rate, that all N data paths proceed almost simultaneously.

[0036] At step S50, data replica Ur1, is constructed using the first basis vector set AFRNC1, in an attempt to reproduce the data U. At step S51 a comparison is made between the replica Ur1 and the data U thereby producing an error value. At step S52 a determination is made, specifically, if the error value is within an acceptable range, then operation moves to step S53. If the error value is not acceptable, operation moves to step S54. At step S53, a determination is made whether or not the data U has been categorized in another data path as falling within any one of data sets C2 through CN. If the determination at step S53 is no, then the data U is identified and categorized as falling within data set C1, and the operation with respect to classification of the data U is done. If the determination at step S53 is yes, specifically, one of the other AFRN array elements identified data U as being of a data set other than C1, then data U is tagged for human analysis. In other words, the system has a problem identifying data U. Returning to step S52, if the error is not acceptable, then the data U is not of the data set C1, and another determination must be made at step S54. At step S54 a determination is made whether the data U has been identified as being in one and only one of the other data categories, such as data C2 through CN. It should be understood that at step S54 identification of the data U must only be confirmable in one and only one data category. If the data U is categorized in more than one data category, for instance more thanone of data sets C2 through CN, then data U is not properly categorized by the system and must be tagged for human analysis.

[0037] Similarly, at step S60, data replica Ur2 is constructed using the first basis vector set AFRNC2 in an attempt to reproduce the data U. At step S61 a comparison is made between the replica Ur2 and the data U thereby producing an error value. At step S62 a determination, is made, specifically, if the error value is within an acceptable range, then operation moves to step S63. If the error value is not acceptable, operation moves to step S64. At step S63, a determination is made whether or not the data U has been categorized in another data path as falling within any one of data sets C1, C3 through CN. If the determination at step S63 is no, is then the data U is identified and categorized as falling within data set C2 and the operation is done. If the determination at step S63 is yes, specifically, one of the other AFRN array elements identified data U as being of a data set other than C2, then data U is tagged for human analysis. Returning to step S62, if the error is not acceptable, another determination is made at step S64. At step S64 a determination is made whether the data U has been identified as being in one and only one of the other data categories, such as data C1, or C3 through CN. It should be understood that at step S64 identification of the data U must only be confirmable in one and only one data category. If the data U is categorized in more than one category, for instance more that one of data sets C1, C3 through CN, then data U is not properly categorized by the system and must be tagged for human analysis.

[0038] The operations are similar for all of the data paths. For the Nth iteration, at step S70, data replica UrN, is constructed using the first basis vector set AFRNCN in an attempt to reproduce the data U. At step S71 a comparison is made between the replica UrN and the data U thereby producing an error value. At step S72 a determination is made. If the error value is within an acceptable range, then operation moves to step S73. If the error value is not acceptable, operation moves to step S74. At step S73, a determination is made whether or not the data U has been categorized in another data path as falling within any one of data sets C1, through CN-1. If the determination at step S73 is no, then the data U is identified and categorized as falling within data set CN and the operation is done. If the determination at step S73 is yes, specifically, one of the other AFRN array elements identified data U as being of a data set other than CN, then data U is tagged for human analysis. Returning to step S72, if the error is not acceptable, another determination is made at step S74. At step S74 a determination is made whether the data U has been identified as being in one and only one of the other data categories, such as data C1 through CN-1. It should be understood that at step S74 identification of the data U must only be confirmable in one and only one data category. If the data U is categorized in more than one category, for instance more that one of data sets C1, through CN-1, then data U is not properly categorized by the system and must be tagged for human analysis. It should be understood that the data sets C1, through CN used in the description above of the invention; may be similar types of data or may be discrete groups of data with little or no relationship to one another. However, to demonstrate the power of the present invention, the inventor has conducted several experiments where the data sets C1, through CN were all of a similar nature.

EXAMPLE ONE: CHIRPS



[0039] In an example prepared by the inventor, an array of AFRNs was constructed using chirps. Chirps are complex exponential waves that vary with time. Chirps are important in a variety of applications such as radar, laser, sonar and devices and in telecommunication links. In the training phase, three separate AFRN array elements were prepared, AFRN I, AFRN II and AFRN III, as is described in greater detail below.

[0040] Chirps were used in this experiment because, as is observed by visual inspection of FIGS. 8A, 8B and 8C, they appear to be very similar. Human intervention is not likely to detect the differences between such three similar data sets. Chirps are also complex signals that can be generated with mathematical precision. The chirps are also 'data-dense' because they are mathematically complex. Compared to real world applications of the present invention, the following example using Chirps is significantly more complex.

[0041] Each chirp in FIGS. 8A, 8B and 8C is made up of 500-element vectors, each individual vector defined by a point located at a point that can be identified visually by a pair of X and Y coordinates. Reference Set I shown in FIG. 8A, has 10 chirps (or waves) altogether, the chirps in reference set I being distinguished from one another by a very slight off-set in frequency. Reference Set II shown in FIG. 8B also has 10 chirps altogether, the chirps distinguished from one another by a very slight offset in frequency. Reference Set III shown in FIG. 8C has 10 chirps as well, the chirps distinguished from one another by a very slight offset in frequency.

[0042] Three sets of known chirp sets used to demonstrate the present invention are depicted in FIGS. 8A, 8B and 8C and are labeled Input Reference Set 1, Input Reference Set II and Input Reference Set III, respectively. Visually, the three chirps in FIGS. 8A, 8B and 8C look similar, but vary in frequency over time. Specifically, each chirp has its own unique frequency profile as can be seen by observing the location of the curves at data points 200 and 300 along the X-axis in FIGS. 8A, 8B and 8C.

[0043] AFRN I, AFRN 11 and AFRN III were created, one AFRN for each of the three sets of chirps. Each AFRN trained to replicate one of the three reference sets. During the training phase, AFRN I was given the ten chirp vectors in Input Reference Set I (FIG. 8A), AFRN II was given the ten chirp vectors in Input Reference Set II (FIG. 8B) and AFRN III was given the ten chirp vectors in Input Reference Set III (FIG. 8C).

[0044] During the training phase, a unique set of basis vectors was derived for each AFRN. Specifically, AFRN I includes the basis vectors shown in FIGS. 18A, 18B 18C, 18D and 18E, and further depicted together in one graph in FIG. 18F. AFRN II includes the basis vectors shown in FIGS. 19A, 19B, 19C, 19D and 19E, and further depicted together in one graph in FIG. 19F. AFRN III includes the basis vectors shown in FIGS. 20A, 20B, 20C, 20D and 20E, and further depicted together in one graph in FIG. 20F.

[0045] As was described above with respect to the flowchart in FIG. 6, AFRN I was trained by inputting reference set I (FIG. 8A and 9A) in order to derive the set of basis vectors depicted in FIGS. 18A, 18B, 18C, 18D and 18E. Thereafter, AFRN I attempted to replicate reference set I, as shown in FIG. 9B. An output error (FIG. 9C) was generated by comparing reference set I and the replicas generated by AFRN I. The error shown in FIG. 9C is flat, indicating that the error was insignificant, and therefore the chirps in FIG. 9A were successfully replicated.

[0046] FIGS. 10A, 10B and 10C show the results of a test where an unknown chirp, Novel Class I was inputted into AFRN 1. FIG. 10Ashows the unknown data, Novel Class I. FIG. 10B shows the replicated data outputted from AFRN 1. FIG. 10C shows the error resulting from a comparison of the output in FIG. 10B and the input shown in FIG. 10A. The error is negligible. Indeed, the error must be amplified to be visible to the naked eye. Therefore, pending further tests, the unknown data depicted in FIG. 10A should clearly be classified in Reference Set I.

[0047] FIGS. 10D, 10E, and 10F show a similar test using AFRN II. Novel Class I vector is again shown in FIG. 10D. FIG. 10E shows an attempted replica ofNovel Class I vector by AFRN II. The error generated by comparing Novel Class I with the attempted replica is shown in FIG. 10F. Clearly, Novel Class I vector is not accurately replicated by AFRN II and cannot be classified by AFRN II.

[0048] FIGS. 10G, 10H and 10I show yet another test using AFRN III. Novel Class I vector is again shown in FIG. 10G. FIG. 10H shows an attempted replica of Novel Class I vector by AFRN III. The error generated by comparing Novel Class I with the attempted replica is shown in FIG. 10I. Clearly, Novel Class I vector is not accurately replicated by AFRN III and cannot be classified by AFRN III.

[0049] To further test AFRN 1, data from Reference Set II shown in FIG. 11A was inputted. AFRN I generated replicas depicted in FIG. 11B. Visual comparisons between the data in FIG. 11A and the replica depicted in FIG. 11B indicate that the two sets of data are not very similar. However, the error output depicted in FIG. 11C shows clearly that the Reference Set II data should not be classified with Reference Set I. Even with no amplification, FIG. 11 C clearly shows gross errors. Therefore, the data curve in FIG. 11A cannot be categorized as being part of a data category corresponding to Reference Set I.

[0050] Similarly, data from Reference Set III shown in FIG. 11D was inputted. AFRN I generated replicas depicted in FIG. 11E. Visual comparisons between the data in FIG. 11D and the replica depicted in FIG. 11E indicate that the two sets of data are not very similar. Further, the error output depicted in FIG. 11F shows clearly that the Reference Set III data should not be classified with Reference Set I. Even with no amplification, FIG. 11F clearly shows gross errors. Therefore, the data curve in FIG. 11D cannot be categorized as being part of a data category corresponding to Reference Set I.

[0051] AFRN II was trained by inputting reference set II (FIG. 8B and 12A) in order to derive the set of basis vectors depicted in FIGS. 19A, 19B, 19C, 19D and 19E. Thereafter, AFRN II attempted to replicate reference set II, as shown in FIG. 12B. An output error (FIG. 12C) was generated by comparing reference set II and the replicas generated by AFRN II. The error shown in FIG. 12C is flat, indicating that the error was insignificant, and therefore the chirps in FIG. 12A were successfully replicated.

[0052] FIGS. 13A, 13B and 13C show the results of a test where an unknown chirp, Novel Class II vector, was inputted into AFRN 11. FIG. 13A shows the unknown Novel Class II vector. FIG. 13B shows the replicated data outputted from AFRN II. FIG. 13C shows the error resulting from a comparison of the output in FIG. 13B and the input shown in FIG. 13A. The error is negligible. Therefore, pending further tests, the unknown data depicted in FIG. 13A should clearly be classified in Reference Set II.

[0053] FIGS. 13D, 13E and 13F show a similar test using AFRN I. Novel Class II vector is again shown in FIG. 13D. FIG. 13E shows an attempted replica of Novel Class II vector generated by AFRN I. The error generated by comparing Novel Class II with the attempted replica is shown in FIG. 13F. Clearly, Novel Class II vector is not accurately replicated by AFRN I and cannot be classified by AFRN I.

[0054] FIGS. 13G, 13H and 13I show yet another test using AFRN III. Novel Class II vector is again shown in FIG. 13G. FIG. 13H shows an attempted replica of Novel Class II vector by AFRN III. The error generated by comparing Novel Class II with the attempted replica is shown in FIG. 131. Clearly, Novel Class II vector is not accurately replicated by AFRN III and cannot be classified by AFRN III.

[0055] To further test AFRN II, data from Reference Set I shown again in FIG. 14A was inputted into AFRN II. AFRN II attempted to generate replicas depicted in FIG. 14B. Visual comparisons between the data in FIG. 14A and the replica depicted in FIG. 14B indicate that the two sets of data are not the same. The error output depicted in FIG. 14C clearly confirms that the Reference Set I data should not be classified with Reference Set II. FIG. 14C clearly shows gross errors in the replication process. Therefore, the data curve in FIG. 14A cannot be categorized as being part of a data category corresponding to Reference Set II.

[0056] Similarly, data from Reference Set III shown again in FIG. 14D was inputted. AFRN II attempted to generate replicas, as is shown in FIG. 14E. Visual comparisons between the data in FIG. 14D and the replicas depicted in FIG. 14E indicate that the two sets of data are not very similar. Further, the error output depicted in FIG. 14F shows clearly that the Reference Set III data should not be classified with Reference Set II. Even with no amplification, FIG. 14F clearly shows gross errors. Therefore, the data curve in FIG. 14D cannot be categorized as being part of a data category corresponding to Reference Set II.

[0057] AFRN III was trained by inputting reference set III (FIG. 8C and 15A) in order to derive the set of basis vectors depicted in FIGS. 20A, 20B, 20C, 20D and 20E. Thereafter, AFRN III attempted to replicate reference set III, as shown in FIG. 15B. An output error (FIG. 15C) was generated by comparing reference set III and the replicas generated by AFRN III. The error shown in FIG. 15C is flat, indicating that the error was insignificant, and therefore the chirps in FIG. 15A were successfully replicated.

[0058] FIGS. 16A, 16B and 16C show the results of a test where an unknown chirp, Novel Class III vector, was inputted into AFRN III. FIG. 16A shows the unknown data. FIG. 16B shows the replicated data outputted from AFRN III. FIG. 16C shows the error resulting from a comparison of the output in FIG. 16B and the input shown in FIG. 16A. The error is negligible. Therefore, pending further tests, the unknown data depicted in FIG. 16A should clearly be classified in Reference Set III.

[0059] FIGS. 16D, 16E and 16F show a similar test using AFRN I. Novel Class III vector is again shown in FIG. 163D. FIG. 16E shows an attempted replica of Novel Class III vector generated by AFRN 1. The error generated by comparing Novel Class III with the attempted replica is shown in FIG. 16F. Clearly, Novel Class III vector is not accurately replicated by AFRN I and cannot be classified by AFRN I.

[0060] FIGS. 16G, 16H and 161 show yet another test using AFRN II. Novel Class III vector is again shown in FIG. 16G. FIG. 16H shows an attempted replica of Novel Class III vector by AFRN II. The error generated by comparing Novel Class III with the attempted replica is shown in FIG. 16I. Clearly, Novel Class III vector is not accurately replicated by AFRN II and cannot be classified by AFRN II.

[0061] To further test AFRN III, data from Reference Set I, shown again in FIG. 17A, was inputted into AFRN III. AFRN III attempted to generate replicas depicted in FIG. 17B. Visual comparisons between the data in FIG. 17A and the replica depicted in FIG. 17B indicate that the two sets of data are not the same. The error output depicted in FIG. 17C clearly confirms that the Reference Set I data should not be classified with Reference Set III. FIG. 17C clearly shows gross errors in the replication process. Therefore, the data curve in FIG. 17A cannot be categorized as being part of a data category corresponding to Reference Set III.

[0062] Similarly, data from Reference Set II shown again in FIG. 17D was inputted. AFRN III attempted to generate replicas, as is shown in FIG. 17E. Visual comparisons between the data in FIG. 17D and the replicas depicted in FIG. 17E indicate that the two sets of data are not very similar. Further, the error output depicted in FIG. 17F shows clearly that the Reference Set II data should not be classified with Reference Set III. FIG. 17F clearly shows gross errors. Therefore, the data curve in FIG. 17D cannot be categorized as being part of a data category corresponding to Reference Set III.

[0063] The above described example demonstrates that once each AFRN element of an array of AFRNs has been trained to replicate a specific category or class of data, each AFRN element will only accurately replicate data that belongs with that class of data. Further each AFRN element fails to accurately replicate data from outside its class of data. Arrays of AFRNs (a group of array elements) can be trained to replicate and categorize groups of data without limitations.

[0064] In the chirp example described above, a 100% accuracy rate was achieved in all tests.

[0065] In general chirps are mathematically complex and it is difficult to recognize the difference between two chirps by most data manipulating systems. In most real world data analysis systems, the data inputs are mathematically much less complex and easier to recognize than chirps. Therefore, it should be clear that AFRN arrays are even more successful classifying data having simpler mathematical representations. In other words, AFRN arrays classify difficult data easily. Simpler types of data are more easily classified using AFRN arrays.

EXAMPLE TWO: GENOTYPING



[0066] FIG. 21 is a graph showing an output from a nucleic acid sequencing device showing representation of indications of four dyes labeling nucleic acids. Such graphs are produced by scanners focused on electrophoresis gels with nucleotide fragments. Such scanners are able to identify nucleic acids by the spires in the graphs, each color representing a different nucleic acid. However, consistency of identification of the nucleic acids and/or length of the fragments is uneven even with the most recent equipments. Using an AFRN array, it is possible to improve identification of the output significantly.

[0067] In the present example of the present invention, AFRNs were trained to assist in genotyping where it is necessary to determine the characteristics of nucleic acid fragments. In this example of the invention, three classes of vectors were used to train an AFRN array having three elements. The three classes I, II and III of reference vectors are depicted in FIGS. 22A, 22B and 22C, respectively. Each class I, II and III of reference vectors corresponds to known characteristics of nucleotide sequences. Specifically, class I reference vectors I in FIG. 22B were predetermined to be a first nucleotide sequence characteristic, class II reference vectors in FIG. 22B were predetermined to be a second nucleotide sequence characteristic, and class III reference vectors in FIG. 22C were predetermined to be a third nucleotide sequence characteristic.

[0068] AFRN I was trained to replicate class I reference vectors, AFRN II was trained to replicate class II reference vectors, and AFRN III was trained to replicate class III reference vectors. Basis vectors depicted in FIG. 23A were derived for AFRN I, basis vectors depicted in FIG. 23B were derived for AFRN II, and basis vectors depicted in FIG. 23C were derived for AFRN III.

[0069] During the training phase, the class I reference vectors depicted again in FIG. 24A, were inputted and AFRN I replicated each vector, as shown in FIG. 24B. An error was determined by comparing the original data with the replicated data, as shown in FIG. 24C. Since all the fines in FIG. 24C are flat, it is clear that AFRN I accurately replicated the vectors and training is complete.

[0070] To test the reliability of AFRN I, a new group of vectors, Test Class II Vectors depicted in FIG. 25A, were initially identified by other means as falling within class II reference vectors. Test Class II Vectors were inputted and AFRN I replicated each inputted vector. The replications were compared to the original vectors and errors were determined, as shown in FIG. 25C. The results in FIG. 25C indicate one vector of Test Class II Vectors falls within the class I reference vectors, as indicated by the single flat fine at the bottom of FIG. 25C. The_remaining vectors clearly do not belong with class I reference vectors. As is seen by comparing FIGS. 25A and 25B, one of the vectors is clearly replicated in FIG. 25B and does indeed appear have the same general amplitude peaks as the class I reference vectors in FIG. 22A. Therefore, the classification of this vector by the previously used method of classification is in doubt. Visual inspection of the accurately reproduced vector confirm that this vector belongs with class I reference vectors.

[0071] Another test of the testing of AFRN I is shown in FIGS. 26A, 26B and 26C. The class III reference vectors are shown again in FIG. 26A. FIG. 26B shows replications of the class III reference vectors generated by AFRN 1. Comparison of the original vectors with the replicas yields the error graph in FIG. 26C. Clearly, none of the vectors replicated by AFRN I shown in FIG. 26B belong with class I reference vectors.

[0072] Similar tests were conducted using AFRNs II and III with identical results. The AFRNs of the present. invention only replicated data that belonged in the class the AFRN was trained to replicate.

[0073] From the above examples of applications of the present invention, it is clear that AFRN arrays may be trained and then used to replicate specific types of data. Data that is accurately replicated is classified within the class used to train the successful AFRN.

[0074] In another embodiment of the present invention, the AFRN array may automatically expand its capabilities by determining that a new AFRN array element needs to be created. Specifically, FIG. 27 is a flowchart showing steps for automatically refining and expanding the data analysis system of the present invention.

[0075] In FIG. 27, at step S80, data inputted is used to train an AFRN array, as described above with respect to FIG. 6. In step S81, the trained AFRN array replicates and thereby classifies data, as is described above with respect to FIG. 7. Repeated iterations of the steps in FIG. 7 generate significant amounts of data classification and error. After a predetermined amount of time or after a predetermined threshold of data has been replicated and classified, a determination is made at step S82 whether or not any array elements in the trained AFRN array has identified and classified an excessive amount of data. If yes, then another determination is made at step S83 whether or not a current AFRN needs a reduction in error threshold and a new AFRN should be created. If yes, then a new error threshold is generated at step S84. Next, in step S85, all of the data classified by that particular AFRN is reevaluated (replicated) with a new smaller threshold of error to determine the amount of data that would remain classified by the AFRN and determine the amount of data that is to be separated and classified in a new classification of data. In step S86, the new classification category is defined and a new AFRN created. Thereafter, the AFRN array returns to a mode where classification of data continues.

[0076] It will be understood from the above description, that the present invention is applicable to a variety of data classification applications. For instance, the AFRN arrays of the present invention may be used in gene sequencing systems, speech recognition systems, optical recognition systems, informatics, radar systems, sonar systems, signal analysis and many other digital signal processing applications.


Claims

1. A method for classifying data, comprising the steps of:

a) training a computer (MP1, MP2...MPN) to replicate data sets (C1, C2...CN), said training step including defining a plurality of replicators (E1, E2...EN), each one replicator (E1, E2...EN) being trained to replicate one data set of said data sets (C1, C2...CN);

b) inputting new data (U) into the computer (MP1, MP2...MPN);

c) replicating (S50; S60; S70) the new data (U) by each of the replicators (E1, E2...EN)

d) comparing (S51) replicated data (Ur1) from a first of the replicators (E1) with the new data (U) to determine accuracy of the replicated data (Ur1) if replicated by the first of the replicators (E1) within a predetermined error threshold;

e) repeating (S61, S71) step d) for each of the replicators (E2...EN)

f) determining (S53, S63, S73) if one of the comparisons in steps d) and e) yielded confirmation of accurate replication for one or more replicator (E1, E2...EN), and

g) classifying the new data (U) in response to the determination of accurate replication by only one replicator (E1, E2...EN) in step f).


 
2. A method for classifying data as set forth in claim 1, wherein the method is performed by a plurality of computers (MP1, MP2...MPN).
 
3. A method for classifying data as set forth in claim 2, wherein the plurality of computers (MP1, MP2...MPN) are connected via a network.
 
4. A method for classifying data as set forth in claim 1, wherein the method is performed by a plurality of microprocessors (MP1, MP2...MPN) working in close communication with one another.
 
5. A method for classifying data as set forth in claim 1, further comprising the step (S83) of determining the need for one or more new replicators in response to a predetermined threshold amount of data classification by any one replicator (E1, E2...EN).
 
6. A method for classifying data as set forth in claim 1, wherein said step a) of training the computer (MP1, MP2...MPN) comprises the steps of:

inputting (S1) several previously identified data sets (C1, C2...CN) into a computer (MP1, MP2...MPN);

creating (S3, S9, S 15) within the computer (MP1, MP2...MPN) a plurality of basis vector sets (AFRNC1, AFRNC2...AFRNCN), one basis vector set for each of said replicators (E1, E2...EN), each of said basis vector sets (AFRNC1, AFRNC2...AFRNCN) comprising a plurality of basis vectors, each of said basis vector sets (AFRNC1, AFRNC2...AFRNCN) used to replicate the corresponding identified data set from said identified data sets (C1, C2...CN);

wherein creating a basis vector set (AFRNC1, AFRNC2...AFRNCN) for each of said replicators (E1, E2...EN) comprises:

constructing (S4, S10, S16) a comparison data set (C1r, C2r...CNr) using the corresponding basis vector set (AFRNC1, AFRNC2...AFRNCN);

comparing the previously identified data set (C1, C2...CN) with the comparison data set (C1r, C2r...CNr);

computing (S5, S11, S17) an error from said comparison;

determining (S6, S12, S18) acceptability of the error; and

repeating said creating of the basis vector set in response to the error being unacceptable.


 
7. A method as set forth in claim 6 wherein in said creating step, the basis vector sets (AFRNC1, AFRNC2...AFRNC3) are generated from any combination of the following: chirplets, walvelets, Fourier based functions, Fractal basis functions, radial basis functions, basis functions generated using multi-layer perceptrons, and other computer generated basis functions.
 
8. A method for classifying data sets (U), comprising the steps of:

a) inputting previously identified data sets (C1, C2...CN) into a computer (MP1, MP2...MPN);

b) creating within the computer (MP1, MP2...MPN) a plurality of basis vector sets (AFRNC1, AFRNC2...AFRNCN) in one to one correspondence with said previously identified data sets (C1, C2...CN), each defining a replicator (E1, E2...EN);

c) inputting a new data set (U) into the computer (MP1, MP2...MPN);

d) for each replicator (E1, E2...EN), generating (S50, S60, S70) a replica (Ur) of the new data set (U) using the basis vector sets (AFRNC1, AFRNC2...AFRNCN);

e) comparing (S51) the replica (Ur1) generated by the first replicator (E1) of the plurality of replicators (E1, E2...EN) with the new data set (U);

f) determining acceptable replication (S52) of the new data set (Ur1) generated by the first replicator (E1) using a predetermined error threshold;

g) repeating (S61, S71, S62, S72) steps e) and f) for each of the replicators (E2...EN) and the replica (Ur2...Urn) generated by the corresponding replicator (E2...EN);

h) determining (S53, S63, S73) if more than one replicator acceptably replicated the new data set (U) in repeated steps f);

i) tagging the new data set (U) for human analysis in response to a determination in step h) that more than one replicator (E1, E2...EN) acceptably replicated the new data set (U);

j) classifying (S53, S63, S73) the new data (U) in response to determination in step h) that only one replicator (E1, E2...EN) acceptably replicated the new data set (U).


 
9. A method for classifying data as set forth in claim 8, further comprising the step (S82, S83) of determining the need for new replicators (E1, E2...EN) in response to a predetermined threshold amount of data classification.
 
10. A method for classifying data as set forth in claim 8, further comprising the step of tagging the new data for human analysis in response to a determination in step h) that no replicator (E1, E2...EN) acceptably replicating the new data (U).
 
11. The method for classifying data as set forth in claim 8, further comprising the step of determining the need to automatically create a new replicator in response to any one of the replicators (E1, E2...EN) having identified and classified an excessive amount of data after a predetermined amount of time or after a predetermined threshold of data has been replicated and classified.
 
12. A computer system for replicating and classifying data, comprising:

a microprocessor (MP1, MP2...MPN) adapted for:

training to replicate identified sets of data (C1, C2...CN) thereby creating a plurality of replicators (E1, E2...EN), each replicator (E1, E2...EN) trained to replicate one set of the data of the identified sets of data (C1, C2...CN),

receiving unknown data (U),

having each of the replicators (E1, E2...EN) attempt to replicate the unknown data (U), for each replicator (E1, E2...EN) comparing the unknown data (U) with the replicated data (Ur);

determining whether each replicator (E1, E2...EN) acceptably reproduced the unknown data (U) within a predetermined threshold error; and

in response to a determination of acceptable replication by only one replicator (E1, E2...EN), classifying the unknown data (U).


 
13. A computer system for replicating and classifying data as set forth in claim 12, wherein said microprocessor (MP1, MP2...MPN) is further adapted for determining if the threshold error for any one replicator (E1, E2...EN) needs revision in response to significant amounts of data classification by that one replicator (E1, E2...EN).
 
14. A computer system for replication and classifying data as set forth in claim 12, wherein said microprocessor (MP1, MP2...MPN) is further adapted for determining whether new replicators (AFRNC1, AFRNC2...AFRNCN) and corresponding basis vector sets (C1, C2...CN) are required for classifying data.
 
15. A computer system for replicating and classifying data as set forth in claim 12, wherein said microprocessor comprises a plurality of microprocessors (MP1, MP2...MPN) linked to one another.
 
16. A computer system for replicating and classifying data as set forth in claim 15, wherein said microprocessors (MP1, MP2...MPN) are linked to one another via a negotiator.
 
17. A computer system for replicating and classifying data as set forth in claim 15, wherein said microprocessors (MP1, MP2...MPN) are linked to one another for symmetric multiprocessing.
 
18. A computer system for replicating and classifying data as set forth in claim 15, wherein said microprocessors (MP1, MP2...MPN) are linked to one another for asymmetric multiprocessing.
 
19. A computer system for replicating and classifying data as set forth in claim 15, wherein said microprocessors (MP1, MP2...MPN) are linked to one another at chip level.
 
20. A computer system for replicating and classifying data as set forth in claim 15, wherein said microprocessors (MP1, MP2...MPN) are linked to one another in a network.
 
21. A computer system for replicating and classifying data as set forth in claim 15, wherein said microprocessors (MP1, MP2...MPN) are linked to one another in clusters.
 
22. A computer system for replicating and classifying data as set forth in claim 15, wherein said computer system is configured for genotyping.
 


Ansprüche

1. Ein Verfahren zum Klassifizieren von Daten, welches die Schritte aufweist:

a) Trainieren eines Computers (MP1, MP2...MPN) auf das Replizieren von Datensätzen (C1, C2...CN), wobei der Schritt des Trainierens das Definieren einer Vielzahl von Replikatoren (E1, E2...EN) umfasst, wobei jeder Replikator (E1, E2...EN) auf das Replizieren von einem der Datensätzen (C1, C2...CN) trainiert wird;

b) Eingeben neuer Daten (U) in den Computer (MP1, MP2...MPN) ;

c) Replizieren (S50; S60; S70) der neuen Daten (U) durch jeden der Replikatoren (E1, E2...EN);

d) Vergleichen (S51) replizierter Daten (Ur1) von einem ersten der Replikatoren (E1) mit den neuen Daten (U), um die Richtigkeit der replizierten Daten (Ur1) festzustellen, falls diese durch den ersten der Replikatoren (E1) innerhalb einer vorgegebenen Fehlergrenze repliziert worden sind;

e) Wiederholen (S61, S71) von Schritt d) für jeden der Replikatoren (E2...EN);

f) Feststellen (S53, S63, S73), ob einer der Vergleiche in Schritten d) und e) eine Bestätigung einer genauen Replikation für einen oder mehrere Replikatoren (E1, E2...EN) ergeben hat;

g) Klassifizieren der neuen Daten (U) anhand der Feststellung einer genauen Replikation durch nur einen Replikator (E1, E2...EN) in Schritt f).


 
2. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 1, wobei das Verfahren durch eine Vielzahl von Computern (MP1, MP2...MPN) ausgeführt wird.
 
3. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 2, wobei die Vielzahl von Computern (MP1, MP2...MPN) über ein Netzwerk verbunden sind.
 
4. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 1, wobei das Verfahren durch eine Vielzahl von Mikroprozessoren ausgeführt wird, die in enger Kommunikation miteinander arbeiten.
 
5. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 1, welches weiter den Schritt (S83) aufweist, den Bedarf an einem oder mehreren neuen Replikatoren anhand eines vorgegebenen grenzwertigen Umfangs der Datenklassifizierung durch irgendeinen Replikator (E1, E2...EN) zu ermitteln.
 
6. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 1, wobei der Schritt a) des Trainierens des Computers (MP1, MP2...MPN) die Schritte umfasst:

Eingeben (S1) von mehreren früher identifizierten Datensätzen (C1, C2...CN) in einen Computer (MP1, MP2...MPN) ;

Erzeugen (S3, S9, S15) einer Vielzahl von Basisvektorsätzen (AFRNC1, AFRNC2...AFRNCN) auf dem Computer (MP1, MP2...MPN), einen Basisvektorsatz für jeden der Replikatoren (E1, E2...EN), wobei jeder der Basisvektorsätze (AFRNC1, AFRNC2...AFRNCN) eine Vielzahl von Basisvektoren umfasst, wobei jeder der Basisvektorsätze (AFRNC1, AFRNC2...AFRNCN) zum Replizieren des zugehörigen identifizierten Datensatzes von den identifizierten Datensätzen (C1, C2...CN) verwendet wird;

wobei das Erzeugen eines Basisvektorsatzes (AFRNC1, AFRNC2...AFRNCN) für jeden der Replikatoren (E1, E2...EN) umfasst:

Erstellen (S4, S10, S16) eines Vergleichsdatensatzes (C1r, C2r...CNr) unter Verwendung des zugehörigen Basisvektorsatzes (AFRNC1, AFRNC2...AFRNCN);

Vergleichen des früher identifizierten Datensatzes (C1, C2...CN) mit dem Vergleichsdatensatz (C1r, C2r...CNr);

Bestimmen (S5, S11, S17) eines Fehlers aus diesem Vergleich;

Ermitteln (S6, S12, S18) der Hinnehmbarkeit des Fehlers; und

Wiederholen des Erzeugens des Basisvektorsatzes, falls der Fehler nicht akzeptabel ist.


 
7. Ein Verfahren gemäß Anspruch 6, wobei die Basisvektorsätze (AFRNC1, AFRNC2...AFRNCN) bei dem Schritt des Erzeugens aus irgendeiner Kombination der folgenden erzeugt werden: Chirplets, Wavelets, Fourier-basierte Funktionen, fraktale Basisfunktionen, radiale Basisfunktionen, Basisfunktionen, die unter Verwendung von mehrschichtiger Perzeptrons erzeugt worden sind, und andere computergenerierte Basisfunktionen.
 
8. Ein Verfahren zum Klassifizieren von Datensätzen (U), welches die Schritte umfasst:

a) Eingeben von früher identifizierten Datensätzen (C1, C2...CN) in einen Computer (MP1, MP2 ... MPN) ;

b) Erzeugen einer Vielzahl von Basisvektorsätzen (AFRNC1, AFRNC2...AFRNCN) auf dem Computer (MP1, MP2...MPN) in einszu-eins Relation mit den früher identifizierten Datensätzen (C1, C2...CN), wobei jeder einen Replikator (E1, E2...EN) definiert;

c) Eingeben eines neuen Datensatzes (U) in den Computer (MP1, MP2...MPN);

d) Erzeugen (S50, S60, S70) einer Replika (Ur) des neuen Datensatzes (U) für jeden Replikator (E1, E2...EN) unter Verwendung der Basisvektorsätze (AFRNC1, AFRNC2...AFRNCN);

e) Vergleichen (S51) der durch den ersten Replikator (E1) der Vielzahl von Replikatoren (E1, E2...EN) erzeugten Replika (Ur1) mit dem neuen Datensatz (U);

f) Überprüfen einer akzeptablen Replikation (S52) des durch den ersten Replikator (E1) erzeugten neuen Datensatzes (Ur1) unter Verwendung einer vorgegebenen Fehlergrenze;

g) Wiederholen (S61, S71, S62, S72) von Schritten e) und f) für jeden der Replikatoren (E2...EN) und die durch den zugehörigen Replikator (E2...EN) erzeugte Replika (Ur2...Urn);

h) Feststellen (S53, S63, S73), ob mehr als ein Replikator den neuen Datensatz (U) in wiederholten Schritten f) in akzeptabler Weise repliziert hat;

i) auf eine Feststellung in Schritt h) hin, dass mehr als ein Replikator (E1, E2...EN) den neuen Datensatz (U) in akzeptabler Weise repliziert hat, Markieren des neuen Datensatzes (U) für die Analyse durch Menschen;

j) Klassifizieren (S53, S63, S73) der neuen Daten (U) anhand der Feststellung in Schritt h), dass nur ein Replikator (E1, E2...EN) den neuen Datensatz (U) in akzeptabler Weise repliziert hat.


 
9. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 8, welches weiter den Schritt (S82, S83) aufweist, den Bedarf an neuen Replikatoren (E1, E2...EN) in Abhängigkeit von einem vorgegebenen grenzwertigen Umfang der Datenklassifizierung zu ermitteln.
 
10. Ein Verfahren zum Klassifizieren von Daten gemäß Anspruch 8, welches weiter den Schritt aufweist, auf eine Feststellung in Schritt h) hin, dass kein Replikator (E1, E2...EN) die neuen Daten (U) in akzeptabler Weise repliziert, die neuen Daten für die Analyse durch Menschen zu markieren.
 
11. Das Verfahren zum Klassifizieren von Daten gemäß Anspruch 8, welches weiter den Schritt aufweist, den Bedarf nach automatischer Erzeugung eines neuen Replikators festzustellen, falls nach einer vorgegebenen Zeitspanne oder nachdem eine vorgegebene Grenzmenge von Daten repliziert und klassifiziert worden ist, irgendeiner der Replikatoren (E1, E2...EN) eine übergroße Menge von Daten identifiziert und klassifiziert hat.
 
12. Ein Computersystem zum Replizieren und Klassifizieren von Daten, welches aufweist:

ein Mikroprozessor (MP1, MP2...MPN), welcher ausgelegt ist:

das Replizieren von identifizierten Datensätzen (C1, C2...CN) zu trainieren, wodurch eine Vielzahl von Replikatoren (E1, E2...EN) erzeugt wird, wobei jeder Replikator (E1, E2...EN) auf das Replizieren eines Satzes von den Daten der identifizierten Datensätze (C1, C2...CN) trainiert ist,

unbekannte Daten (U) zu empfangen,

jeden der Replikatoren (E1, E2...EN) zu veranlassen, die unbekannten Daten (U) zu replizieren,

die unbekannten Daten (U) mit den replizierten Daten (Ur) für jeden Replikator (E1, E2...EN) zu vergleichen;

festzustellen, ob jeder Replikator (E1, E2...EN) die unbekannten Daten (U) innerhalb einer vorgegebenen Fehlergrenze in akzeptabler Weise reproduziert hat; und

die unbekannten Daten (U) anhand einer Feststellung einer akzeptablen Replikation durch nur einen Replikator (E1, E2...EN) zu klassifizieren.


 
13. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 12, wobei der Mikroprozessor (MP1, MP2...MPN) des weiteren ausgelegt ist festzustellen, ob die Fehlergrenze für irgendeinen Replikator (E1, E2...EN) aufgrund eines erheblichen Umfangs von Datenklassifizierung durch diesen Replikator (E1, E2...EN) einer Veränderung bedarf.
 
14. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 12, wobei der Mikroprozessor (MP1, MP2...MPN) des weiteren ausgelegt ist festzustellen, ob neue Replikatoren (AFRNC1, AFRNC2...AFRNCN) und zugehörige Basisvektorsätze (C1, C2...CN) zum Klassifizieren von Daten benötigt werden.
 
15. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 12, wobei der Mikroprozessor eine Vielzahl von miteinander verbundenen Mikroprozessoren (MP1, MP2...MPN) umfasst.
 
16. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei die Mikroprozessoren (MP1, MP2...MPN) über einen Vermittler miteinander verbunden sind.
 
17. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei die Mikroprozessoren (MP1, MP2...MPN) für symmetrischen Multiprozessorbetrieb miteinander verbunden sind.
 
18. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei die Mikroprozessoren (MP1, MP2...MPN) für asymmetrischen Multiprozessorbetrieb miteinander verbunden sind.
 
19. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei die Mikroprozessoren (MP1, MP2...MPN) auf Chipebene miteinander verbunden sind.
 
20. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei die Mikroprozessoren (MP1, MP2...MPN) in einem Netzwerk miteinander verbunden sind.
 
21. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei die Mikroprozessoren (MP1, MP2...MPN) in Clustern miteinander verbunden sind.
 
22. Ein Computersystem zum Replizieren und Klassifizieren von Daten gemäß Anspruch 15, wobei das Computersystem für die Genotypisierung ausgelegt ist.
 


Revendications

1. Procédé pour classer des données, comprenant les étapes suivantes :

a) entraîner un ordinateur (MP1, MP2 ... MPN) à répliquer des ensembles de données (C1, C2 ... CN), ladite étape d'entraînement comprenant la définition d'une pluralité de réplicateurs (E1, E2 ... EN), chacun des réplicateurs (E1, E2 ... EN) étant entraîné à répliquer un ensemble de données parmi lesdits ensembles de données (C1, C2 ... CN) ;

b) entrer de nouvelles données (U) dans l'ordinateur (MP1, MP2 ... MPN) ;

c) faire répliquer (S50 ; S60 ; S70) les nouvelles données (U) par chacun des réplicateurs (E1, E2 ... EN) ;

d) comparer (S51) les données répliquées (Ur1) d'un premier des réplicateurs (E1) avec les nouvelles données (U) pour déterminer la fidélité des données répliquées (Ur1) si elles ont été répliquées par le premier des réplicateurs (E1) sous un seuil d'erreur prédéterminé ;

e) répéter (S61, S71) l'étape d) pour chacun des réplicateurs (E2 ... EN) , ;

f) déterminer (S53, S63, S73) si l'une des comparaisons dans les étapes d) et e) apporte la confirmation d'une réplication fidèle pour un ou plusieurs réplicateurs (E1, E2 ... EN), et

g) classer les nouvelles données (U) en réponse à la détermination de la fidélité de la réplication par seulement un réplicateur (E1, E2 ... EN) dans l'étape f).


 
2. Procédé pour classer des données selon la revendication 1, dans lequel le procédé est réalisé par une pluralité d'ordinateurs (MP1, MP2 ... MPN).
 
3. Procédé pour classer des données selon la revendication 2, dans lequel la pluralité d'ordinateurs (MP1, MP2 ... MPN) est connectée par l'intermédiaire d'un réseau.
 
4. Procédé pour classer des données selon la revendication 1, dans lequel le procédé est réalisé par une pluralité de microprocesseurs (MP1, MP2 ... MPN) travaillant en relation étroite les uns avec les autres.
 
5. Procédé pour classer des données selon la revendication 1, comprenant de plus l'étape (S83) de détermination du besoin d'un ou plusieurs nouveaux réplicateurs en réponse à une quantité de seuil prédéterminée de classement de données par un réplicateur quelconque (E1, E2 ... EN).
 
6. Procédé pour classer des données selon la revendication 1, dans lequel ladite étape a) d'entraînement de l'ordinateur (MP1, MP2 ... MPN) comprend les étapes suivantes :

l'entrée (S1) de plusieurs ensembles de données antérieurement identifiés (C1, C2 ... CN) dans un ordinateur (MP1, MP2 ... MPN) ;

la création (S3, S9, S15) dans l'ordinateur (MP1, MP2 ... MPN) d'une pluralité d'ensembles de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN), un ensemble de vecteurs de base pour chacun desdits réplicateurs (E1, E2 ... EN), chacun desdits ensembles de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN) comprenant une pluralité de vecteurs de base, chacun desdits ensembles de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN) étant utilisé pour répliquer l'ensemble de données identifié correspondant parmi lesdits ensembles de données identifiées (C1, C2 ... CN) ;

   dans lequel la création d'un ensemble de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN) pour chacun desdits réplicateurs (E1, E2 ... EN), comprend :

la construction (S4, S10, S16) d'un ensemble de données de comparaison (C1r, C2r ... CNr) utilisant l'ensemble de vecteurs de base correspondant (AFRCN1, AFRNC2 ... AFRNCN) ;

la comparaison de l'ensemble de données antérieurement identifiées (C1, C2 ... CN) à l'ensemble de données de comparaison (C1r, C2r... CNr) ;

le calcul (S5, S11, S17) d'une erreur à partir de ladite comparaison ;

la détermination (S6, S12, S18) de l'acceptabilité de l'erreur ; et

la répétition de ladite création de l'ensemble de vecteurs de base en réponse à une erreur inacceptable.


 
7. Procédé selon la revendication 6 dans lequel dans ladite étape de création, les ensembles de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN) sont générés à partir d'une combinaison quelconque des items suivants :

trains d'onde à fréquence variable, transitoires, gazouillis de Fourier, fonctions à base fractale, fonctions à base radiale, fonctions de base générées en utilisant des perceptrons multicouches, et autres fonctions de base générées par ordinateur.


 
8. Procédé pour classer des ensembles de données (U) comprenant les étapes suivantes :

a) entrer des ensembles de données antérieurement identifiés (C1, C2 ... CN) dans un ordinateur (MP1, MP2 ... MPN) ;

b) créer dans l'ordinateur (MP1, MP2 ... MPN) une pluralité d'ensembles de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN) en correspondance un à un avec lesdits ensembles de données antérieurement identifiés (C1, C2 ... CN), chacun définissant un réplicateur (E1, E2 ... EN) ;

c) entrer un nouvel ensemble de données (U) dans l'ordinateur (MP1, MP2 ... MPN) ;

d) pour chaque réplicateur (E1, E2 ... EN), générer (S50, S60, S70) une réplique (Ur) du nouvel ensemble de données (U) en utilisant les ensembles de vecteurs de base (AFRCN1, AFRNC2 ... AFRNCN) ;

e) comparer (S51) la réplique (Ur1) générée par le premier réplicateur (E1) de la pluralité de réplicateurs (E1, E2 ... EN) au nouvel ensemble de données (U) ;

f) déterminer une réplication acceptable (S52) du nouvel ensemble de données (Ur1) généré par le premier réplicateur (E1) en utilisant un seuil d'erreur prédéterminé ;

g) répéter (S61, S71, S62, S72) les étapes e) et f) pour chacun des réplicateurs (E2 ... EN) et chacunes des répliques (Ur2 ... Urn) générées par le réplicateur correspondant (E2 ... EN) ;

h) déterminer (S53, S63, S73) si plus d'un réplicateur a répliqué de manière acceptable le nouvel ensemble de données (U) dans les étapes répétées f) ;

i) marquer le nouvel ensemble de données (U) en vue d'une analyse humaine en réponse à une détermination dans l'étape h) que plus d'un réplicateur (E1, E2 ... EN) a répliqué de manière acceptable le nouvel ensemble de données (U) ;

j) classer (S53, S63, S73) les nouvelles données (U) en réponse à la détermination dans l'étape h) qu'un seul réplicateur (E1, E2 ... EN) a répliqué de manière acceptable le nouvel ensemble de données (U).


 
9. Procédé pour classer des données selon la revendication 8, comprenant de plus l'étape (S82, S83) de détermination du besoin de nouveaux réplicateurs (E1, E2 ... EN) en réponse à une quantité de seuil prédéterminée de classement de données.
 
10. Procédé pour classer des données selon la revendication 8, comprenant de plus l'étape de marquage des nouvelles données en vue d'une analyse humaine en réponse à la détermination dans l'étape h) du fait qu'aucun réplicateur (E1, E2 ... EN) n'a répliqué de manière acceptable les nouvelles données (U).
 
11. Procédé pour classer des données selon la revendication 8, comprenant de plus l'étape de détermination du besoin de création automatique d'un nouveau réplicateur en réponse à l'identification et au classement d'une quantité excessive de données par l'un quelconque des réplicateurs (E1, E2 ... EN) après un temps prédéterminé ou après un seuil prédéterminé de données répliquées et classées.
 
12. Système informatique pour répliquer et classer des données, comprenant :

un microprocesseur (MP1, MP2 ... MPN) configuré pour :

s'entraîner à répliquer des ensembles identifiés de données (C1, C2 ... CN), créant de ce fait une pluralité de répliquateurs (E1, E2 ... EN), chaque réplicateur (E1, E2 ... EN) étant entraîné pour répliquer un ensemble de données parmi les ensembles identifiés de données (C1, C2 ... CN),

recevoir des données inconnues (U),

faire en sorte que chacun des réplicateurs (E1, E2 ... EN) tente de répliquer les données inconnues (U) , pour chaque réplicateur (E1, E2 ... EN) , comparer les données inconnues (U) avec les données répliquées (Ur) ;

déterminer si chaque réplicateur (E1, E2 ... EN) a reproduit de manière acceptable les données inconnues (U) en restant sous une erreur de seuil prédéterminée ; et

en réponse à la détermination d'une réplication acceptable par seulement un réplicateur (E1, E2 ... EN) classer les données inconnues (U).


 
13. Système informatique pour répliquer et classer des données selon la revendication 12, dans lequel ledit microprocesseur (MP1, MP2 ... MPN) est de plus adapté pour déterminer si l'erreur de seuil pour chacun des réplicateurs (E1, E2 ... EN) nécessite une révision en réponse à des quantités significatives de classement de données par ce réplicateur (E1, E2 ... EN).
 
14. Système informatique pour répliquer et classer des données selon la revendication 12, dans lequel ledit microprocesseur (MP1, MP2 ... MPN) est de plus adapté pour déterminer si de nouveaux réplicateurs (AFRCN1, AFRNC2 ... AFRNCN) et les ensembles de vecteurs de base correspondants (C1, C2 ... CN) sont nécessaires pour classer les données.
 
15. Système informatique pour répliquer et classer des données selon la revendication 12, dans lequel ledit microprocesseur comprend une pluralité de microprocesseurs (MP1, MP2 ... MPN) liés l'un à l'autre.
 
16. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel lesdits microprocesseurs (MP1, MP2 ... MPN) sont liés l'un à l'autre par l'intermédiaire d'un négociateur.
 
17. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel lesdits microprocesseurs (MP1, MP2 ... MPN) sont liés l'un à l'autre pour un multitraitement symétrique.
 
18. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel lesdits microprocesseurs (MP1, MP2 ... MPN) sont liés l'un à l'autre pour un multitraitement asymétrique.
 
19. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel lesdits microprocesseurs (MP1, MP2 ... MPN) sont liés l'un à l'autre au niveau boîtier.
 
20. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel lesdits microprocesseurs (MP1, MP2 ... MPN) sont liés l'un à l'autre dans un réseau.
 
21. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel lesdits microprocesseurs (MP1, MP2 ... MPN) sont liés l'un à l'autre en grappes.
 
22. Système informatique pour répliquer et classer des données selon la revendication 15, dans lequel ledit système informatique est configuré pour le génotypage.
 




Drawing