JP2002503002A - N-tuple or RAM based neural network classification system and method - Google Patents

Abstract

SUMMARY OF THE INVENTION The present invention provides a sampled training input data, each defined by a vector having cells with values, each containing a fixed number of rows corresponding to at least one subset of possible classes. A method and system for training a computer classification system, which may be constituted by a network comprising a fixed number of n-tuples or look-up tables (LUTs), comprising columns addressed by example elements or signals. Example training input data for different classes in a manner such that at least a portion of the cells include or point to information based on the number of times corresponding cell addresses are sampled from the set or sets of training input examples. Determining a column vector cell value based on the training set or sets; Determine the weight cell value for the column vector cell or cells being addressed or sampled by the training example, thus also allowing for weighting of the positive value column cell or cells during the classification process And determining the weighted cell value based on at least a portion of the determined column vector cell value and by using at least a portion of a training set of input examples. A second aspect of the invention is a method and system for determining a weight cell value corresponding to one or more column vector cells addressed by a training example in a computer classification system, the determination comprising: A method and system based on information of at least a portion of the calculated vector cell values, which allows for weighting of column vector cells having positive or non-positive values. Finally, the present invention provides a method and system for classifying example input data into a plurality of classes using a computer classification system.

Description

DETAILED DESCRIPTION OF THE INVENTION

BACKGROUND OF THE INVENTION FIELD OF THE INVENTION The present invention relates generally to n-tuple or RAM-based neural network classification systems, and more particularly to n-tuple or RAM-based neural network elements whose element values are determined during a training process and have weight vectors. Related to classification system.

[0002] 2. 2. Description of the Prior Art Known methods of classifying objects or patterns represented by electrical signals or binary codes and, more precisely, signal vectors applied to the inputs of a neural network classification system are known for performing so-called learning or training phases. is there. This step generally involves using one or more signals, referred to as a training or training set, in which each such signal is known to belong to one of the classes of interest for which signal classification is desired. And the configuration of a classification network that performs the function of performing the classification of the object under consideration as efficiently as possible. This method
Also known as supervised or supervised learning.

A subclass of the classification network using supervised learning is a network using memory-based learning. Here, the oldest memory-based networks were proposed by Bledsoe and Browning (Bledsoe, WW and Brownin
g, 7.1959, "Pattern Recognition and Reading by Machine", Proceedings of the Eastern Computer Conference, p225-232, and recently described by Morciniec and Rohwer (Morciniec, M. and Rohwer, R., 1996, "n- Theoretical and experimental reports on the performance of tuple classifiers ”Neural Comp., P. 629-642),“ n
-Tuple network ".

[0004] One of the advantages of such a memory-based system is the very fast computation time, both during the learning phase and during classification. For known types of n-tuple networks, also known as "RAM networks" or "weightless neural networks", the learning is done in random access memory (RAM), which requires only one presentation to the system. This can be achieved by storing the features of the pattern.

A training procedure for a conventional RAM-based neural network is described in Jorgense
n (co-inventor of the present invention) et al. (Jorgensen, TM Christensen, SS and Lii
sberg, C., 1995, "Cross-validation and information measures for RAM-based neural networks" Minutes of Neural Network Workshop WN Network 95 (Canterbury, Kent, UK) Minutes, edited by D. Bisset, pp. 76-8.
1), which describes how a RAM-based neural network can be considered as including a fixed number of look-up tables (LUTs). Each LUT can probe a subset of the binary input data vector. In conventional schemes, the bits used are chosen at random. The sampled bit system is used to construct one address. This address corresponds to a specific entry (column) in the LUT. The number of rows in the LUT corresponds to the number of possible classes.
For each class, the output can have a value of 0 or 1. A value of 1 corresponds to one vote on that particular class. When performing classification, the input vectors are sampled and the output vectors from all LUGs are added, after which the winner makes all decisions to classify the input vectors. To perform simple training of the network, the output value can be initially set to zero. For each example in the training set,
The following steps must be performed: present the input vector and the target class to the network, calculate its corresponding column entry for every LUT, and in every "active" column, Set the output value of the class to 1.

[0006] By using such a training strategy, it can be ensured that each training pattern always gets the maximum number of votes. As a result, such a network does not make any classification errors for the training set, but ambiguous decisions can occur. Here, the generalization capability of the network is directly related to the number of input bits for each LUT. If the LUT samples all input bits, it will act as a pure memory device and will not provide any generalization. As the number of input bits decreases, the generalization increases at the cost of increasing the number of ambiguous decisions. Furthermore, the classification and generalization performance of the LUT
It depends heavily on the actual subset of input bits examined. Thus, the purpose of the "intelligent" training procedure is to select the most appropriate subset of the input data.

[0007] Jorgensen et al. Further suggest a method for selecting the optimal number of input connections to use per LUT in order to obtain a low classification error rate with a short overall computation time, "Mutually valid. Sex Test ".
In order to perform such a cross-validation test, it is necessary to visit the elements or cells corresponding to the addressed columns and classes and to know about the number of actual training examples addressed. Therefore, it is suggested that these numbers be stored in the LUT. Similarly, Jorgencen et al. Suggests that by continuously training a new set of LUTs and performing a cross-validation test on each LUT, it is possible to select the LUTs in the network in a more optimal way. I have. Thus, by presenting the training set to the system several times, L
It has been found that a RAM network is obtained in which the UT is selected.

In one paper by Jorgensen (co-inventor of the present invention) (Jorgensen, IM “R
Classification of Handwritten Numbers Using AM Neural Network Architecture ", 1997
February, International Journal of neural Systems, Vol. 8, No. 1 p17
-25), suggesting how class-recognition of RAM-based networks can be further improved by extending conventional RAM architectures to include what is termed "inhibition." This approach addresses the problem that, in many situations, two different classes may differ only in some of their features. Thus, to address this problem, it is necessary to weight different features differently for a given class. Thus, methods have been suggested in which the network includes inhibitory factors for some classes of addressed columns. Here, a measure of reliability is introduced, and the suppression factor is calculated in such a way that the reliability after suppression corresponds to the desired level.

[0009] The result of the preferred suppression scheme is that all addressed LUT cells or elements are set to 1 in the correction pattern, as would be set to 1 in a simple system, but modified. In the version, the column cell set to 1 may further include information about the number of times the cell has been visited by the training set. However, in a simple system, the portion of the cell that contains a zero will cause its content to change to a negative value in the modified network. In other words, conventional networks are extended to allow deterrence from one class to another.

[0010] One bit per cell or element is not enough to encode negative values in LUT cells, as in conventional RAM networks. Thus, values less than 128 are used to represent different negative values while values greater than 128 are used to store information about the number of training examples that visited or addressed the cell. Preferably, one byte is used per cell. At this time, at the time of classifying one object, an addressed cell having a value of 1 or more can be counted as having a value of 1.

By using inhibition, the cells of the LUT are given different values that can be considered a kind of “weighting”. However, only cells that have never been visited by the training set can be suppressed by changing its value from 0 to a negative value. When performing input data classification, there is no boost of cells with positive values. LUTs or columns of LUTs that perform very well in this way can easily be overwhelmed with the rest of the network.

Thus, while allowing a very fast training or learning phase, at the same time, based on the training set access information, the actual weights are LUTs to obtain the proper ability to generate a sampled number of input bits. There is a need for a RAM classification network that allows both boosting and suppression of column cell values.

According to a first aspect of the present invention, each includes a fixed number of rows corresponding to at least one subset of possible classes, and is further defined by vectors each having a cell with a value. A computer classification system that can be defined by a network that includes a fixed number of n-tuples or look-up tables (LUTs), comprising a fixed number of columns addressed by elements or signals of sampled training input data Wherein at least a portion of the cells include or point to information based on the number of times corresponding cell addresses are sampled from the set or sets of training input examples, and input for different classes. Column vectors based on one or more training sets of example data Determining the cell values, and the method comprising the step of determining the weight cell values corresponding to one or more column vector cells are being trained examples addressed or sampled is provided.

According to a second aspect of the invention, each includes a fixed number of rows corresponding to at least one subset of the possible classes, and at least a portion has a corresponding weight vector and each is sampled. Based on one or more training sets of example input data for a class comprising a fixed number of column vectors addressed by the elements or signals of the example training input data, each column vector and weight vector being different. A method for determining a weight cell value in a computer classification system that may be defined by a network including a fixed number of n-tuples or look-up tables (LUTs) having cells with a value that has been determined. At least a portion of the corresponding cell address from the training input example set. Determining column vector cell values based on the training set of input examples in such a way as to include or point to information based on the number of times the sample is sampled, and a weight vector corresponding to one or more column vector cells A method is provided that comprises determining a cell value.

Preferably, the weight cell values are determined based on information of at least a portion of the determined column vector cell values and by using at least a portion of the training set of input examples. According to the invention, the example training input data is preferably presented to the network as an input signal vector. The determination of the weight cell value is preferably implemented to allow weighting of the positive value of the column vector cell or cells and to allow boosting of the column vector cell or cells during the classification process. . Additionally or alternatively, the weight cell values may be determined to allow for suppression of one or more column vector cells during the classification process.

The present invention also provides for determining one or more column vector cells having a positive value (greater than zero) and one or more column vector cells having a non-positive value (less than or equal to 0) by determining a weight cell value. Is also provided. Preferably, the determination of the weight cell enables weighting of an arbitrary column vector cell. To determine or calculate the weight cell values, determining these values includes initializing one or more sets of weight cells corresponding to at least a portion of the column cells, and determining at least a portion of the determined cell value. Adjusting at least a portion of the weight cell values based on the information and by using at least a portion of the training set of input examples may be included. At the time of determining the weight cell values, they are preferably arranged in the form of a weight vector corresponding to at least one part of the column vector.

In order to determine or adjust a weight cell value according to the present invention, a column cell value must be determined. Here, it is preferred that at least a portion of the column cell values is determined as a function of the number of times the corresponding cell address is sampled from the set of training input examples. Alternatively, the column cell information has an associated value that is a function of the number of times the corresponding cell address is sampled from the training set of input examples, even though the maximum column cell value is one. It may be determined in such a way as to do. Preferably, the column vector cell value is determined before adjustment of the weight vector cell value and stored in the storage means.

According to the present invention, a preferred method for determining column vector cell values includes: a) applying a known class of training input data examples to a classification network, thus addressing one or more column vectors. B) incrementing, preferably by one, the value or vote of a cell of an addressed column vector corresponding to a row of a known class; and c) until all training examples have been applied to the network. A training step of repeating the steps (a) and (b).

However, it should also be understood that the invention equally covers embodiments in which the column cell information is determined by an alternative function on the number of times the cell has been addressed by the input training set. . Thus, the cell information need not include a count of the total number of times the cell has been addressed, for example, when the cell has received zero, one, multiple and / or two and more than two visits, etc. Can be included.

Although it has been described above that the weight cell value can be determined for one or more column cells, in a preferred embodiment, every column vector has a corresponding weight vector. When initializing weight cell values according to embodiments of the present invention, the initialization may include setting each weight cell value to a predetermined specific cell value. These values may be different for different cells, but it is also possible to set all weight cell values to predetermined constant values. Such a value can be 0 or 1, but other values may be preferred.

In order to determine the weight cell values, these values are preferably adjusted, which may include one or more iteration steps. To adjust the weight cell value,
Determining a global quality value based on at least a portion of the weights and column vector cell values; determining whether the global quality value meets a required quality criterion;
And adjusting at least a portion of the weight cell value until a global quality criterion is met.

The adjustment process also includes determining a local quality value for each sampled training input, where the local quality value is defined or required for the selected input example. If the quality criteria are not met, one or more weight cell adjustments are performed. By way of example, adjusting the weight cell values may include the following steps: a) selecting input data from a training set; b) local quality corresponding to a sampled training input example. Determining a value, wherein the local quality value is a function of at least a portion of the addressed weights and column vector cell values; c) the local quality value meets required local quality criteria Determining whether or not, and if not, adjusting one or more of the addressed weight vector cell values if the local quality criterion is not met; c) defining a training set E) repeating the local quality test steps (b)-(d) for all given training input examples; f) during the local quality test Determining a global quality value based on at least a portion of the weights and column vectors being dressed; g) determining whether the global quality value meets required global quality criteria; and h) global. Repeating steps (a)-(g) until the quality criteria are met. Preferably, steps (b)-(d) of the adjustment process described above may be performed for all examples of the training set.

The local and / or global quality values may be defined as a function of the weight and / or at least a portion of the column cell. Correspondingly, the global and / or local quality criterion can likewise be a function of the weight and / or the column cell. Thus, the quality criterion or criterion does not need to be a predetermined constant threshold and can be changed during the adjustment iteration process. However, the invention likewise covers embodiments in which the quality criterion or measures are given by a constant threshold.

When adjusting the weight cell values by using one or more quality values, each with a corresponding quality criterion, if the quality criterion is not met after a given number of iterations, the adjustment iterative process is performed. It should be understood that stopping may be preferable. Similarly, during the adjustment process, the adjusted weight cell values are preferably stored after each adjustment, and if the adjustment process includes a determination of the global quality value, after the step of determining the global quality value, If the global quality value is closer to meeting the global quality criteria than the global quality value corresponding to the previously separately stored weight cell value or configuration value, the weight cell value or classification system configuration value thus obtained is It should also be understood that a separate storage step may follow.

The main reason for training a classification system according to one embodiment of the present invention is to obtain high reliability in the subsequent classification process of input examples of unknown classes. Thus, according to a further aspect of the present invention, there is provided a method of classifying an example input data into at least one of a plurality of classes using a computer classification system configured according to any of the above-described methods of the present invention, thus comprising: The column cell values and the corresponding weight cell values for each n-tuple or LUT based on one or more training sets of example input data, wherein: a) determining the input data to classify; Applying to the constructed classification network and thus addressing the column vectors and the corresponding weight vectors in the n-tuple or LUT b) selecting one class and thus identifying in the set of n-tuples or LUTs Addressing the rows of b) the output value as a function of the value of the addressed weight cell Constant stages, d) until the output of all of the classes has been determined, steps (b) - (c)
And d) comparing the calculated output values; and f) selecting the class or classes having the largest output value.

When classifying unknown input examples, multiple functions can be used to determine the output value from the addressed weight cell. However, it is preferred that the parameters used to determine the output value include the values of both the addressed weight cell and the addressed column cell. Thus, by way of example, the output value is determined as a first sum of all addressed weight cell values corresponding to column cell values greater than or equal to a predetermined value. In another preferred embodiment, determining the output value comprises determining a first sum of all addressed weight cell values corresponding to column vector cell values greater than or equal to a predetermined value. Determining a second sum of the addressed weighted cell values and determining an output value by dividing the first sum by the second sum.
The predetermined value is preferably set to one.

The present invention also provides a training and classification system according to the training and classification method described above. Thus, in accordance with the present invention, sampled training input data each including a fixed number of rows corresponding to at least one subset of possible classes, and further defined by vectors having cells with values, respectively. In a training system of a computer classification system, which can be defined by a network comprising a fixed number of n-tuples or look-up tables (LUTs) comprising a fixed number of columns addressed by elements or signals of Means for receiving an example training input data, means for sampling the received example input data and addressing a column vector within a stored set of n-tuples or LUTs, n-tuples or LUTs Address a specific row corresponding to a known class within a set of Storage means for storing the determined n-tuple or LUT, a column for including or pointing to information based on the number of times the corresponding cell address is sampled from the training input example set. A system comprising: means for determining a vector cell value; and means for determining a weight cell value corresponding to one or more column vector cells being addressed and sampled by the training example. I have.

[0028] The present invention likewise comprises a fixed number of rows, each corresponding to at least one subset of the number of possible classes, and at least a portion having a corresponding weight vector, each of which is sampled. A number of column vectors addressed by the elements or signals of the example input data for training, and each column vector and weight vector is determined during the training process based on the set or sets of training input data examples. Training a known class in a system for determining weight cell values of a classification network that can be defined by stored numbers of n-tuples or look-up tables (LUTs) having cell values being determined Means for receiving input data examples for use, sampling the received input data examples Means for addressing a column vector and a corresponding weight vector within a stored set of n-tuples or LUTs, addressing a particular row corresponding to a known class within a set of n-tuples or LUTs Means for designating, storage means for storing the determined n-tuple or LUT, a column vector to contain or point to information based on the number of times the corresponding cell address is sampled from the training input example set. There is also provided a system comprising means for determining a cell value and means for determining a weight vector cell value corresponding to one or more column vector cells.

Here, the means for determining the weight cell values are based on information of at least a portion of the determined column vector cell values and by using at least a portion of the training set of the input examples. Preferably, it is adapted to determine a value. Preferably, the means for determining the weight cell value is such that one or more column cells having a positive value can be weighted and / or the column cell or cells can be boosted during the classification process. Is adapted to determine the value of The means for determining may additionally or alternatively be adapted to determine the weight cell value to enable suppression of one or more column vector cells during the classification process.

According to one embodiment of the invention, the weight determining means comprises one or more column vector cells having a positive value (greater than zero) and one or more column vector cells having a non-positive value (0 or less). It can be adapted to determine weight cell values to allow for weighting of cells. Preferably, the means may be further adapted to determine a weight cell value to allow for weighting of any column cell. Similarly, it is also preferred that the means for determining the weight cell values are adapted to determine these values such that the weight cell values are arranged in the form of a weight vector corresponding to at least a part of the column vector. .

To determine the weight cell value according to a preferred embodiment of the present invention, the means for determining the weight cell value includes initializing one or more sets of weight vectors corresponding to at least a portion of the column vector. Means for adjusting a weight vector cell value of at least a portion of the weight vector based on information of at least a portion of the determined column vector cell value and by using at least a portion of the training input example set. Means may be included.

As described above, in order to determine a weight cell value, a column cell value must be determined. Here, the means for determining the column vector cell value is
Preferably, it is adapted to determine these values as a function of the number of times the corresponding cell address is sampled from the training input example set. Alternatively, the means for determining the column vector cell value may be such that at least a portion of the cells having a maximum value of 1 is a function of the number of times the corresponding cell address is sampled from the training input example set. It may be adapted to determine these values in such a way as to have certain associated values.

According to one embodiment of the present invention, the example training input data belonging to a known class is applied to a classification network, thus determining a column vector cell value when one or more column vectors are addressed. The means for adapting is preferably adapted to increment the value or vote of the cell of the addressed column vector corresponding to a known class of rows, this value being preferably incremented by one.

To initialize a weight cell according to an embodiment of the present invention, the means for initializing the weight vector is adapted to set the weight cell value to one or more predetermined values. Is preferred. For the weight cell adjustment process, the means for adjusting the weight vector cell value determines a global quality value based on the weight and at least a portion of the column vector cell value, wherein the global quality value is needed. Preferably, it is adapted to determine whether a quality criterion is met and to adjust at least a portion of the weight vector cells until the global quality criterion is met.

As an example of a preferred embodiment according to the present invention, the means for adjusting the weight vector cell value comprises: a) weight and column vector cell values addressed corresponding to the sampled training input. B) determining whether the local quality value satisfies a required local quality criterion; b) determining an address if the local quality criterion is not satisfied. Adjusting one or more of the specified weight vector cell values; c) repeating the local quality test for a predetermined number of training input examples; d) being addressed during the local quality test. Determining a global quality value based on at least a portion of the weights and the column vector; e) determining whether the global quality value meets required global quality criteria; f) It may be adapted to repeat the local and global quality tests until the global quality criteria are met.

The means for adjusting the weight vector cell value may be adapted to stop the iterative process after a given number of iterations if the global quality criterion is not met. In a preferred embodiment, the means for storing the n-tuple or LUT comprises means for storing the adjusted weight cell values and storing the highest weight cell value or the highest system configuration value so far. Separate means for doing so are included. Wherein the means for adjusting the weight vector cell value further comprises: determining that the determined global quality value satisfies a global quality criterion higher than the global quality value corresponding to the highest weight value previously stored separately. , Is adapted to replace the previously weighted cell value previously stored separately with the adjusted cell value obtained. Thus, even if the system cannot meet the global quality criteria within a given number of iterations, the system can always include the “best ever” system configuration.

According to a further aspect of the present invention, a system for classifying an example of input data of an unknown class into at least one of a plurality of classes, wherein each system includes at least one subset of a certain number of possible classes. Contains a corresponding number of rows,
Further, a fixed number of n-
Comprising storage means for storing tuples or look-up tables (LUTS) or sets, wherein each column vector is addressed by an element or signal of the sampled input data example, and each column vector and weight vector is A system having cells with values determined during a training process based on one or more training input data set sets, the input means for receiving the training input data examples to be classified; Means for sampling the received input data example and addressing the columns and corresponding weight vectors in the stored n-tuple or LUT set, a particular row corresponding to a particular class, an n-tuple or Means for addressing in the LUT set, the output value as a function of the addressed weight cell There is also provided a system further comprising means for determining and comparing the calculated output values corresponding to all classes and selecting one or more classes having a maximum output. .

According to a preferred embodiment of the classification system of the present invention, the output determining means includes all of the addressed weight vector cell values corresponding to a column vector cell value equal to or greater than a predetermined value and corresponding to a specific class. Means for generating a first sum are included. Similarly, the output determining means includes means for generating a second sum of all addressed weight vector cell values corresponding to the particular class, and dividing the first sum by the second sum. Preferably, means for determining the output value is further included.

It is understood that the cell values of the columns and weight vectors of the classification system according to the invention are preferably determined by using a training system according to any of the systems described above. Should be. Thus, these cell values may be determined during the training process according to any of the methods described above. For a better understanding of the present invention and to show how it may be carried into effect, reference is now made by way of example to the accompanying drawings.

DETAILED DESCRIPTION OF THE INVENTION In the following, a more detailed description of the architecture and concepts of a classification system according to the present invention is provided, including an example of a training process and an example of a classification process for the column cells of the architecture. . Further, different examples of the learning process for weight cells according to embodiments of the present invention are described.

Notation The notation used in the following description and examples is as follows. X: Training set

[0042]

(Equation 1)

N _X : number of examples in training set X

[0044]

(Equation 2)

C: Class label

[0046]

(Equation 3)

C _W : winner class obtained by classification C _R : runner-up class obtained by classification

[0048]

(Equation 4)

N _C : number of training classes corresponding to the maximum number of rows in the LUT Ω: LUT set (each LUT can contain only a subset of all possible address columns, and different columns can be existing classes N _LUT : number of LUTs N _COL : number of different columns that can be addressed within a particular LUT (LUT-dependent) S _C : set of training examples labeled class C W _iC : weight for cell addressed by i-th column and C-th class V _iC : entry counter for cell addressed by i-th column and C-th class

[0050]

(Equation 5)

DESCRIPTION OF ARCHITECTURE AND CONCEPT In the following, FIG. 1 shows a block diagram of a RAM classification network with a look-up table (LUTS) and a single look-up table (LU) according to the invention.
Please refer to FIG. 2 showing a detailed block diagram of T). The RAM-net or LUT-net has a fixed number of lookup tables (LUTs).
) (1.3). Let the number of _LUTs be denoted as N _LUT . An example of the input data vector -y to be classified is the input module (1) of the LUT network.
. 1). Each LUT can sample a portion of the input data, where a different number of input signals can be sampled for different LUTs (1.2) (in principle, one that samples the entire input space).
It is also possible to have two LUTs). The output of the LUT can be provided to the output module (1.4) of the RAM classification network.

In FIG. 2, for each LUT, the sampled input data (2.1) presented on the LUT-net can be supplied to the address selection module (2.2). It is shown. Address selection module (2.2
) Can calculate the address of one or more specific columns (2.3) in the LUT from the input data. One example is input example

[0053]

(Equation 6)

The index of the column in the i-th LUT addressed by

[0055]

(Equation 7)

It is assumed to be calculated as The number of addressable columns in a particular LUT can be represented as N _COL , which generally varies between LUTs. The information stored in a particular row of the LUT may correspond to a particular class C (2.4). At this time, the maximum number of rows may correspond to the number of classes N _C. In a preferred embodiment, every column in the LUT contains two sets of cells. The number of cells in each set corresponds to the number of rows in the LUT. The first set of cells may be represented as column vector cells, and the cell value may correspond to a class specific entry counter for the column in question. Another set of cells can be denoted as weight cells or weight vector cells with cell values that can correspond to weight factors that can be associated with one entry counter value or column vector cell value, respectively. The entry counter value for the cell addressed by the i-th column and class C is denoted as V _iC ( _2.5 ). The weight value for the cell addressed by the i-th column and class C is labeled W _iC ( _2.6 ).

The V _iC and W _iC values of the activated LUT column ( _2.7 ) are stored in the output module (
1.5) is provided (1.4), where a vote number can be calculated for each class, and finally a winner-takes-all (WTA) decision can be performed.

[0058]

(Equation 8)

Represents the input data used for training,

[0060]

(Equation 9)

Represents an example of input data that does not belong to the training set. or

[0062]

(Equation 10)

Is

[0064]

[Equation 11]

Represents the class to which belongs. An example

[0066]

(Equation 12)

The class assignment given to is then obtained by calculating the number of votes for each class. The number of votes obtained for class C is an example

[0068]

(Equation 13)

Calculated as a function of V _iC and W _iC addressed by:

[0070]

[Equation 14]

From the calculated number of votes, the winner class C _w can be obtained as follows:

[0072]

(Equation 15)

[0073]

(Equation 16)

An example of a sensitive choice for is:

[0075]

[Equation 17]

In the equation, δ _{i, j} is a Kronecker symbol (if i = j, δ _{i, j} =
1 and 0 otherwise),

[0077]

(Equation 18)

Ω describes the set of LUTs that make up the entire LUT network, and Sc
4 represents a set of training examples labeled class C. In the special case where all _WiC values are set to 1, a conventional LUT network is obtained.

[0079]

[Equation 19]

FIG. 3 shows an example of a block diagram of a computer classification system according to the present invention.
Here, a source, such as a video camera or a database, provides one or more input data signals (3.0) describing the examples to be classified. These data are supplied to a preprocessing module (3.1) of a type capable of extracting features in a predetermined manner and reducing and converting the input data. An example of such a pre-processing module is an FFT (Fast Fourier Transform) board. The transformed data is then provided to a classification unit (3.2) comprising a RAM network according to the invention. The classification unit (3.2) outputs a ranked classification list that may include associated reliability. The classification unit is implemented by using software to program a standard personal computer or by programming a hardware device, for example, by using a programmable gate array combined with a RAM circuit and a digital signal processor. Can be done. These data can be interpreted in a post-processing device (3.3), which can be a computer module that combines the obtained classification with other relevant information. Ultimately, the result of this interpretation is an output device such as an actuator (
3.4).

Initial Training of the Architecture The flow diagram of FIG. 4 is for determining a column vector entry counter or cell distribution that is a V _iC distribution (4.0) according to one embodiment of the present invention, which can be described as follows. Illustrates a one-pass learning scheme or process for: Cell value

[0082]

(Equation 20)

Initialize all entry counters or column vectors by setting

[0084]

(Equation 21)

Is initialized. This can be done by setting all weight values to constant factors or by choosing random values from within a certain range (4.1). 2. A first training input example from training set X;

[0086]

(Equation 22)

Is presented to the network (4.2, 4.3). 3. Compute the addressed column for the first LUT (4.4, 4.
5). 4.

[0088]

(Equation 23)

Add 1 to the entry counter in the row of the addressed column corresponding to the class label (in all LUTs)

[0090]

(Equation 24)

Is incremented) (4.6). 5. Repeat step 4 for the remaining LUTs (4.7, 4.8). 6. Repeat steps 3-5 for the remaining training input examples (4.9, 4.10
). The number of training examples is represented as N _X. Classification of Unknown Input Examples Once the RAM network of the present invention has been trained, and once the values for the column cells and weight cells that can define the LUT have been determined, the network can be used to classify the unknown input data examples. it can.

In a preferred example according to the invention, the classification is performed by determining the class with the largest number of votes, ie, Vote No, obtained by the following equation:

[0093]

(Equation 25)

If the denominator is zero, Vote No can be defined as being zero. Thus, an example of a classification can be described as follows with reference to FIGS. 1 and 2: An unknown input example

[0095]

(Equation 26)

Is presented to the network (1.1).・ For all LUTs,

[0097]

[Equation 27]

The column addressed by

[0099]

[Equation 28]

Is calculated (2, 3). For each class (corresponding to a particular row in each of the addressed columns)

[0101]

(Equation 29)

The sum (sum — 1) is generated.

[0103]

[Equation 30]

The term is

[0105]

(Equation 31)

In the case where

[0107]

(Equation 32)

It means that only the component is contained (1.5). For each class (corresponding to a particular row in each of the addressed columns)

[0109]

[Equation 33]

The sum (sum — 2) is generated (1.5). Calculate the output value corresponding to class C as Out [C] = sum_1 / sum_2 (1.5). Select the class (s) that maximizes Out [C] (1.5). FIG. 7 shows a block diagram of the operation of the computer classification system in which the classification process (7.0) is performed. The system acquires one or more input signals using, for example, an optical sensor system (7.1). The obtained input data is pre-processed in a pre-processing module, for example a low-pass filter (7.2) and presented to a classification module (7.3), which may be a LUT network according to one embodiment of the invention. The output data from the classification module is then post-processed in a post-processing module (7.4), for example a CRC algorithm that calculates a cyclic redundancy check sum, and the result is:
It is forwarded to an output device that can be a monitor screen (7.5).

Weight Adjustment Normally, the initially determined weight cell value does not suggest an optimal value selection. Thus, according to a preferred embodiment of the present invention, optimization or adjustment of the weight value must be performed. In order to select and adjust the weight values for the purpose of improving the performance of the classification system, it is suggested according to an embodiment of the present invention to define an appropriate quality function for measuring the performance of the weight values. . Thus, the local quality function

[0112]

(Equation 34)

Can be defined, where

[0114]

(Equation 35)

_Represents a vector containing all V _iC elements of the LUT network,

[0116]

[Equation 36]

_Represents a vector including all W _iC elements of the LUT network. The local quality function is a specific example

[0118]

(37)

A confidence measure for the output classification of can be provided. If the quality value does not meet certain given criteria (which may have changed dynamically between iterations), the weight

[0120]

(38)

Is that the quality value meets or is closer to meeting the criteria (if possible)
It is adjusted to make it so. In addition, the global quality function:

[0122]

[Equation 39]

Can also be defined. The global quality function can measure the performance of the input training set as a whole. FIG. 5 shows a flow chart for adjusting or learning the weight cells according to the present invention. FIG.
Flowchart illustrates a more general tuning or learning process that can be simplified for a particular embodiment.

Example 1 Input example

[0125]

(Equation 40)

The vote number function for is obtained from the following equation.

[0127]

[Equation 41]

According to this definition of the VotoNo () function, an input example of the training set

[0129]

(Equation 42)

A leave-one-out cross-validation for can be calculated as follows:

[0131]

[Equation 43]

This equation is:

[0133]

[Equation 44]

When is

[0135]

[Equation 45]

Equal to

[0137]

[Equation 46]

In the case of

[0139]

[Equation 47]

A factor equal to

[0141]

[Equation 48]

Except for, it is actually described above.

[0143]

[Equation 49]

Is

[0145]

[Equation 50]

It is 1 only when, and 0 in other cases. This merely ensures that one example cannot obtain a contribution from itself when calculating the leave-one-out cross-validation. An example

[0147]

(Equation 51)

Let the local quality function calculated for is defined as follows:

[0149]

(Equation 52)

Here, Q _L is

[0151]

(Equation 53)

Is 0 if generates a cross-validation error, and 1 otherwise.
Therefore, when Q _L = 0, weight change is performed. Let the global quality function be defined as:

[0153]

(Equation 54)

The global quality function is:

[0155]

[Equation 55]

Since each term in the sum over the training set is 1 if and 0 otherwise, it is properly classified if left outside the training cent. The number of examples from training set X is determined. Global quality standards are Leave One
The parameters that determine the fraction of the training example required to be properly classified in the out-cross-validation test may be those that satisfy Q _G > εN _X , with ε as the parameter.

An updating scheme to improve Q _G can be implemented according to the following rules. All training set input examples with incorrect cross-validation

[0158]

[Equation 56]

For, adjust the weights according to the following equation:

[0160]

[Equation 57]

[0161] Note that k is a small constant in the expression. A feasible choice of k may be one tenth of the average of the absolute values of the _WiC values. This update rule:

[0162]

[Equation 58]

In the case where

[0164]

[Equation 59]

That is, and

[0166]

[Equation 60]

When is

[0168]

[Equation 61]

Implies that The max () function ensures that the weight cannot be negative. Referring now to FIG. 5, the weight update or adjustment step of Example 1 can be described as follows. : Initialize all _WiC values to zero (5.1).

• Loop through all examples in the training set (5.2, 5.10, 5.3). Calculate the local quality value for each example (can the examples be properly classified if excluded from the training set?) (5.4, 5.5). • If the answer is yes, proceed to the next example (5.10). If the answer is no, then the corresponding column cell adds a positive contribution to the VoteNo () function (
Ie

[0171]

(Equation 62)

Increase the addressed weight of the “true” class, otherwise (ie,

[0173]

[Equation 63]

If), reduce the weight of the "true" class (5.6-5.9). Calculate the global quality value. If the quality is the best you have ever obtained,
Store the LUT network (5.11). • Repeat until the global quality value is satisfactory or other exit conditions are satisfied (5.12, 5.13).

Example 2 Input example

[0176]

[Equation 64]

Assume that the vote function for has been obtained as follows:

[0178]

[Equation 65]

The true class

[0180]

[Equation 66]

With respect to a given V value, using the following function:

[0182]

[Equation 67]

Sum the values of:

[0184]

[Equation 68]

The parameter I runs over all possible values of V _iC 0 <V _ic ≦ N _x .
The confidence Conf between the winning class C _W and the runner-up class C _R can then be defined as:

[0186]

[Equation 69]

The value m can be determined by the following function:

[0188]

[Equation 70]

The upper limit of the total index n is

[0190]

[Equation 71]

It can vary from 1 to the maximum V _iC value in the vector. This expression is

[0192]

[Equation 72]

It states that m is selected as the maximum value of n. The local quality function can now be defined as:

[0194]

[Equation 73]

In the equation, m _thresh is a threshold constant. If a Q _L <0, the weight W _ij is updated to increase the Q _L by adjusting the weights on the runner-up class C _R:

[0196]

[Equation 74]

The update rule is:

[0198]

[Equation 75]

When is

[0200]

[Equation 76]

[0201]

[0202]

[Equation 77]

When is

[0204]

[Equation 78]

Implies that Global quality criteria may be based on two functions:

[0206]

[Expression 79]

Here, Θ _O (Q _L ) is 1 when Q _L ≧ 0, and 0 when O _L <0.
It is. O _G1 measures the number of examples from the training set that can pass the leave-one-out cross-validation test, and Q _G2 measures the number of examples that can pass the local quality criteria. These two quality functions can then be combined into one quality function based on the following logical expression (true expressions are given a value of 1 and false expressions are given a value of 0): :

[0208]

[Equation 80]

Here, ε ₁ and Ε ₂ are the two parameters that determine the fraction of the training example required to pass the leave-one-out cross-validation test and local quality criteria, respectively. If you pass both of these criteria, you have passed the global quality criteria,

[0210]

(Equation 81)

Is 1; otherwise it is 0. Referring to FIG. 5, the weight update or adjustment step of Example 2 can be described as follows: Initialize all _WiC values to zero (5.1). • Loop through all examples in the training set (5.2, 5.10, 5.3)
. Calculate the local quality value for each example (5.4). (Does the example have sufficient "support"? (5.5).-If the answer is "yes," proceed to the next example (5.10). If the answer is "no," run-up class. Reduce the weight associated with voting for,

[0212]

(Equation 82)

Increase the weight associated with the cell with • Calculate global quality values. If the quality is the best obtained so far, remember the network (5.11). • Iterate until the global quality value is satisfactory or other exit conditions are met (5, 12, 5, 13).

Example 3 Also in this case, an input example

[0215]

[Equation 83]

The number of votes function for is obtained as follows.

[0219]

[Equation 84]

Local Quality Function

[0219]

[Equation 85]

Are examples of input training

[0221]

[Equation 86]

Is defined as a measure of vote reliability. An example

[0223]

[Equation 87]

For the true class

[0225]

[Equation 88]

The confidence Conf between the runner class C _R and the runner class C _R can be determined as follows:

[0227]

[Equation 89]

The confidence may be zero indicating that the next class has a vote level equal to that of the true class (the class or classes have a vote level equal to that of the true class) In that case, we would define one of the classes different from the true class as the next class). The local quality function can then be defined as:

[0229]

[Equation 90]

A threshold can be determined for the calculated local quality value, and Q_L<Q_{threshol d} , The weight is Q_LIs updated to increase. Q_thresholdConsideration of
The value obtained is that the difference between the vote level of the true class and the vote level of the next class is small.
Both seem to be 0.1 indicating that they should be 10% of the maximum vote level
. The weight is the weight C for the next class._RAdjusting the

[0231]

[Equation 91]

(Where k is a small constant) and adjusting the weight for the true class;

[0233]

(Equation 92)

Can be updated by The small constant k determines the relative change in weight to be adjusted. Possible 1
One option would be k = 0.05. Again, the number of cross-validation errors is a possible global quality measure.

[0235]

[Equation 93]

The global quality criterion may be one that satisfies Q _G > εN _X , where ε is a training example that is required to be properly classified within the leave-one-out cross-validation test Is a parameter for determining the fractionation. Referring to FIG. 5, the weight update or adjustment step of Example 3 can be described as follows: Initialize all _WiC values to zero (5.1).

Loop through all examples in the training set (5, 2, 5, 10, 5, 3)
). Calculate the local quality value for each example (5, 4). (Can the example be properly classified and at the same time have sufficient "support" when excluded from the training set? "(5, 5)).

If the answer is “yes”, proceed to the next example, if the answer is “no”,
Update the weight associated with the cell voting for the next class, update the weight associated with the cell voting for the true class, increase the vote level on the true class, and change the vote level for the next class. Decrease. • Calculate global quality values. If the quality is the best obtained so far, remember the network (5.11).

• Iterate (5, 12, 5, 13) until the global quality value is satisfactory or other exit conditions are satisfied. Example 4 Input example

[0240]

[Equation 94]

The vote number function for is defined as follows.

[0242]

[Equation 95]

The vote levels obtained for the training example when performing a cross-validation test are then as follows:

[0244]

[Equation 96]

Also here

[0246]

(97)

The next class obtained using can be denoted as C _R (if the class or classes have a vote level equal to that of the true class, we will use a different class than the true class Are defined as runner-up classes). Then the local quality function

[0248]

[Equation 98]

Can be defined by the following logical formula:

[0250]

[Equation 99]

In the formula, k ₁ and k ₂ are two constants between 0 and 1, and k ₁ > k ₂ . All three criteria

[0252]

[Equation 100]

If is satisfied,

[0254]

[Equation 101]

Is 1 and otherwise it is 0. Two criteria, leave-one-out cross-validation that vote level and that the runner-up class vote level of the true class is larger than k ₁ is lower than the k ₂ in the test and the level k ₁ level corresponding to seek to become greater than k _2. The Vote No () function used in this example would have a value between 0 and 1 if the weights were restricted to have positive values, in which case the possible choices for k are: it is equal k ₂ to k ₁ and 0.6 equivalent to 0.9.

[0256]

[Equation 102]

[0257] If the criteria given for the local quality value given is not satisfied by the weight W _iCR is to meet the criteria for possible Q _L, by adjusting the weights for the runner-up class C _R,

[0258]

[Equation 103]

( _Where k ₃ is a small constant) and is updated by adjusting the weight W _iCR for the true class.

[0260]

[Equation 104]

K ₃ determines the relative change in weight to be adjusted for the next class. One possible option is that k ₃ is 0.1. A possible choice of k ₄ may be one-tenth of the average of the absolute values of the values of W _iC . A suitable global quality function may be defined by the sum of the local quality values for all training input examples:

[0262]

[Equation 105]

The global quality criterion may be one that satisfies Q _G > εN _X , where ε is a parameter that determines the fraction of the training example required to pass the local quality test. Referring to FIG. 5, the weight update or adjustment step of Example 4 can be described as follows: Initialize all _WiC values to zero (5.1).

Loop through all examples in the training set (5.2, 5.10, 5.3)
). • Calculate the local quality for each example (5.4). (Can the example be properly classified if excluded from the training set, and at the same time have sufficient “voting” of votes? ”(5.5)).

If the answer is “yes”, proceed to the next example; if the answer is “no”, update the weight associated with the cell voting for the next class and vote for the true class The weight associated with the cell being updated is updated to increase the vote level on the true class and decrease the vote level for the next class. • Calculate the global quality function. If the quality is the best obtained so far, remember the network (5.11).

• Iterate (5.12, 5.13) until the global quality value is satisfactory or other exit conditions are satisfied. Example 5 In this example, input example

[0267]

[Equation 106]

The function of the number of votes for is defined as follows.

[0269]

[Equation 107]

The local quality function and the threshold criterion are defined such that the answer to the question “isQlocal OK” is always “No”. Thus, the local quality function can be defined as: In the Q _L = FALSE of these definitions pickled, since the answer is always "no" to (5.5),

[0271]

[Equation 108]

To adjust, all training examples will be used. The weight update rules are as follows:

[0273]

(Equation 109)

In the formula,

[0275]

[Equation 110]

Is

[0277]

(Equation 111)

Where α is the number of iterations. The global quality function for the αth iteration can be defined as:

[0279]

[Equation 112]

In the formula,

[0281]

[Equation 113]

Referring to FIG. 5, the weight update or adjustment step of Example 5 can be described as follows: Initialize all _WiC values to zero (5.1). Loop through all examples in the training set (5.2, 5.10, 5.3
).

• Calculate the local quality value for each example (5.4). (In this example it is always false, ie it does not meet the quality standards). If _QL = TRUE (5.5), proceed to the next example (which will never be true in this example), otherwise depend on the actual iteration

[0284]

[Equation 114]

The weight designated by using is set (5.6-5.9). Calculate global quality values. If the quality is the best obtained so far, remember the network (5.11). • Repeat until the last iteration (5.12, 5.13). Thus, Example 5 above is compatible with the flow diagram structure shown in FIG. However, since the answer to (5.5) is always "No", the weight assignment procedure will in this case refer to FIG. 6, which shows a flow chart of a more simplified weight cell adjustment according to the invention. And can be simplified as described below.

The number α _MAX of schemes for setting the W _iC value can be defined as (6
1,6.6,6.7): Scheme α: (6.2) For all LUTs, do the following: For all i, do the following: For all C, do the following :

[0287]

[Outside 1]

For each α scheme, a global quality function for classification performance can be calculated (6.3). One possibility for the global quality function consists in calculating the cross-validation errors.

[0289]

[Equation 115]

In the formula,

[0291]

[Equation 116]

At this time, the highest quality value

[0293]

[Formula 117]

(6.4, 6.5). Here, it should be understood that another number of iterations can be selected, and other suitable quality functions can be defined. The foregoing description of a preferred embodiment of the invention has been presented for purposes of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise forms disclosed, and obviously to those skilled in the art
Many modifications and variations are possible in light of the present invention. All such modifications that retain the basic principles disclosed and claimed herein are within the scope of the invention.

[Brief description of the drawings]

FIG. 1 is a block diagram of a RAM classification network with a look-up table (LUT).

FIG. 2 illustrates a single look-up table (LUT) according to one embodiment of the present invention.
FIG. 2 shows a detailed block diagram of FIG.

FIG. 3 shows a block diagram of a computer classification system according to the present invention.

FIG. 4 shows a flowchart of a learning process for an LUT column cell according to an embodiment of the present invention.

FIG. 5 is a flowchart of a learning process for a weight cell according to the first embodiment of the present invention.

FIG. 6 is a flowchart of a learning process for a weight cell according to the second embodiment of the present invention;

FIG. 7 shows a flowchart of a classification process according to the present invention.

──────────────────────────────────────────────────続 き Continuation of front page (81) Designated country EP (AT, BE, CH, CY, DE, DK, ES, FI, FR, GB, GR, IE, IT, LU, MC, NL, PT, SE ), OA (BF, BJ, CF, CG, CI, CM, GA, GN, GW, ML, MR, NE, SN, TD, TG), AP (GH, GM, KE, LS, MW, SD, SZ, UG, ZW), EA (AM, AZ, BY, KG, KZ, MD, RU, TJ, TM), AL, AM, AT, AU, AZ, BA, BB, BG, BR, BY, CA, CH, CN, CU, CZ, DE, DK, EE, ES, FI, GB, GD, GE, GH, GM, HR, HU, ID, IL, IN, IS, JP, KE , KG, KP, KR, KZ, LC, LK, LR, LS, LT, LU, LV, MD, MG, MK, MN, MW, MX, NO, NZ, PL, PT, RO, RU, SD, SE, SG, SI, SK, SL, TJ, TM, TR, TT, UA, UG, UZ, VN, YU, ZW It is determined by using at least a portion. A second aspect of the invention is a method and system for determining a weight cell value corresponding to one or more column vector cells addressed by a training example in a computer classification system. The method and system are based on information of at least a portion of the determined vector cell values, thereby enabling weighting of column vector cells having positive or non-positive values. Finally, the present invention provides a method and system for classifying example input data into a plurality of classes using a computer classification system.

Claims (49)

[Claims] 1. An element of the sampled training input data, each comprising a fixed number of rows corresponding to at least one subset of the possible classes and further defined by a vector having cells with values, respectively. Alternatively, in a method of training a computer classification system, which may be defined by a network comprising a fixed number of n-tuples or look-up tables (LUTs) comprising a fixed number of columns addressed by signals, at least a portion of the cells comprises: One or more training sets of example input data for different classes in such a way as to include or point to information based on the number of times corresponding cell addresses are sampled from one or more sets of training input examples. Determining column vector cell values based on the Determining the weight cell value for the column vector cell or cells being addressed or sampled, thus also allowing the weighting of the positive value column cell or cells during the classification process. Wherein the weighted cell value is determined based on at least a portion of the determined column vector cell value and by using at least a portion of a training set of input examples. 2. Elements of the sampled training input data each defined by a vector having a number of rows each corresponding to at least one subset of possible classes and each having a cell with a value. Alternatively, in a method of training a computer classification system, which may be defined by a network comprising a fixed number of n-tuples or look-up tables (LUTs) comprising a fixed number of columns addressed by signals, at least a portion of the cells comprises: One or more training sets of example input data for different classes in such a way as to include or point to information based on the number of times corresponding cell addresses are sampled from one or more sets of training input examples. Determining a column vector cell value based on the at least Determining a weight cell value corresponding to a subset of the vector cells, thus enabling the boosting of one or more column vector cells during the classification process, wherein the weight cell value is determined by the determined column vector A training method determined based on information of at least a portion of the cell values and by using at least a portion of a training set of input examples. 3. The method of claim 2, wherein a weight cell value is determined during the classification process to enable suppression of one or more column vector cells. 4. The weight cell value determination allows weighting of one or more column vector cells having a positive value (greater than zero) and one or more column vectors having a non-positive value (0 or less). The method according to claim 1, wherein 5. The method according to claim 1, wherein the weight cells are arranged in a weight vector corresponding to at least a part of the column vector. 6. Elements of the example training input data each comprising a fixed number of rows corresponding to at least one subset of the possible classes, at least in part having corresponding weight vectors, and each being sampled. Or includes a fixed number of column vectors addressed by the signal, each column vector and weight vector with a value determined based on one or more training sets of example input data for different classes. A method for determining a weight cell value in a computer classification system that may be defined by a network including a fixed number of n-tuples or look-up tables (LUTs) having cells,
Determine column vector cell values based on the training set of input examples such that at least a portion of the values include or point to information based on the number of times the corresponding set address is sampled from the training input example set. Determining a weighted vector cell value corresponding to one or more column vector cells based on information of at least a portion of the determined column vector cell values and using at least a portion of the training set of input examples. A method wherein the determination allows weighting of column vector cells having positive values (greater than zero) and column vector cells having non-positive values (less than or equal to zero). 7. The method according to claim 1, wherein the determination of the weight cells allows weighting of any column vector cells. 8. The method for determining weight cell values, wherein the weight cells are arranged in the form of weight vectors and initializing one or more sets of weight vectors corresponding to at least a portion of the column vectors, and 8. A method as claimed in any preceding claim, comprising adjusting a weight vector cell value of at least a portion of the weight vector based on information of at least a portion of the value and by using at least a portion of the training set of the input example. The method described in the section. 9. The method of claim 9, wherein at least a portion of the column cell values is determined as a function of the number of times a corresponding cell address is sampled from the set of training input examples.
A method according to any one of the preceding claims. 10. The maximum column vector value is one, but at least a portion of the values have an associated value that is a function of the number of times the corresponding cell address is sampled from the training set of the input example. The method according to any one of claims 1 to 9. 11. The method according to claim 8, wherein the column vector cell value is determined before adjusting the weight vector cell value and stored in the storage means. 12. Determine column vector cell values by: a) applying a known class of training input data examples to a classification network;
Thus addressing one or more column vectors; b) incrementing, preferably by one, the value or vote of the cell of the addressed column vector corresponding to a known class of rows; and c) Until the training example is applied to the network, (a) to (b)
12. The method according to any one of the preceding claims, comprising a training step of repeating the steps of. 13. All column vectors have corresponding weight vectors.
A method according to any one of claims 5 to 12. 14. The method of claim 8, wherein initializing the weight vector includes setting all weight vector cell values to a predetermined constant value, wherein the predetermined value is preferably one. Or the method of claim 1. 15. The method of claim 8, wherein initializing the weight vector includes setting each weight vector cell to a predetermined specific cell value.
The method described in the section. 16. Adjusting the weight vector cell value includes determining a global quality value based on at least a portion of the weight and the column vector cell value, determining whether the global quality value meets a required quality criterion. The method of any of claims 8 to 15, comprising the steps of: adjusting at least a portion of the weight cell value until a global quality criterion is met. 17. Adjusting the weight cell values: a) selecting input data from a training set; b) determining a local quality value corresponding to the sampled training input. The local quality value is a function of at least a portion of the addressed weights and column vector cell values; c) determining if the local quality value meets the required local quality criteria, and if not Adjusting one or more of the addressed weight vector cell values if the local quality criterion is not met; c) selecting a new input example from a predetermined number of examples of the training set D) Local quality test steps (b)-(d) for all given training input examples
E) determining a global quality value based on at least a portion of the column vector and the weight being addressed during the local quality test; f) if the global quality value meets the required global quality criteria 16. The method according to any one of claims 8 to 15, comprising determining whether or not there is, and h) repeating steps (a)-(g) until a global quality criterion is met. 18. The method of claim 17, wherein steps (b)-(d) are performed for all examples of the training set. 19. The method according to claim 16, wherein global and / or local quality criteria are changed during the adjustment iteration process. 20. The method according to claim 16, wherein the adjusted iteration process is stopped if the global quality criterion is not met after a given number of iterations.
The method described in the section. 21. The adjusted weight cell value is stored after each adjustment, and after the determination of the global quality value, the determined global quality value is further stored in a weight cell value or configuration previously stored separately. 21. The method according to any of claims 16 to 20, wherein the step of separately storing the weight cell values or the classification system configuration values thus obtained is closer to meeting the global quality criteria than the global quality values corresponding to the values. Or the method of claim 1. 22. A method for classifying an example of input data into at least one of a plurality of classes using a computer classification system configured according to the method of any one of claims 1 to 21. The column vector cell values and corresponding weight vector cell values are determined for each n-tuple or LUT based on one or more training sets of example input data, comprising: a) input data to be classified; Applying the example to the constructed classification network, thus addressing the column vectors and the corresponding weight vectors in the n-tuple or LUT, b) selecting one class and thus the set of n-tuples or LUTs Addressing a particular location in c) the output value as a function of the value of the addressed weight cell Determining step d) steps (b)-(c) until the output has been determined for all classes
E) comparing the calculated output values; and f) selecting the class or classes with the highest output values. 23. The method of claim 22, wherein the output value is further determined as a function of the value of the addressed column cell. 24. The output value is determined as a first sum of all addressed weight vector cell values corresponding to column vector cell values greater than or equal to a predetermined value, wherein the predetermined value is preferably one. 24. The method of claim 23, wherein: 25. The step of determining an output value comprises: determining a first sum of all addressed weight vector cell values corresponding to column vector cell values greater than or equal to a predetermined value. 24. The method of claim 23, comprising determining a second sum of designated weight vector cell values and determining an output value by dividing the first sum by the second sum. . 26. Elements of the sampled training input data each defined by a vector having a number of rows corresponding to at least one subset of the possible classes and each having a cell with a value. Or training a known class in a training system of a computer classification system, which can be defined by a network comprising a fixed number of n-tuples or look-up tables (LUTs), comprising a fixed number of columns addressed by signals. Means for receiving an example of input data for use, sample the received example of input data, and
Means for addressing a column vector within a stored set of tuples or LUTs, means for addressing a particular row corresponding to a known class within a set of n-tuples or LUTs, storage means for storing an n-tuple or LUT, means for determining a column vector cell value to include or point to information based on the number of times a corresponding cell address is sampled from a training input example set; Determining the weight cell value corresponding to the column vector cell or cells being addressed and sampled by the training example, thus allowing the weighting of the positive value of the column vector cell or cells during the classification process Means for determining the weighted cell value of the column vector section value. Kutomo it based on a portion of the information, and is determined by using at least a portion of the training set of input examples, the training system. 27. Elements of the sample training input data each defined and defined by a vector having cells with values, each comprising a fixed number of rows corresponding to at least one subset of possible classes. Or training a known class in a computer classification system training system, which can be defined by a network comprising a fixed number of n-tuples or look-up tables (LUTs), comprising a fixed number of colors addressed by signals. Means for receiving an example input data, means for sampling the received example input data and addressing a column vector within a stored set of n-tuples or LUTs, a set of n-tuples or LUTs To address a specific row within a class that corresponds to a known class Means for storing the determined n-tuple or LUT, column column cell values to include or point to information based on the number of times corresponding cell addresses are sampled from the training input example set. And means for determining a weight cell value corresponding to at least a subset of the column vector cells and thus enabling boosting of one or more column vector cells during the classification process. A training system wherein cell values are determined based on information of at least a portion of the determined column vector values and by using at least a portion of a training set of input examples. 28. The method according to claim 27, wherein the means for determining weight cell values is adapted to determine these values to allow suppression of one or more column vector cells during the classification process. The described system. 29. The means for determining a weight cell value comprises: one or more column vector cells having a positive value (greater than zero) and one or more column vector cells having a non-positive value (less than or equal to 0) 29. The system according to any one of claims 26 to 28, adapted to determine these values in order to allow a weighting of. 30. The means for determining a weight cell value, wherein the weight cell value is arranged in the form of a weight vector corresponding to at least a portion of the column vector.
30. The system according to any one of claims 26 to 29, adapted to determine these values. 31. Example training input data each including a fixed number of rows corresponding to at least one subset of the number of possible classes, at least in part having corresponding weight vectors, and each being sampled. And a respective column vector and weight vector are being determined during the training process based on one or more sets of example training input data. Example training input data of a known class in a system for determining a weighted cell value of a classification network defined by a stored number of n-tuples or look-up tables (LUTs) having cell values Input means for receiving the input data, sample the received input data, and perform n-tuple or Means for addressing column vectors and corresponding weight vectors in a stored set of LUTs, n-tuples or LUTs.
Means for addressing a particular row corresponding to a known class, storage means for storing the determined n-tuple or LUT, and corresponding cell addresses sampled from the training input example set. Means for determining a column vector cell value to include or point to information based on the number of times
Determining a weight vector cell value corresponding to one or more column vector cells based on at least a portion of the determined column vector cell values and using at least a portion of a training set of input examples. Means, wherein said determination allows weighting of one or more column vector cells having a positive value (greater than zero) and one or more column vector cells having a non-positive value (less than or equal to zero). You, the system. 32. The system according to any one of claims 26 to 31, wherein the means for determining a weight cell is adapted to allow weighting of any column vector cell. 33. The means for determining a weight cell value includes means for initializing one or more weight vector sets corresponding to at least a portion of a column vector, and at least one of the determined column vector cell values. 33. A method according to any one of claims 26 to 32, comprising means for adjusting a weight vector cell value of at least a portion of the weight vector based on the portion of information and by using at least a portion of the training input example set. The system according to paragraph. 34. The means for determining column vector cell values is adapted to determine these values as a function of the number of times corresponding cell addresses are sampled from the training input example set. The system according to any one of claims 26 to 33. 35. The means for determining a column vector cell value is such that at least a portion of the maximum one values are a function of the number of times the corresponding cell address is sampled from the training input example set. 34. A system according to any one of claims 26 to 33, adapted to determine these values in such a way as to have certain associated values. 36. An example of training input data belonging to a known class is applied to a classification network, such that at the point of addressing one or more column vectors, the means for determining a column vector cell value comprises a known 36. A method as claimed in any one of claims 26 to 35, wherein the address corresponding to a row of a class is adapted to increment a value or vote of a cell of a specified column vector, the value preferably being incremented by one. System. 37. All column vectors have a corresponding weight vector.
The system according to any one of claims 26 to 36. 38. The means for initializing a weight vector is adapted to set all weight vector cell values to a predetermined constant value, wherein said predetermined value is preferably one. The system according to any one of claims 33 to 37. 39. The means for initializing a weight vector is adapted to set each weight vector cell to a predetermined particular value.
The system according to claim 1. 40. The means for adjusting a weight vector cell value determines a global quality value based on at least a portion of the weight and the column vector cell value, and determines whether the global quality value meets a required global quality criterion. 40. The system of any one of claims 33-39, wherein the system is adapted to determine whether or not and adjust at least a portion of the weight cell value until a global quality criterion is met. 41. The means for adjusting the weight vector cell value comprises: a) a local quality value corresponding to the sampled training input and being a function of at least a portion of the addressed weight and column vector cell values. B) determining whether the local quality value satisfies the required local quality criterion; and c) determining which of the weight vector cell values was addressed if the local quality criterion was not met. D) iterating the local quality test for a predetermined number of training input examples, d) based on at least a portion of the weights and column vectors being addressed during the local quality test. Determining global quality values; e) determining whether the global quality values meet the required global quality criteria; and g) determining local and large values until the global quality criteria are met. Quality test and associated is adapted to repeat the weight adjustment system according to any one of claims 33 to 39. 42. The means for adjusting the weight vector cell value is adapted to stop the iterative process if the global quality criterion is not met after a given number of iterations. 42. The system according to 40 or 41. 43. The means for storing an n-tuple or LUT includes means for storing an adjusted weight cell value and separate means for storing a highest weight cell value so far. And the means for adjusting the weight vector cell value further comprises the determined global quality value satisfies a global quality criterion that is higher than the global quality value corresponding to the highest weight value previously stored separately. 42. The system of claim 40 or 41, wherein the system is adapted to replace an adjusted weight cell value that previously obtained the highest weight cell value previously stored separately when it is near. 44. A system for classifying an example input data into at least one of a plurality of classes, each comprising a number of rows corresponding to at least one subset of a number of possible classes, A storage means for storing a fixed number of n-tuples or look-up tables (LUTS) or sets comprising a fixed number of column vectors with corresponding weight vectors, wherein each column vector is sampled input data. A system addressed by an example element or signal, wherein each column vector and weight vector has a cell having a value determined during a training process based on one or more example training input data sets. Input means for receiving an example of input data to be classified,
Means for sampling the received example input data and addressing the empty and corresponding weight vectors in a stored n-tuple or LUT set, n-tuple or LUT for a particular row corresponding to a particular class Means for addressing in the set, means for determining the output value as a function of the addressed weight cell, and a singular with maximum output, comparing the calculated output values corresponding to all classes Or a system further comprising means for selecting a plurality of classes. 45. The system of claim 44, wherein the output value is further determined as a function of the value of the addressed column cell. 46. The output determining means includes means for generating a first sum of all addressed weight vector cell values corresponding to column vector cell values greater than or equal to a predetermined value and corresponding to a particular class. 46. The system according to claim 44 or claim 45, comprising: 47. An output determining means comprising: means for generating a second sum of all addressed weight vector cell values corresponding to a particular class; and
47. The system of claim 46, further comprising means for determining an output value by dividing the sum of the second by the second sum. 48. The cell value of the column vector and the weight vector, wherein
48. The system of any one of claims 44-47, wherein the system has been determined by using a training system of any one of -43. 49. The cell value of the column vector and the weight vector,
21. The method according to any one of claims 21 to 24, wherein the determination has been made during the training process.
8. The system according to claim 7.