KR101067376B1 - method for associative information processing using multisets and the associative memory device therof - Google Patents

Abstract

According to the present invention, data is continuously introduced into a stream by learning to code data into clusters of various patterns using a multiset structure, which is a set concept allowing overlap, and to reconstruct memory through a combination of these patterns. The present invention relates to an associative information processing method using a multiset and a memory device capable of efficiently modeling a new pattern and generating an infinite virtual memory capacity using a finite physical memory capacity.

Associative information processing method using a multiset of the present invention comprises the steps of (a) receiving the training data; (B) extracting a plurality of relatively low-dimensional patterns from the learning data input in the step (a) compared to the learning data; (C) dividing the pattern extracted in the step (b) into a matchset which is an existing pattern and a newset which is a new pattern by matching a pattern previously stored in a memory; (D) updating a memory using the matchset and the newset; (E) generating one virtual data having the highest probability by predicting a probability using the associative interaction from the matchset and learning a memory using the virtual data generated in the step (e) (f) Step), wherein the pre-stored pattern and the pattern extracted in step (b) are composed of a multiset, which is an aggregation concept to allow duplication.

Association, Memory, Multiset, Learning Data, Pattern, Cluster Pattern Code, Virtual Data

Description

Method for associative information processing using multisets and the associative memory device therof}

According to the present invention, data is continuously introduced into a stream by learning to code data into clusters of various patterns using a multiset structure, which is a set concept allowing overlap, and to reconstruct memory through a combination of these patterns. The present invention relates to an associative information processing method using a multiset and a memory device capable of efficiently modeling a new pattern and generating an infinite virtual memory capacity using a finite physical memory capacity.

Associative memory is a content-based data storage technology that retrieves the entire data item by presenting a portion of the data to be searched for. So far, associative memory technology has assumed that the data set to be stored is fixed. However, these methods are not suitable for modeling ever-changing data. For example, in case of surveillance camera or mobile multimedia device, data is continuously generated and it is difficult to solve by existing method to model it in real time and reflect it to service immediately.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-mentioned problem, and uses a multiset structure, which is a set concept that allows duplication, to code data into clusters of various patterns and to reconstruct memory through a combination of these patterns. Method to process associative information using a multiset to efficiently model data continuously flowing in a stream form to generate a new pattern, and to calculate an infinite virtual memory capacity using a finite physical memory capacity. It is an object to provide the memory device.

Associative information processing method using a multiset of the present invention comprises the steps of (a) receiving the training data; (B) extracting a plurality of relatively low-dimensional patterns from the learning data input in the step (a) compared to the learning data; (C) dividing the pattern extracted in the step (b) into a matchset which is an existing pattern and a newset which is a new pattern by matching a pattern previously stored in a memory; (D) updating a memory using the matchset and the newset; (E) generating one virtual data having the highest probability by predicting a probability using the associative interaction from the matchset and learning a memory using the virtual data generated in the step (e) (f) Step), wherein the pre-stored pattern and the pattern extracted in step (b) are composed of a multiset, which is an aggregation concept to allow duplication.

In the above configuration, the extraction of the pattern in the step (a) is characterized in that it is made at random.

The matching in the step (c) is implemented through a set operation of two multisets, the previously stored pattern and the pattern extracted in the step (b).

The update in step (d) is characterized in that the matchset increases its weight to enhance memory, and the newset includes an addition to the memory.

The updating in the step (d) includes the removal of the matchset previously existing in the memory, characterized in that the removal is made in consideration of the age and weight of the matchset.

The generation of the virtual data in the step (e) is characterized in that the base rule.

The learning in the step (f) is characterized in that the memory is reconfigured by comparing the virtual data and the learning data, deleting the memory elements contributing to the incorrectly generated portion of the virtual data and duplicating the memory elements contributing to the correct generation. .

On the other hand, the associative memory device using the multiset of the present invention comprises: an extraction unit for extracting a plurality of relatively low-dimensional pattern from the input learning data; A match unit for matching the extracted pattern with a pattern previously stored in a memory and classifying the matched pattern into an existing pattern and a new set as a new pattern; A storage unit to update a memory using the matchset and the newset; And a learning unit for generating one virtual data having the highest probability by predicting a probability using the associative action from the match set, and a learning unit for learning a memory using the virtual data. And the extracted pattern consists of a multiset, which is a set concept that allows duplication.

In the above configuration, the extraction of the pattern is characterized in that it is made at random.

The match may be implemented through a set operation of two multisets, the previously stored pattern and the extracted pattern.

The storage unit may include a reinforcing unit for increasing memory by increasing the weight of the matchset and an adding unit for adding the newset to the memory.

The storage unit may further include a remover which removes the matchset existing in the memory in consideration of its age and weight.

The virtual data is generated based on a base rule.

The learning unit compares the virtual data with the learning data; And a deletion unit for deleting a memory element contributing to the wrongly generated portion of the virtual data as a result of the comparison, and a copying unit for copying the memory element contributing to the correct generation of the virtual data as a result of the comparison.

According to the associative information processing method using the multiset of the present invention and its memory device, the following operations / effects are exhibited.

The ability to form patterns allows the creation of new patterns similar to observed data.

-The noise removal effect can remove the noise included in the input data to restore the clean original image.

Priming effects can be reflected, allowing them to adapt quickly to changing environments and predict new conditions.

-Efficiently store large amounts of changing data and retrieve it based on content.

Similar to a person's memory ability, very old memories can be recalled in real time.

-The concept of population coding is used to accommodate short-term changes while maintaining long-term stability, which is useful for online learning.

-By learning and changing only the parts that are changed by sparse coding, the memory utilization efficiency can be maximized.

-By reducing the spurious states through glocal coding method, there is an effect of improving information storage than the conventional associative memory.

Hereinafter, an associative information processing method and a memory device using the multiset according to the present invention will be described in detail with reference to the accompanying drawings.

First, the present invention efficiently expresses the training data set D on a multiset pattern memory M having a multiset data structure, and the original training data by recombination of various patterns from the multiset pattern memory M. Reconstruct set D. This method is especially useful for learning data Efficient storage and retrieval is possible even if set D changes, typically | M | << | D |, that is, it is effective when the memory capacity is extremely small compared to the total capacity of the training data.

FIG. 1 is a diagram illustrating a configuration of a multiset pattern memory in the associative information processing method using the multiset of the present invention, in which 20 patterns of N training data are extracted and stored in the multiset pattern memory M, respectively. To illustrate. In addition, it is assumed that four different types of patterns are duplicated for one training data (the number of duplicate samplings is weighted w). In general, the same pattern may be duplicated among training data. Cases become more frequent. 1 illustrates a case where there is no overlap in order to explain the pattern extraction method, and the pattern generation by the associative action uses the overlap pattern.

The content of the multiset pattern memory of the present invention (hereinafter also referred to simply as 'pattern memory') M = { m _k } may be represented by an incidence matrix as shown in FIG. 1. The pattern constituting one memory element is shown in row m _k , and if there is a value of 0 or 1 shown in the row, that variable x _i is included in this pattern. Variables left blank will not be included in this pattern.

The pattern memory M of the present invention has the following characteristics.

One. | m _k | << N. In other words, a sparse coding scheme in which the number of dimensions of the pattern is much smaller than the number of dimensions N of the data is used.

2. | m _i | ≠ | m _j | for i ≠ j. That is, variable codewords having various lengths of patterns stored in memory are used.

3. The distributed code is used by randomly extracting n low-dimensional patterns from one learning data.

4. Allow m _i = m _j for i ≠ j . That is, it has a multiset structure in which the same pattern can be stored repeatedly.

5. | M | >> N. That is, the number of patterns stored in memory is much larger than the number of data dimensions.

6. | M | << | D |. That is, the memory capacity (the number of patterns to be stored) is significantly smaller than the amount of data (the number of learning data), so that the compression ratio is large.

For example, consider storing 10000 image data consisting of 10000 pixels 100 x 100. In this case N = 10,000 and | D | = 10,000. If you store this data as a normal image, you need 10,000 x 10,000 = 10 ⁸ or 100 Mega memory elements. When storing this data in multiset associative memory, for example, 1 million (= 10 ⁶ ) memory elements m _i of fixed size k = 100 are used (that is, | M | = 10 ⁶ ) for a total of 10 ⁸ It can be represented using memory elements. As can be seen from the results of backward experiments, when stored in associative memory with sparse code, the detailed information of the original image is removed, and instead, the information on the common parts of the images is maintained to express the characteristics of the entire data. There is a characteristic.

Associative memory storage is particularly advantageous when the number of observed data is very large or changes frequently. If the number of images observed in this example is 10 images per second, this is 600 images per minute, 36000 images per hour, and 24 x 36000 = 864,000 images per day (about 0.86 Mega). If we wanted to store all the data we observed in a day, the total capacity would require about 10 ⁴ x 0.86 Mega = 8.6 Giga memory elements. In this case, if you are not just storing the data, but analyzing it in real time, this is not practical. In this case, the problem can be solved by utilizing multiset associative memory. In other words, for this large problem, as in the previous small problem, using 1 million memory elements m _i (= 10 ⁶ ) consisting of low-dimensional data of fixed size k = 100 , | M | If you configure associative memory with 10 ⁸ memory elements, or 100 Mega elements, you can analyze image data online. Of course, the degree of problem solving effect may vary depending on the difficulty level of the specific problem. However, as shown in the experimental analysis below, since many data structures exist and the observed data are redundant, information loss is not significant even when stored in associative memory.

In addition to the fact that the amount of observation data is so large that it is difficult to store, the multiset associative memory is well suited when the data is constantly changing and the model must be adapted online. This is because learning can be performed quickly only by extracting a sample from observation data and storing it in memory during learning. This replacement of some memory elements does not significantly affect the global solution because learning is based on cluster code.

2 is an overall block diagram of an associative memory device using the multiset according to the present invention. The multiset pattern memory M, the extractor 20, the matcher 30, the storage 40, and the generator 50 are shown in FIG. And a learning unit 60.

First, the extractor 20 extracts n k- dimensional patterns from the input learning data x (sampling unit) 22 to form a cluster pattern code S (coding unit) 24.

The match unit 30 matches the group pattern code S thus configured with the stored pattern memory M (matching unit; 32) to form a match set M '= S ∩ M and a new set S' = S-M '(selection unit; 34). )do. Here, the matchset M 'is a collection of patterns extracted from the training data x that are already stored in the memory and the newset S' is a collection of new patterns not previously observed.

Next, the storage unit 40 updates the pattern memory M using the matchset M 'and the newset S'. Specifically, since the pattern in matchset M 'is in the existing pattern memory M, the weight is increased to increase the weight (consolidation) (42), and since the pattern in newset S' is a new pattern, a new pattern memory M Add (addition); However, in this case, since the memory capacity may be insufficient, the existing pattern is deleted and replaced (removed) 46 (using the deletion strategy described later).

Next, the generation unit 50 predicts the probability using the associative action from the match set M '(prediction unit; 52) to synthesize one virtual data y having the highest probability (synthesis unit) 52 (specific method). Where the virtual data y has the same dimension and shape as the training data x .

Finally, the learning section 60 includes a learning data x and the virtual data y compared to the next virtual data y is the to calibrate the memory pattern in the direction in which more similarly generated in the learning data x, a specifically generated virtual data y and the original the By comparing (comparing with; 62) the training data x , the memory elements contributing to the correctly generated part of the virtual data y are duplicated and reinforced (duplicating part 66), and the memory elements contributing to generating the wrong part are deleted. The pattern memory M is reconfigured (deleting section 64) to reduce it.

FIG. 3 is a flowchart for explaining in detail the associative information processing method using the multiset of the present invention together with the processing of each part of the associative memory device shown in FIG.

First, the extraction section; the (20, step S20) to extract a random pattern of low-order from the given learning data x sampling (step S22) and, S = from a learning data x to repeat the sampling process n times again { s ₁ , s ₂ ,. , cluster pattern code S composed of a set of n patterns of s _n } is constructed (step S24). Here, it is assumed that the input training data x takes a vector form of binary variable values. However, this method can be applied to any variable value. If the actual data is not in binary form, it is also possible to convert it to this form via preprocessing, which is assumed below.

4 is a road standing, N = 64-dimensional number of image data (learning data) x from k = 7-D of a variable for explaining a method for extracting patterns from the training data in the image of the information processing method using a multi-set of the invention The process of randomly extracting values to construct a pattern s of order 7 is illustrated.

The cluster pattern code S is a population code composed of several code words . This code also has a small portion of the dimension N of data, a sparse code composed of k- dimensional patterns, i.e. k << N.

If the training data vector x is composed of N dimensions, the total number of samples that can be generated therefrom is 2 ^N equal to the number of subsets. Applying the example of the image data above, N = 10,000 and the number of all samples that can be generated is a large number of 2 ¹⁰⁰⁰⁰ > 10 ¹² . If you set the sample pattern dimension to k = 100, then ₁₀₀₀₀ C ₁₀₀ is also a very large number. However, if the dimensions are extracted without overlap, then 100 samples with k = 100 dimensions can be extracted. Therefore, the sample number n can be determined based on the following equations (1) and (2).

[Equation 1]

[Equation 2]

However, considering the loss of information due to low-dimensional sampling, the number of samples should be greater than this N. Therefore, let cK but c is determined according to the amount of memory space available for the pattern memory M.

Next, the match unit 30 (step S30) matches the contents of the group pattern code S with the contents of the pattern memory M (step S32) to select and match the matchset M 'and the newset S' (step S34). Here, S = { s _i } and M = { m _j } are represented as multisets, so a match can be implemented through a set operation. However, because it is a multiset, multiple patterns of the same element can be overlapped, which is represented in the pattern memory M with a weight w. The results found by the match are divided into two, matchset M 'and newset S', as shown in Equations 3 and 4 below.

&Quot; (3) &quot;

M '= M ∩ S

&Quot; (4) &quot;

S '= S-M'

The two sets of Equations 3 and 4 above are separate sets having no common part, and thus satisfy Equation 5 below, whereby Equation 6 below is derived.

[Equation 5]

S = S '∪ M'

&Quot; (6) &quot;

S '∩ M' = {}

Some errors can be tolerated during the match process. In other words, rather than assuming a perfect match between the two patterns, it is possible to calculate the similarity sim ( m _j , s _i ) and see that the match is greater than a certain threshold θ _sim . In this case, the match set is represented by Equation 7 below.

[Equation 7]

For example, the following Equation 8 may be used as a unit of measure of similarity.

[Equation 8]

, 단 σ²은 데이터의 분산값

Where σ ² is the variance of the data

Next, the storage unit 40 (step S40) strengthens by increasing the weight of the memory pattern in the match set M '(step S42), and removes the old pattern inside the pattern memory M (step S44) if necessary, The patterns of the new set S 'are added to the pattern memory M (step S46). There are two ways of changing the weight of the existing memory patterns as shown in Equations 9 and 10 below.

[Equation 9]

[Equation 10]

First, Equation 9 is a method of adding weights, and Equation 10 is a method of multiplying the current weight by some percentage. In the former case, the stable learning is possible, but the learning may be slow. In the latter case, the important pattern may be quickly reflected, but there is a concern that certain patterns may dominate the characteristics of the entire pattern.

On the other hand, when the pattern in the newset S 'is stored in the pattern memory M, a case where the capacity of the memory is already filled can no longer accommodate a new pattern may occur. This can happen frequently if new data is constantly being entered or the environment is changing after observing a lot of training data (but if many training data have been observed but similar data continues to flow in, this is not the case). In this case, you need to decide which existing patterns to delete or replace. To do this, specify a time stamp for every storage pattern and specify the time t _k when the pattern is newly stored. This time continues to age and considers whether this age is greater than the threshold θ _age when looking for patterns to replace later. However, if only age is considered, important patterns can be deleted. Therefore, delete only when the weight w _k is below a certain value θ _weight as follows. The replacement strategy may be performed by selecting a candidate set R to be replaced as shown in Equation 11 below and deleting the oldest or least weighted pattern among them.

[Equation 11]

In the associative memory method of the present invention, it is necessary to note that there is an automatic generalization, that is, a learning effect only through the process of extracting and storing a pattern from the learning data as described above. That is, learning happens in two parts

1. By updating the contents of the pattern memory with a new sample, the weight value of the pattern stored in the pattern memory is corrected (learning effect by memory enhancement).

2. Spontaneous generalization occurs because the dimension of the stored sample is lower than the original learning data (learning effect by reducing the number of dimensions).

In particular, the spontaneous generalization feature of paragraph 2 has many advantages for learning in associative memory devices. In this case, it is possible to track and adapt the data characteristics recently by simply extracting a low-dimensional sample from the observed data and updating the memory contents without performing the learning unit that requires more time.

Next, the generation unit 50 (step S50) evaluates the associative memory capability by generating the virtual data y , and for this purpose, the most probable one is predicted by using the associative action from the matchset M '(step S52). One virtual data y is synthesized (step S54). In the above-described storage unit 40 (step S40), the storage unit 40 (step S40) is added as it is or the weight is updated without considering the interaction of the contents previously stored in the pattern memory M. However, in the associative memory, the retrieved data may serve as a new cue to extract the next memory trace.

In order to increase the associative memory capability of associative memory, the interaction between different data variables is considered by repeating the multiple retrieval process instead of performing the single memory retrieval for a given learning data x . This process can be seen as a probabilistic reasoning process, and the probability of new data, ie, virtual data y , to be generated for a given clue, learning data x , can be described by the Bayes' rule as shown in Equation 12 below. Can be.

[Equation 12]

The problem is to find the most probable virtual data y for a given training data x , which can be optimized using the equation (13) below, taking advantage of the fact that the multiset associative memory M stores a large number of different patterns. Can be.

[Equation 13]

In Equation 13 above, the first identity is taken into account that P ( x ) has no effect on maximizing the above value, and in the second identity, P ( y ) is assumed to be equal. The third identity uses the ability to predict the value of y using a large number of m _k stored in memory. This equation can be decomposed into a product of probabilities as shown in Equation 14 below.

[Equation 14]

In Equation 14 above, P ( x | m _k ) represents a relation between a given learning data x and a pattern stored in the pattern memory M, and P ( m _k | y ) represents the stored pattern m _k and the newly generated virtual data y Indicates an association of. The associative memory structure of the present invention can synthesize new virtual data y using the characteristics of the multiset storage structure in the process of calculating the above values.

Finally, the learning unit 60 (step S60) compares the generated virtual data y with the learning data x (step S62) to delete the memory elements contributing to the incorrectly generated portion of the virtual data x (step S64) and to correct generation. By modifying the contributing memory elements (step S66), M is modified to generate virtual data y more similar to the training data x .

Equation 15 below compares the virtual data y with the training data x to determine whether each memory element value is correct.

[Equation 15]

Meanwhile, the memory element that caused the error is removed, and the memory element that contributes to the correct data generation is added. This learning process can be summarized as in Equation 16 below.

[Equation 16]

M _pos is the set of m _k that contributed to the creation of the correct virtual data y , and M _neg is the set of m _k that contributed to the generation of the wrong virtual data y . As shown by the set symbol in the above equation, memory calibration can be implemented by performing parallel set operation.

FIG. 5 is a diagram illustrating the entire process of configuring a memory by extracting a pattern from training data in the method of processing association information using the multiset of the present invention as described above, and illustrates simple 64-bit numeric image data as training data x . Doing.

Hereinafter, various experimental examples for explaining the usefulness of the associative information processing method using the multiset of the present invention will be described.

Experimental Example 1

FIG. 6 is a diagram experimentally showing association results of arbitrary numerical image data according to the associative information processing method using the multiset of the present invention. FIG. For this experiment, 3000 numerical images were collected, each consisting of 16 x 16 = 256 binary digits. Among them, 6 and 7 images were selected and learned by the multiset association information processing method. As can be seen in FIG. 6, the result of processing by the method of the present invention on an image in which original image data, which is learning data, is arbitrarily damaged by 50% and 80%, respectively, is shown on the right. As described above, a restored image having a shape similar to the original image was obtained.

Experimental Example 2)

7 is a diagram showing experimental results of associative results with respect to face image data according to the associative information processing method using the multiset according to the present invention. The Yale data set consists of 165 face images taken of 11 pictures of 15 faces each. In this experiment, we reduced the face image to 60 x 45 = 2700 pixels and constructed a multiset associative memory using the training data for two faces. The total number of training images is 22. N = 20 low-dimensional patterns of k = 50 were sampled and stored in the pattern memory for each.

As can be seen in FIG. 7, the result of processing by the method of the present invention on an image in which original image data, which is training data, is arbitrarily damaged by 60% and 80%, respectively, is shown on the right. As described above, a restored image having a shape similar to the original image was obtained.

Experimental Example 3)

8 is an experimental diagram showing the noise removal effect according to the associative information processing method using the multiset of the present invention. As shown in FIG. 8, the image of the number 6 on the image of the random number or the image of the number 7 is randomly modified so that, for example, 30% is mixed on the image of the number 6 by modifying the original image which is the training data. As a result of processing by the method of the present invention on an image made to have a dot made to be mixed at 30%, for example, as shown on the right, a restored image having a shape similar to the original image was obtained.

Experimental Example 4)

FIG. 9 is an experimental diagram showing priming effects according to the associative information processing method using the multiset of the present invention. FIG. 9 shows a low dimensional pattern of k = 5 for each of numerical image data as used in Experimental Example 1. FIG. Associative memory was constructed by sampling N = 10 pieces. In the early stage of the configuration, since the information occupied by the information about the numbers 0 to 9 in the associative memory is similar, when the partial image of 6 characters is used as a stimulus, as shown in the far left of the bottom of FIG. The image reminds me of a mixture of used images. Afterwards, as we continued to repeat the learning of only 6 characters, we could observe a progressive improvement in the associative ability of the 6 characters.

Experimental Example 5)

FIG. 10 is an experimental diagram showing priming effects for arbitrary drama scene data according to the method for processing association information using the multiset of the present invention. Priming effects are more observed when dealing with continuous data such as movie scene data. Can be. The priming effect on the image capturing the scene of the drama Friends was observed in the same manner as in Experimental Example 4. Each scene in the drama was captured and converted into a black and white image of 80 x 60 pixels for use as data. The total number of training images is 10. For each of them, N = 100 low-dimensional patterns of k = 5 are sampled and stored in memory, and then the priming effect of repeatedly learning a specific scene is observed. As shown below, it can be observed that as the number of repetitive learning increases, the associative ability gradually improves.

Associative information processing method using the multiset of the present invention as described above The memory device can be used in the following fields.

-A device that stores video data such as IPTV and searches contents based quickly.

-A device attached to a digital camera that automatically retrieves past shots similar to the scene captured by the current camera.

-An apparatus that detects abnormal state by storing large-scale image data attached to the security surveillance camera device and searching at high speed based on the content.

A device for retrieving similar past events in real time from current sensor data in a lifelogging system.

-Noise reduction device of image data.

Associative information processing method using the multiset of the present invention The memory device is not limited to the above-described embodiments, and can be modified in various ways within the scope of the technical idea of the present invention.

1 is an exemplary configuration diagram of a multiset pattern memory in the associative information processing method using the multiset according to the present invention;

2 is an overall block diagram of an associative memory device using a multiset according to the present invention;

FIG. 3 is a flowchart illustrating a method of processing associative information using a multiset according to the present invention together with a process of each unit of the associative memory device shown in FIG. 1;

4 is a view for explaining a method of extracting a pattern from training data in the associative information processing method using the multiset of the present invention;

5 is a view for explaining a process of configuring a memory by extracting a pattern from the training data in the associative information processing method using the multiset of the present invention;

6 is an experimental diagram showing an associative result for arbitrary numerical image data according to the associative information processing method using the multiset of the present invention;

7 is an experimental view showing association results for face image data according to the associative information processing method using the multiset of the present invention;

8 is an experimental view showing a noise removal effect according to the associative information processing method using the multiset of the present invention;

9 is an experiment showing the priming effect according to the associative information processing method using the multiset of the present invention;

FIG. 10 is a diagram experimentally showing a priming effect on arbitrary drawing scene data according to the associative information processing method using the multiset according to the present invention.

*** Explanation of symbols for the main parts of the drawing ***

M: multiset pattern memory, 20: extraction section,

22: sampling part, 24: coding part,

30: matching unit, 32: matching unit,

34: sorting unit, 40: storage unit,

42: reinforcement part, 44: removal part,

46: add, 50: generator,

52: prediction unit, 54: synthesis unit,

60: learning unit, 62: comparison unit,

64: deletion unit, 66: replication unit

Claims (15)

(A) receiving N-dimensional learning data; (B) extracting a plurality of patterns having a lower dimension (k, where N> k) relative to the learning data from the learning data input in the step (a); (C) dividing the pattern extracted in the step (b) into a matchset which is an existing pattern and a newset which is a new pattern by matching a pattern previously stored in a memory; (D) updating contents of a memory using the matchset and the newset; (E) generating one virtual data having the highest probability by predicting a probability using the associative action from the matchset; and (F) learning a memory using the virtual data generated in the step (e), The prestored pattern and the pattern extracted in the step (b) are associative information processing method using a multiset consisting of a multiset which is a set concept allowing redundancy. The method of claim 1, Associating information processing method using a multiset, characterized in that the extraction of the pattern in the step (a) is performed at random. The method of claim 2, And the matching in step (c) is implemented by a set operation of two multisets, the previously stored pattern and the pattern extracted in step (b). The method of claim 3, wherein And the matchset increases its weight to enhance the memory, and the newset includes an addition to the memory. The method of claim 4, wherein The updating in the step (d) includes the removal of the matchset that previously existed in the memory, wherein the removal is performed by considering the age and weight of the matchset. Way. The method of claim 5, The associative information processing method using the multiset, wherein the generation of the virtual data in the step (e) is performed based on a Bayes rule. The method of claim 6, The learning in the step (f) is characterized by reconstructing the memory by comparing the virtual data and the learning data to delete the memory element that contributed to the incorrectly generated portion of the virtual data and duplicate the memory element that contributed to the correct generation Associative information processing method using multiset. An extraction unit for extracting a plurality of relatively low dimensional (k, N> k) patterns from the input N-dimensional learning data; A match unit for matching the extracted pattern with a pattern previously stored in a memory and classifying the matched pattern into an existing pattern and a new set as a new pattern; A storage unit for updating contents of a memory by using the matchset and the newset; A generation unit for generating one virtual data having the highest probability by predicting a probability using the associative action from the matchset; It includes a learning unit for learning the memory using the virtual data, The prestored memory device and the extracted pattern are associative memory devices using a multiset, which is a set concept that allows duplication. The method of claim 8, The associative memory device using a multiset, characterized in that the extraction of the pattern is performed at random. The method of claim 9, The match is associative memory device using the multi-set, characterized in that implemented through the set operation of the two stored sets and the extracted pattern of the multi-set. 11. The method of claim 10, And the storage unit includes a reinforcing unit for enhancing memory by increasing the weight of the matchset and an adding unit for adding the newset to the memory. The method of claim 11, And the storage unit further includes a remover which removes the matchset existing in the memory in consideration of its age and weight. 13. The method of claim 12, The associative memory device using the multiset, wherein the virtual data is generated based on a Bayes rule. The method of claim 13, The learning unit compares the virtual data with the learning data; A deletion unit for deleting a memory element contributing to an incorrectly generated portion of the virtual data as a result of the comparison; And a copying unit for copying a memory element contributing to the correct generation of the virtual data as a result of the comparison. A computer-readable recording medium having a program for executing the method of any one of claims 1 to 7.

Images