JP7467979B2 - Classification system, information processing device, classification method, and program - Google Patents

Description

The present invention relates to a classification system, an information processing device, a classification method, and a program for causing a computer to execute the method, which extracts features and classifies signal sources based on time-frequency analysis.

Conventionally, various methods have been proposed for classifying acoustic signals output from multiple signal sources according to the signal sources (see, for example, Non-Patent Document 1). For example, a classification system for classifying acoustic signals from multiple signal sources performs time-frequency analysis processing on the time waveforms of the acoustic signals. The classification system then performs feature analysis processing using matrix operations on the matrix data obtained as a result of the analysis, and acquires the feature quantities contained in the time waveforms.

Keigo Kimura, Yuzuru Tanaka, and Mineichi Kudo, “A Fast Hierarchical Alternating Least Squares Algorithm for Orthogonal Nonnegative Matrix Factorization”, JMLR 39(2014), pp.129-141

However, conventional classification systems require a memory area capable of storing both the matrix data used when performing the matrix operation in the feature analysis process and the matrix data that is the result of the matrix operation. In particular, to improve classification capabilities, it is necessary to handle matrix data with a large data size, and a memory area with a correspondingly large capacity is required. If the capacity of the memory area is small, the matrix data required for the matrix operation cannot be stored in the memory area, and the feature analysis process is interrupted, making it difficult to properly perform the feature analysis process.

Therefore, there is a need for a classification system, information processing device, classification method, and program that can perform feature analysis processing appropriately even when the storage capacity is reduced.

The classification system according to the present invention is a classification system that performs feature extraction and class classification based on time-frequency analysis, and includes a time-frequency analysis processing means for converting a time waveform indicating the amplitude of a time series of a known signal whose signal source is known into data indicating a time series frequency and signal intensity, a feature analysis processing means for generating feature amounts and feature extraction parameters by performing a submatrix operation on the data, a learning processing means for estimating classification parameters from the feature amounts, and a learning unit having a storage means having a main storage area for storing processing and an auxiliary storage area for storing processing results and the data as matrices, and and a classification unit having a feature extraction processing means for calculating a feature amount of an unknown signal whose signal source is unknown, and a classification processing means for estimating a signal source of the unknown signal from the feature amount of the unknown signal by referring to the classification parameters , wherein the feature analysis processing means performs the submatrix operation, which includes reading out a plurality of submatrices of the matrix from the auxiliary storage area to the main storage area, calculating a product of two submatrices read out to the main storage area and storing the product in the auxiliary storage area, and further reading out the product and a submatrix of a matrix to be output from the auxiliary storage area, and sequentially calculating a sum to calculate a submatrix product .

The information processing device according to the present invention is an information processing device that performs feature extraction and class classification based on time-frequency analysis, and includes a time-frequency analysis processing means that converts a time waveform indicating the amplitude of a time series of a known signal whose signal source is known into data indicating a time series frequency and signal intensity, a feature analysis processing means that generates feature amounts and feature extraction parameters by performing a submatrix operation on the data, a learning processing means that estimates classification parameters from the feature amounts, a feature extraction processing means that refers to the feature extraction parameters and calculates feature amounts of an unknown signal whose signal source is unknown, and a feature extraction processing means that refers to the classification parameters and calculates feature amounts of the unknown signal. and a storage means having a main storage area for storing processing and an auxiliary storage area for storing processing results and the data as matrices , wherein the feature analysis processing means performs the submatrix operation, which includes reading out a plurality of submatrices of the matrix from the auxiliary storage area into the main storage area, calculating a product of two submatrices read out to the main storage area and storing the product in the auxiliary storage area, and further reading out the product and a submatrix of the matrix to be output from the auxiliary storage area and sequentially calculating a sum to calculate a submatrix product .

A classification method according to the present invention has a storage means having a main storage area for storing processing and an auxiliary storage area for storing data and processing results, and performs feature extraction and class classification based on time-frequency analysis, converting a time waveform indicating a time series amplitude of a known signal whose signal source is known into the data indicating a time series frequency and signal intensity, generating feature quantities and feature extraction parameters using a submatrix operation on the data converted into a matrix , estimating classification parameters from the feature quantities, referring to the feature extraction parameters, calculating feature quantities of an unknown signal whose signal source is unknown, referring to the classification parameters, and estimating the signal source of the unknown signal from the feature quantities , wherein the submatrix operation includes reading a plurality of submatrices of the matrix from the auxiliary storage area into the main storage area, calculating a product of the two submatrices read into the main storage area, storing the product in the auxiliary storage area, and further reading out the product and a submatrix of a matrix to be output from the auxiliary storage area, sequentially calculating a sum to calculate a submatrix product .

The program according to the present invention causes a computer that performs feature extraction and class classification based on time-frequency analysis to function as: time-frequency analysis processing means that converts a time waveform indicating a time-series amplitude of a known signal whose signal source is known into the data indicating a time-series frequency and signal intensity; feature analysis processing means that generates feature amounts and feature extraction parameters by using a submatrix operation on the matrix -arranged data; learning processing means that estimates classification parameters from the feature amounts; feature extraction processing means that refers to the feature extraction parameters and calculates feature amounts of an unknown signal whose signal source is unknown ; and classification processing means that refers to the classification parameters and estimates the signal source of the unknown signal from the feature amounts of the unknown signal, and the feature analysis processing means performs the submatrix operation, which includes the feature analysis processing means reading a plurality of submatrices of the matrix from the auxiliary storage area into the main storage area, calculating a product of the two submatrices read into the main storage area and storing the product in the auxiliary storage area, and further reading the product and a submatrix of a matrix to be output from the auxiliary storage area and sequentially calculating a sum to calculate a submatrix product .

According to the present invention, by using submatrix operations during feature analysis processing, the memory area used during the calculation can be made smaller than when matrix operations are used, so that feature analysis processing can be performed appropriately even when the memory area capacity is reduced.

1 is a block diagram showing an example of a configuration of a classification system according to a first embodiment. 2 is a hardware configuration diagram showing an example of the configuration of a learning unit in FIG. 1 . 1. FIG. 4 is a hardware configuration diagram showing another example of the configuration of the learning unit in FIG. FIG. 1 is a block diagram showing an example of a configuration of a conventional classification system. 4 is a schematic diagram showing an example of a result of time-frequency analysis processing by a time-frequency analysis processing means. FIG. 10 is a schematic diagram showing an example of a result of a feature analysis process by a feature analysis processing means. FIG. 11 is a schematic diagram showing an example of a result of a learning process by a learning processing means. FIG. FIG. 11 is a schematic diagram for explaining a matrix sum operation process. FIG. 11 is a schematic diagram for explaining a matrix multiplication operation process. FIG. 11 is a schematic diagram for explaining a first problem caused by a shortage of a main memory area. FIG. 11 is a schematic diagram for explaining a second problem caused by a shortage of the main memory area. FIG. 11 is a schematic diagram for explaining a submatrix. FIG. 11 is a schematic diagram for explaining a sum operation process using a submatrix. 14 is a block diagram showing an example of the configuration of a partial matrix sum calculation means in FIG. 13 . FIG. 13 is a schematic diagram showing an example of the state of a main memory area and an auxiliary memory area in a sum operation process using a submatrix. FIG. 11 is a schematic diagram for explaining a multiplication process using a submatrix. 17 is a block diagram showing an example of the configuration of a partial matrix product calculation means in FIG. 16 . FIG. 13 is a schematic diagram showing an example of the state of a main memory area and an auxiliary memory area in a multiplication process using a submatrix.

The following describes an embodiment of the present invention with reference to the drawings. The present invention is not limited to the following embodiment, and various modifications are possible without departing from the spirit of the present invention. In addition, in each drawing, the same reference numerals are used to denote the same or equivalent parts, and this is common throughout the entire specification.

Embodiment 1.
A classification system according to the first embodiment will be described. The configuration of the classification system as shown in FIG. 1 described below is realized by executing a feature analysis processing program loaded into an auxiliary storage device on a computer (such as a personal computer). This feature analysis processing program is stored in an information storage medium such as a CD-ROM (Compact Disc-Read Only Memory) or distributed via a network such as the Internet, and installed in the computer.

[Configuration of Classification System 1]
A classification system 1 according to the first embodiment will be described. FIG. 1 is a block diagram showing an example of the configuration of the classification system according to the first embodiment. As shown in FIG. 1, the classification system 1 according to the first embodiment includes a learning unit 10 and a classification unit 20. The learning unit 10 and the classification unit 20 are provided in separate devices, for example. However, this is not limited to this, and the learning unit 10 and the classification unit 20 may be provided in a single device such as an information processing device.

(Learning Unit 10)
The learning unit 10 has a time-frequency analysis processing means 11, a feature analysis processing means 12, a learning processing means 13, and a storage means 14. The learning unit 10 receives time waveforms L101-n (n=1, ..., N) of N acoustic signals recorded in the past. A time waveform is a waveform that indicates the amplitude of a time series of a signal to be classified. It is assumed that the time waveform L101-n of an acoustic signal is a signal emitted from one of C types of known signal sources. Here, the type of this known signal source is called a label, and the time waveform L101-n is represented as a signal from a signal source with label c (c indicates one of the types 1, ..., C).

The time-frequency analysis processing means 11 receives the time waveform L101-n, performs time-frequency analysis processing to convert the time waveform L101-n into data indicating a time series of frequency and signal intensity, and outputs the time-frequency analysis result as the processing result. A representative example of time-frequency analysis processing is, for example, a short-time Fourier transform. In this case, the time-frequency analysis result is vector data or matrix data.

The feature analysis processing means 12 receives the N time-frequency analysis results and performs feature analysis processing using a matrix operation. The feature analysis processing means 12 then generates features and labels for each time waveform L101-n to be supplied to the learning processing means 13, and feature extraction parameters to be supplied to the feature extraction processing means 22, and outputs the generated features, labels, and feature extraction parameters.

The input to the feature analysis processing means 12 is an arrangement of N time-frequency analysis results, which are matrix data or tensor data. Representative examples of feature analysis processing include nonnegative matrix factorization (NMF) and nonnegative tensor factorization (NTF). In the first embodiment, the feature analysis processing means 12 performs feature analysis processing by using submatrix operations as matrix operations. Details of feature analysis processing using submatrix operations will be described later.

The learning processing means 13 receives the features and the labels of each time waveform L101-n, estimates classification parameters to be supplied to the classification processing means 23 from the distribution of the features, and outputs the estimated classification parameters.

The storage means 14 has a main storage area 14a and an auxiliary storage area 14b. The main storage area 14a is an area for temporarily storing data required when performing various types of processing. For example, when the feature analysis processing means 12 performs a submatrix operation, matrix data and the like required for processing are read from the auxiliary storage area 14b and stored in the main storage area 14a. The auxiliary storage area 14b is an area for storing various types of data, and data is read and written as necessary. The auxiliary storage area 14b stores matrix data, for example, when the feature analysis processing means 12 performs a submatrix operation.

(Classification Unit 20)
The classification unit 20 has a time-frequency analysis processing means 21, a feature extraction processing means 22, and a classification processing means 23. A time waveform C101 is input to the classification unit 20. Unlike the time waveform L101-n, it is unknown from which of C types of signal sources the time waveform C101 is radiated. The classification unit 20 estimates the label of this unknown signal.

The time-frequency analysis processing means 21 has a configuration similar to that of the time-frequency analysis processing means 11, performs time-frequency analysis processing on the time waveform C101, and outputs the time-frequency analysis result as the processing result.

The feature extraction processing means 22 refers to the feature extraction parameters supplied from the feature analysis processing means 12 and calculates the feature amount from the time-frequency analysis result. The feature extraction processing means 22 then outputs the calculated feature amount.

The classification processing means 23 refers to the classification parameters supplied from the learning processing means 13 and estimates the label c of the unknown time waveform C101 from the feature quantity of the unknown time waveform C101. The classification processing means 23 then outputs an estimated label C102, which is the estimated label c.

Although not shown, the classification unit 20, like the learning unit 10, is provided with a memory means having a main memory area and an auxiliary memory area.

The learning unit 10 and classification unit 20 in such a classification system 1 are composed of a computing device such as a microcomputer that realizes various functions by executing software, or hardware such as a circuit device corresponding to various functions.

Figure 2 is a hardware configuration diagram showing an example of the configuration of the learning unit in Figure 1. When the various functions of the learning unit 10 are executed by hardware, the learning unit 10 in Figure 1 is configured by a processing circuit 31 as shown in Figure 2. In the learning unit 10 in Figure 1, the functions of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13, and the storage means 14 are realized by the processing circuit 31.

When each function is executed by hardware, the processing circuit 31 corresponds to, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, an ASIC (Application Specific Integrated Circuit), an FPGA (Field-Programmable Gate Array), or a combination of these. The learning unit 10 may realize the functions of each of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13, and the storage means 14 by the processing circuit 31, or may realize the functions of each of the parts by a single processing circuit 31.

Figure 3 is a hardware configuration diagram showing another example of the configuration of the learning unit in Figure 1. When the various functions of the learning unit 10 are executed by software, the learning unit 10 in Figure 1 is configured with a processor 41 and a memory 42 as shown in Figure 3. In the learning unit 10, the functions of the time-frequency analysis processing means 11, the feature analysis processing means 12, the learning processing means 13, and the storage means 14 are realized by the processor 41 and the memory 42.

When each function is executed by software, the functions of the time-frequency analysis processing means 11, feature analysis processing means 12, learning processing means 13 and storage means 14 in the learning unit 10 are realized by software, firmware, or a combination of software and firmware. The software and firmware are written as programs and stored in the memory 42. The processor 41 realizes the functions of each unit by reading and executing the programs stored in the memory 42.

For example, non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM, flash memory, EPROM (Erasable and Programmable ROM), and EEPROM (Electrically Erasable and Programmable ROM) may be used as the memory 42. In addition, for example, removable recording media such as magnetic disks, flexible disks, optical disks, CDs, MDs (Mini Discs), and DVDs (Digital Versatile Discs) may also be used as the memory 42.

In this way, the learning unit 10 can realize each of the above-mentioned functions by hardware, software, firmware, or a combination of these. Note that, like the learning unit 10, the classification unit 20 can also realize each of its functions by hardware, software, firmware, or a combination of these.

[Operation of Classification System 1]
The following describes the operation of the classification system 1 according to the present embodiment 1. First, in order to facilitate understanding of the operation of the classification system 1 according to the present embodiment 1, the configuration and operation of a conventional classification system will be described.

(Configuration of conventional classification system 100)
4 is a block diagram showing an example of the configuration of a conventional classification system 100. As shown in FIG. 4, the conventional classification system 100 includes a learning unit 110 and a classification unit 120.

The learning unit 110 has a time-frequency analysis processing means 111, a feature analysis processing means 112, and a learning processing means 113. The learning unit 110 receives, as in the learning unit 10 in the classification system 1 according to the first embodiment, time waveforms L101-n (n=1, ..., N) of N acoustic signals recorded in the past.

The time-frequency analysis processing means 111, like the time-frequency analysis processing means 11, receives the time waveform L101-n, performs time-frequency analysis processing on the time waveform L101-n, and outputs the time-frequency analysis result as the processing result.

The feature analysis processing means 112 receives the N time-frequency analysis results and performs feature analysis processing using a matrix operation. The feature analysis processing means 112 then generates features and labels for each time waveform L101-n to be supplied to the learning processing means 113, and feature extraction parameters to be supplied to the feature extraction processing means 122, and outputs the generated features, labels, and feature extraction parameters.

Like the learning processing means 13, the learning processing means 113 receives the features and the labels of each time waveform L101-n, estimates classification parameters to be supplied to the classification processing means 123 from the distribution of the features, and outputs the estimated classification parameters.

The classification unit 120 has a time-frequency analysis processing means 121, a feature extraction processing means 122, and a classification processing means 123. Similar to the classification unit 20 in the classification system 1 according to the first embodiment, the classification unit 120 receives the time waveform C101 as input and estimates the label of the signal based on the input time waveform C101.

The time-frequency analysis processing means 121 has a configuration similar to that of the time-frequency analysis processing means 111, performs time-frequency analysis processing on the time waveform C101, and outputs the time-frequency analysis result as the processing result.

The feature extraction processing means 122, like the feature extraction processing means 22, refers to the feature extraction parameters supplied from the feature analysis processing means 112 and calculates features from the time-frequency analysis results. Then, the feature extraction processing means 122 outputs the calculated features.

Similar to the classification processing means 23, the classification processing means 123 refers to the classification parameters supplied from the learning processing means 113 and estimates the label c of the unknown time waveform C101 from the feature quantity of the unknown time waveform C101. Then, the classification processing means 123 outputs an estimated label C102, which is the estimated label c.

Operation of Conventional Classification System 100
Fig. 5 is a schematic diagram showing an example of a result of time-frequency analysis processing by the time-frequency analysis processing means. Fig. 6 is a schematic diagram showing an example of a result of feature analysis processing by the feature analysis processing means. Fig. 7 is a schematic diagram showing an example of a result of learning processing by the learning processing means. Figs. 5 to 7 show images of a time-frequency analysis result 201, a feature extraction parameter 202, a distribution of feature amounts 203, and a classification parameter 204. In Figs. 5 to 7, the time-frequency analysis result 201 has a plurality of data for each label. Also, Figs. 5 to 7 show an example in which the type of signal source is C=4.

As shown in Fig. 5 and Fig. 6, the feature extraction parameters 202 are in the form of M vectors, and are obtained by decomposing the common frequency components contained in the time-frequency analysis result 201. Note that Fig. 6 illustrates an example in which the number of vectors in the feature extraction parameters 202 is M = 2.

If the feature extraction parameters 202 are appropriately obtained, the distribution of the feature quantities 203 forms groups for each label as shown in FIG. 7. The number of axes of the distribution of the feature quantities 203 is M, which is the same as the number of vectors of the feature extraction parameters 202. The feature quantities 203 represent the degree to which the frequency components of the feature extraction parameters 202 are included in each time-frequency analysis result 201.

In Figures 5 to 7, the time-frequency analysis result 201 for label #1 will be used as an example. The time-frequency analysis result 201 for label #1 does not include the first frequency component of the feature extraction parameters 202, but does include the second frequency component. Therefore, looking at the distribution of features 203 in Figure 7, the features for label #1 are plotted as points at coordinates where the value of feature 1 is small and the value of feature 2 is large. As shown in Figure 7, the classification parameters 204 are parameters that represent the boundary lines for each label in the distribution of features 203.

In the conventional classification system 100 that operates in this manner, the feature analysis processing means 112 calculates and outputs feature extraction parameters and feature quantities from the time-frequency analysis results, as described above. At this time, the feature analysis processing means 112 performs various calculations on the time-frequency analysis results, which are matrix data or tensor data, using at least one of matrix operations, matrix sum and matrix multiplication.

Figure 8 is a schematic diagram for explaining the matrix sum operation process. As shown in Figure 8, when calculating the sum of matrices, the sum operation unit 301 is used. The sum operation unit 301 adds two input matrices A and B, and outputs the sum result, matrix C (= A + B).

Figure 9 is a schematic diagram for explaining the matrix multiplication process. As shown in Figure 9, when calculating the matrix multiplication, the multiplication unit 302 is used. The multiplication unit 302 multiplies two input matrices A and B, and outputs the multiplication result, matrix C (=AB).

(Conventional Issues)
Incidentally, when the feature analysis processing means 112 performs a matrix calculation, a main memory area and an auxiliary memory area are used to store the calculation results. The matrix data as the time-frequency analysis result input to the feature analysis processing means 112 is stored in the auxiliary memory area, and is read into the main memory area during the feature analysis processing. The feature analysis processing means 112 then performs the feature analysis processing based on the matrix data read into the main memory area, and writes the matrix data obtained by the processing into the main memory area.

At this time, in the matrix calculations performed by the conventional feature analysis processing means 112, at least one of the row size and column size of the input matrix data is very large, and the following two problems may occur due to insufficient capacity of the main memory area.

Figure 10 is a schematic diagram for explaining the first problem caused by insufficient main memory area. The first problem is that when the capacity of the main memory area is insufficient, the matrix data stored in the auxiliary memory area cannot be read into the main memory area. When performing a matrix operation, the matrix data of both matrices A and B are read into the main memory area. However, in the example shown in Figure 10, when the matrix data of matrix A is read from the auxiliary memory area into the main memory area, the remaining area of the main memory area becomes smaller than the matrix data of matrix B, and the matrix data of matrix B cannot be read into the main memory area.

Figure 11 is a schematic diagram for explaining a second problem caused by a shortage of main memory area. The second problem is that when the main memory area is insufficient, the matrix data obtained by the matrix operation cannot be loaded into the main memory area. In the example shown in Figure 11, the matrix data of matrices A and B are loaded from the auxiliary memory area into the main memory area. After that, when a matrix operation is performed using matrices A and B and matrix C is calculated, the remaining capacity of the main memory area becomes smaller than the capacity of the matrix data of matrix C, and the matrix data of matrix C cannot be loaded into the main memory area.

As such, in the conventional classification system 100, the main memory area of the feature analysis processing means 112 needs to be at least an area that can store both the matrix data used in the matrix calculation and the matrix data obtained by the matrix calculation. In other words, in the conventional classification system 100, the main memory area needs to have a capacity at least three times the capacity of the matrix data.

Therefore, the feature analysis processing means 12 of the classification system 1 according to the present embodiment 1 is capable of performing matrix calculations even when the capacity of the main memory area 14a is smaller than the capacity that would conventionally be required.

(Matrix Calculation by Feature Analysis Processing Means 12 of Classification System 1)
A description will now be given of the matrix calculation performed by the feature analysis processing means 12 according to the present embodiment 1. In the present embodiment 1, the feature analysis processing means 12 performs feature analysis processing using submatrix calculations instead of conventional matrix calculations.

In this embodiment 1, the feature analysis processing means 12 performs various calculations on the time-frequency analysis results, which are matrix data or tensor data, using at least one of the matrix operations "matrix sum using submatrices" and "matrix multiplication using submatrices."

Fig. 12 is a schematic diagram for explaining a submatrix. As shown in Fig. 12, the submatrix is, for example, an M _i ×N _j matrix obtained by dividing an M ×N matrix A into I parts in the row direction and J parts in the column direction. In this case, i = 1, ..., I, and j = 1, ..., J. The sum of M _i is M, and the sum of N _j is N.

(Matrix addition using submatrices)
First, the matrix sum operation using the submatrix will be described. Fig. 13 is a schematic diagram for explaining the sum operation process using the submatrix. As shown in Fig. 13, when the matrix sum operation process is performed using the submatrix, a first division unit 221 and a plurality of submatrix sum operation means 222 are used.

The first division unit 221 divides the two input matrices A and B into submatrices A _ij and B _ij (i=1,...,I, j=1,...,J) of the same size M _i ×N _j . The submatrix sum calculation means 222 are provided in accordance with the number of divided submatrices. The submatrix sum calculation means 222 adds the two input submatrices A _ij and B _ij and outputs the submatrix C _ij that is the addition result.

When calculating the sum of matrices using submatrices, the first division unit 221 divides the matrix A and matrix B of size M×N into I×J submatrices, as shown in formulas (1) and (2). In formulas (1) and (2), i=1,...,I, and j=1,...,J. The sum of M _i is M, and the sum of N _j is N. The submatrix A _ij and the submatrix B _ij are both matrices of the same size M _i ×N _j .

Here, the first division unit 221 holds index numbers indicating which portions of data of matrices A and B submatrix A _ij and submatrix B _ij have been extracted from, and does not read all of the data of matrices A and B into the main memory area 14 a at once.

Next, the submatrix sum calculation means 222 receives the submatrix A _ij and the submatrix B _ij and outputs the submatrix C _ij . The correspondence between the input and output in this case is shown in the circled portion in FIG.

Fig. 14 is a block diagram showing an example of the configuration of the submatrix sum calculation means of Fig. 13. As shown in Fig. 14, the submatrix sum calculation means 222 is configured to include a sum calculation unit 223. In this way, when calculating the sum of submatrices, the sum calculation unit 223 is used. The sum calculation unit 223 adds two input submatrices A _ij and B _ij , and outputs a submatrix C _ij which is the addition result.

Fig. 15 is a schematic diagram showing an example of the state of the main memory area and the auxiliary memory area in a sum operation process using submatrices. At this time, as shown in Fig. 15, the characteristic analysis processing means 12 reads the submatrix _Aij and the submatrix _Bij from the auxiliary memory area 14b to the main memory area 14a, and calculates the submatrix _Cij using equation (3). Then, the characteristic analysis processing means 12 writes the calculated submatrix _Cij from the main memory area 14a to the auxiliary memory area 14b.

Finally, when the submatrices C _ij for all i and j are calculated, the matrix sum C=A+B is obtained as shown in equation (4).

(Matrix multiplication using submatrices)
Next, the matrix multiplication using the submatrices will be described. Fig. 16 is a schematic diagram for explaining the multiplication process using the submatrices. As shown in Fig. 16, when the matrix multiplication process is performed using the submatrices, a second division unit 224 and a plurality of submatrix multiplication means 225 are used.

The second division unit 224 receives an M×L matrix A and an L×N matrix B. The second division unit 224 divides the input M×L matrix A into submatrices A _ik (i=1,...,I, k=1,...,K) of M _i × _Lk size. The second division unit 224 also divides the input L×N matrix B into submatrices B _kj (k=1,...,K, j=1,...,J) of L _k × _Nj size. The submatrix product calculation means 225 calculates a submatrix C _ij based on the two input submatrices A _ik and B _kj .

When calculating the matrix product using submatrices, the second division unit 224 divides the matrix A of M×L size into I×K submatrices as shown in formula (5). The second division unit 224 also divides the matrix B of L×N size into K×J submatrices as shown in formula (6). In formulas (5) and (6), i=1,...,I, j=1,...,J, and k=1,...,K. The sum of M _i is M, the sum of N _j is N, and the sum of L _k is L. Furthermore, the submatrix A _ik is a matrix of M _i ×L _k size, and the submatrix B _kj is a matrix of L _k ×N _j size.

Here, the second division unit 224 holds index numbers indicating which portions of data from matrices A and B the submatrix A _ik and submatrix B _kj represent, and does not read all of the data of matrices A and B into the main memory area 14 a at once.

Next, the submatrix product calculation means 225 receives the submatrix A _ik and the submatrix B _kj and outputs the submatrix C _ij . The correspondence between the input and output in this case is shown in the circled area in Fig. 16. The submatrix product calculation means 225 calculates the submatrix C _ij from the input submatrix A _ik and the submatrix B _kj using equation (7).

Figure 17 is a block diagram showing an example of the configuration of the submatrix multiplication means of Figure 16. Figure 18 is a schematic diagram showing an example of the state of the main memory area and auxiliary memory area in a multiplication process using submatrices. As shown in Figure 17, the submatrix multiplication means 225 is configured to include a multiplication unit 226 and a summation unit 227. In this way, when calculating the product of submatrices, the multiplication unit 226 and the summation unit 227 are used.

The multiplication unit 226 multiplies the two input submatrices A _ij and B _ij together and outputs the multiplication result, submatrix A _ij B _ij . The addition unit 227 adds the two input submatrices together and outputs the addition result, submatrix C _ij .

18, first, the submatrix A _ik and the submatrix B _kj are read from the auxiliary memory area 14b to the main memory area 14a, and the submatrix C _ij is calculated based on the formula (7). Then, the calculated submatrix C _ij is written from the main memory area 14a to the auxiliary memory area 14b.

In the actual calculation by the submatrix multiplication calculation means 225, first, the submatrix multiplication calculation means 225 initializes the submatrix C _ij to be output by the zero matrix O. Next, as a first multiplication calculation process, the submatrix A _ik and the submatrix B _kj are read from the auxiliary storage area 14b to the main storage area 14a, and the multiplication calculation unit 226 calculates A _ik B _kj . As a second multiplication calculation process, the submatrix multiplication calculation means 225 calculates "C _ij +A _ik B _kj " in the addition calculation unit 227, and updates the submatrix C _ij read into the main storage area 14a by overwriting it. That is, the submatrix multiplication calculation means 225 accumulates the calculated A _ik B _kj to the submatrix C _ij read into the main storage area 14a.

The submatrix multiplication means 225 repeats the first and second multiplication processes for all k, and writes the finally obtained submatrix C _ij from the main memory area 14 a to the auxiliary memory area 14 b, thereby performing the calculation corresponding to the formula (7).

Finally, when the calculation of the submatrix C _ij for all i and j is completed, the submatrix product calculation means 225 calculates the matrix product "C=AB" as shown in equation (8).

In this way, in the first embodiment, when submatrices are used in a matrix addition operation, matrix C = A + B is output, similar to a conventional matrix addition operation. Also, when submatrices are used in a matrix multiplication operation, matrix C = AB is output, similar to a conventional matrix multiplication operation.

In addition, in this embodiment 1, if the capacity of the main memory area 14a is at least approximately three times the size of the submatrix used in the calculation, matrix C can be obtained by a matrix sum operation using the submatrices. Furthermore, if the capacity of the main memory area 14a is at least approximately four times the size of the submatrix used in the calculation, matrix C can be obtained by a matrix multiplication operation using the submatrices.

This improves problems that could occur with conventional technology, such as the inability to load matrix data due to insufficient main memory area, or the inability to store the matrix data resulting from calculations in the main memory area. Furthermore, since it becomes possible to process huge amounts of matrix data, the classification ability of the classification unit 20 can be improved.

As described above, in the classification system 1 according to the first embodiment, submatrix calculations are used during feature analysis processing. Therefore, even when huge matrix data is handled in the learning unit 10 in order to improve the classification ability of the classification unit 20, it is only necessary to provide a main memory area according to the size of the submatrix, and therefore matrix calculations can be performed appropriately even when the capacity of the main memory area is restricted.

1 Classification system, 10 Learning unit, 11 Time-frequency analysis processing means, 12 Feature analysis processing means, 13 Learning processing means, 14 Storage means, 14a Main memory area, 14b Auxiliary memory area, 20 Classification unit, 21 Time-frequency analysis processing means, 22 Feature extraction processing means, 23 Classification processing means, 31 Processing circuit, 41 Processor, 42 Memory, 100 Classification system, 110 Learning unit, 111 Time-frequency analysis processing means, 112 Feature analysis processing means, 113 Learning processing means, 120 Classification unit, 121 Time-frequency analysis processing means, 122 Feature extraction processing means, 123 Classification processing means, 201 Time-frequency analysis result, 202 Feature extraction parameters, 203 Feature amount, 204 Classification parameters, 221 First division unit, 222 Submatrix sum operation means, 223, 227, 301 Sum operation unit, 224 Second division unit, 225 submatrix multiplication means, 226, 302 multiplication unit.

Claims (4)

A classification system for performing feature extraction and class classification based on time-frequency analysis, comprising:
a time-frequency analysis processing means for converting a time waveform indicating a time series amplitude of a known signal, the signal source of which is known, into data indicating a time series frequency and signal intensity;
a feature analysis processing means for generating feature quantities and feature extraction parameters by performing a submatrix operation on the data;
A learning processing means for estimating classification parameters from the feature amounts ;
a storage means having a main storage area for storing processing results and an auxiliary storage area for storing the processing results and the data as matrices;
A learning unit having
a feature extraction processing means for calculating a feature amount of an unknown signal, the signal source of which is unknown, by referring to the feature extraction parameters;
a classification unit having a classification processing means for estimating a signal source of the unknown signal from a feature amount of the unknown signal by referring to the classification parameters ,
The feature analysis processing means reads out a plurality of submatrices of the matrix from the auxiliary memory area into the main memory area, calculates the product of the two submatrices read out into the main memory area, stores the product in the auxiliary memory area, and further reads out the product and a submatrix of the output matrix from the auxiliary memory area, sequentially calculates the sum, and calculates the submatrix product . An information processing device that performs feature extraction and class classification based on time-frequency analysis,
a time-frequency analysis processing means for converting a time waveform indicating a time series amplitude of a known signal, the signal source of which is known, into data indicating a time series frequency and signal intensity;
a feature analysis processing means for generating feature quantities and feature extraction parameters by performing a submatrix operation on the data;
A learning processing means for estimating classification parameters from the feature amounts;
a feature extraction processing means for calculating a feature amount of an unknown signal, the signal source of which is unknown, by referring to the feature extraction parameters;
a classification processing means for estimating a signal source of the unknown signal from a feature quantity of the unknown signal by referring to the classification parameters ;
a storage means having a main storage area for storing processing results and an auxiliary storage area for storing the processing results and the data as matrices;
Equipped with
The feature analysis processing means performs the submatrix operation by reading a plurality of submatrices of the matrix from the auxiliary memory area into the main memory area, calculating the product of the two submatrices read into the main memory area and storing the product in the auxiliary memory area, and further reading the product and a submatrix of the output matrix from the auxiliary memory area and sequentially calculating the sum to calculate the submatrix product . A classification method comprising : a storage means having a main storage area for storing processing data and an auxiliary storage area for storing data and processing results; and performing feature extraction and class classification based on time-frequency analysis, the method comprising:
A time waveform indicating a time series amplitude of a known signal, the signal source of which is known, is converted into the data indicating a time series frequency and signal intensity;
generating feature quantities and feature extraction parameters using a submatrix operation on the matrixed data;
Estimating classification parameters from the features;
Calculate a feature amount of an unknown signal whose signal source is unknown by referring to the feature extraction parameters;
A signal source of the unknown signal is estimated from a feature quantity of the unknown signal by referring to the classification parameters,
The submatrix operation includes reading a plurality of submatrices of the matrix from the auxiliary memory area to the main memory area, calculating a product of the two submatrices read into the main memory area and storing the product in the auxiliary memory area, and further reading the product and a submatrix of an output matrix from the auxiliary memory area and sequentially calculating a sum to calculate the submatrix product . A computer comprising a storage means having a main storage area for storing processing data and an auxiliary storage area for storing data and processing results, and performing feature extraction and class classification based on time-frequency analysis,
a time-frequency analysis processing means for converting a time waveform indicating a time series amplitude of a known signal, the signal source of which is known, into said data indicating a time series frequency and signal intensity;
A feature analysis processing means for generating feature quantities and feature extraction parameters by performing a submatrix operation on the matrix-arranged data;
A learning processing means for estimating classification parameters from the feature amounts;
a feature extraction processing means for calculating a feature quantity of an unknown signal, the signal source of which is unknown, by referring to the feature extraction parameters;
functioning as a classification processing means for estimating a signal source of the unknown signal from a feature amount of the unknown signal by referring to the classification parameters ;
A program for performing the submatrix operation, the program including: the feature analysis processing means reading a plurality of submatrices of the matrix from the auxiliary memory area into the main memory area; calculating the product of the two submatrices read into the main memory area and storing the product in the auxiliary memory area; and further reading the product and a submatrix of an output matrix from the auxiliary memory area, and sequentially calculating the sum to calculate the submatrix product .

Images