Processing device

Abstract

This processing device comprises: an acquisition unit that acquires region-of-interest information indicating a region of interest from among training data used when carrying out machine learning for ordered classification of classes; and a class division calculation unit that determines the manner of class division in the training data in accordance with the region-of-interest information acquired by the acquisition unit. For example, using the configuration as described above, the processing device adjusts and optimizes the cost or the like of performing machine learning.

Description

Processing equipment

[0001] The present disclosure relates to a processing device, a processing method, and a recording medium.

[0002] Machine learning models can be used to classify classes, etc.

[0003] For example, Patent Literature 1 describes a machine learning program that discretizes training data (learning data). According to Patent Literature 1, for multiple pieces of learning data, the program converts data item values ​​for each data item into discretized data values ​​that are discretized based on a predetermined criterion, and executes a learning process to learn a model that makes a judgment regarding label information using the discretized data values ​​of the multiple data items as input. The program also acquires multiple pieces of feature information and index values ​​indicating the importance of each piece of feature information from the execution result of the learning process, selects one or more pieces of feature information from the multiple pieces of feature information based on the index values, and changes the criterion used to discretize the data item values ​​based on the selected one or more pieces of feature information.

[0004] International Publication No. 2021 / 075029

[0005] When classifying ordered classes, there may be important regions or regions of interest, such as regions where more detailed classification is desired. However, in the case of the technology described in Patent Literature 1, regions of interest and regions other than the important regions are treated equally, and learning is performed at the same cost for both the important and uninterested regions. As a result, the balance between accuracy and cost may be poor, making it difficult to perform appropriate machine learning.

[0006] Therefore, one of the objects of the present disclosure is to provide a processing device, a processing method, and a recording medium that can solve the above-mentioned problems.

[0007] In order to achieve this object, the processing device in the present disclosure has a configuration including: an acquisition unit that acquires area-of-interest information indicating an area of ​​interest in training data used when performing machine learning to classify ordered classes; and a class division calculation unit that determines how to divide the training data into classes according to the area-of-interest information acquired by the acquisition unit.

[0008] Furthermore, the processing method in the present disclosure is configured such that an information processing device acquires area of ​​interest information indicating an area of ​​interest in training data used when performing machine learning to classify ordered classes, and determines how to divide the training data into classes according to the acquired area of ​​interest information.

[0009] Furthermore, the recording medium in the present disclosure is a computer-readable recording medium having recorded thereon a program for causing an information processing device to perform the following process: acquire area of ​​interest information indicating an area of ​​interest in training data used when performing machine learning to classify ordered classes; and determine how to divide the training data into classes according to the acquired area of ​​interest information.

[0010] According to the above-described configurations, machine learning can be performed using more appropriate training data.

[0011] FIG. 1 is a diagram for explaining an overview of a learning system in the present disclosure. FIG. 1 is a diagram illustrating an example configuration of a learning system. FIG. 2 is a block diagram illustrating an example configuration of a WDM terminal device. FIG. 3 is a block diagram illustrating an example configuration of a learning device. FIG. 4 is a diagram illustrating an example class division policy. FIG. 5 is a diagram illustrating an example class division calculation. FIG. 6 is a diagram illustrating an example class division calculation. FIG. 7 is a flowchart illustrating an example operation of a learning device. FIG. 8 is a diagram illustrating another example configuration of a learning system. FIG. 9 is a diagram illustrating another example configuration of a learning system. FIG. 10 is a diagram illustrating another example application of a learning system. FIG. 11 is a diagram illustrating an example hardware configuration of a processing device in the present disclosure. FIG. 12 is a block diagram illustrating an example configuration of a processing device. FIG. 13 is a flowchart illustrating an example operation of a processing device.

[0012] [First Embodiment] A configuration example of a learning system 100 according to the present disclosure will be described with reference to Figs. 1 to 12. Fig. 1 is a diagram for explaining an overview of the learning system 100. Fig. 2 is a diagram showing a configuration example of the learning system 100. Fig. 3 is a block diagram showing a configuration example of a WDM terminal device 200. Fig. 4 is a block diagram showing a configuration example of a learning device 300. Fig. 5 is a diagram showing an example of a class division policy 342. Figs. 6 and 7 are diagrams showing examples of class division calculations. Fig. 8 is a flowchart showing an example of the operation of the learning device 300. Figs. 9 and 10 are diagrams showing other configuration examples of the learning system 100. Figs. 11 and 12 are diagrams showing other application examples of the learning system. Note that in the present disclosure, the drawings may be associated with one or more embodiments.

[0013] In a first embodiment of the present disclosure, a learning system 100 is described that determines how to divide classes during machine learning in accordance with region-of-interest information indicating a region of interest when classifying classes with an order. For example, when classifying classes with an order, it may be important whether an inferred value is equal to or greater than a predetermined value, and so on, and so on. Furthermore, when performing classification with an order, i.e., when the order is meaningful, it may be necessary to focus on an arbitrary region, such as when an arbitrary region becomes important for various other reasons. The learning system 100 can determine how to divide classes in accordance with region-of-interest information indicating the region of interest as described above.

[0014] For example, as shown in FIG. 1 , the learning system 100 can merge multiple classes corresponding to regions determined to be unfocused based on the region-of-interest information. This allows the learning system 100 to reduce the number of training data selected for machine learning in regions not of interest. Furthermore, the learning system 100 can perform more dense class division in regions determined to be focused based on the region-of-interest information. In other words, the learning system 100 can narrow the width of classes corresponding to regions determined to be focused compared to the width of classes corresponding to regions determined to be unfocused based on the region-of-interest information. In this case, the learning system 100 may combine the above-described class merging and dense division processing. For example, by performing the above-described processing while keeping the number of training data selected within each class constant, the learning system 100 can reduce the number of training data in regions not of interest, thereby reducing costs, or enable learning in regions of interest in denser steps. The learning system 100 may also determine a class division method other than the above-described example. For example, the learning system 100 may decide to perform dense division processing while maintaining the number of classes, or may decide to increase or decrease the number of learning data to select within each class depending on whether the area is of interest.

[0015] Furthermore, when determining how to divide classes, the learning system 100 can refer to guideline information that indicates guidelines for class division, such as cost priority or accuracy priority, in addition to the area of ​​interest information. For example, when the guideline information indicates cost priority, the learning system 100 merges multiple classes corresponding to areas determined to be unfocused. Furthermore, when the guideline information indicates accuracy priority, the learning system 100 performs more dense class division in areas determined to be focused. The learning system 100 may combine the above methods, for example, when prioritizing both cost and accuracy. For example, as described above, the learning system 100 can be configured to determine how to divide classes according to the area of ​​interest information and guideline information.

[0016] The specific configuration of the training system 100 will be described below. For example, the training system 100 infers quality values ​​such as an optical signal-to-noise ratio (OSNR) based on an input constellation waveform as ordered class classification. Furthermore, the training system 100 uses, as region-of-interest information, a service level requirement (SLR) value, which specifies that the OSNR value inferred by the model must be equal to or greater than a predetermined value. FIG. 2 shows an example configuration of the training system 100 when the present invention is applied in the above-described manner. Referring to FIG. 2, the training system 100 may include a WDM (Wavelength Division Multiplexing) terminal device 200 and a training device 300. As shown in FIG. 2, the WDM terminal device 200 and the training device 300 may be connected to each other via wired or wireless communication so as to be able to communicate with each other.

[0017] A constellation, also known as a signal space diagram, is a representation of digitally modulated data signal points on a two-dimensional complex plane. The quality value to be inferred may be any value corresponding to degradation of the signal waveform, such as chromatic dispersion, or any other value that affects the signal quality.

[0018] The WDM terminal device 200 is a device that receives an optical signal transmitted using wavelength division multiplexing. The WDM terminal device 200 may also be configured to measure quality values ​​such as OSNR. FIG. 3 shows an example of the configuration of the WDM terminal device 200. Referring to FIG. 3, the WDM terminal device 200 has a typical configuration of a terminal device, including a demultiplexer 210, a receiver 220, and a received light data collection unit 230. The WDM terminal device 200 also has a storage device, in which learning data information 240 and a learned model 250 are stored.

[0019] The demultiplexer 210 demultiplexes multiple optical signals of different wavelengths transmitted through an optical fiber into individual wavelengths. For example, in optical wavelength division multiplexing communications, multiple wavelengths are superimposed and transmitted through an optical fiber. The demultiplexer 210 can demultiplex the multiple wavelengths that are superimposed and transmitted into individual wavelengths.

[0020] The receiver 220 receives the optical signal demultiplexed by the demultiplexer 210 and performs OE (Optical signal / Electrical signal) conversion and AD (Analog / Digital) conversion. Referring to FIG. 3 , the receiver 220 includes, for example, an OE converter 221, an AD converter 222, and a DSP (Digital Signal Processor) 223. For example, the OE converter 221 converts the optical signal received from the demultiplexer 210 into an electrical signal. The AD converter 222 converts the analog signal output by the OE converter 221 into a digital signal. The DSP 223 performs predetermined processing based on the digital signal. For example, the DSP 223 can perform error correction coding processing and various processing required to estimate the optical signal-to-noise ratio. The processing performed by the DSP 223 may be general processing.

[0021] The received light data collection unit 230 collects signal information such as constellation waveforms for optical signals of each wavelength. The received light data collection unit 230 can also store the collected constellation waveforms as learning data information 240 in a storage device.

[0022] The learning data information 240 includes learning data to be selected when the learning system 100 performs machine learning. In other words, the learning system 100 can select a portion of the learning data included in the learning data information 240 to perform machine learning.

[0023] For example, the training data information 240 includes multiple pieces of training data that associate constellation waveforms, which are data, collected by the received light data collector 230 or the like, with quality values, such as OSNR, which are labels. Here, the constellation waveforms may be data that expresses symbols of an optical signal received by the WDM terminal device 200 as coordinates on an IQ plane. Furthermore, the quality values, such as OSNR, may be values ​​measured using the WDM terminal device 200 or an external measuring device. For example, the training data information 240 includes multiple pieces of training data for each region within the range where the OSNR is 10 dB or more and less than 40 dB, such as a region where the OSNR is 10 dB or more and less than 15 dB, and a region where the OSNR is 15 dB or more and less than 20 dB. The training data information 240 may include more training data than the number of pieces of training data that the training device 300, described later, may select in each region, so that the training device 300 can make a selection.

[0024] The trained model 250 includes information about a model trained by the learning device 300 (described later). The trained model 250 may include weight values ​​of the trained model, etc. The trained model 250 is updated as the learning device 300 (described later) performs machine learning using training data selected according to a predetermined criterion from the training data included in the training data information 240.

[0025] The above is an example of the configuration of the WDM terminal device 200. Note that the WDM terminal device 200 may have a configuration other than that illustrated in FIG. 3 , such as a communication quality estimation unit that estimates the signal-to-noise ratio by inputting the constellation waveform collected by the received light data collection unit 230 into the trained model 250.

[0026] The learning device 300 is an information processing device (processing device) that selects learning data to be used for learning from the learning data information 240 and performs machine learning using the selected learning data. Fig. 4 shows an example configuration of the learning device 300. Referring to Fig. 4, the learning device 300 has, as main components, for example, an operation input unit 310, a screen display unit 320, a communication I / F unit 330, a storage unit 340, and an arithmetic processing unit 350.

[0027] 4 illustrates an example in which the functions of the learning device 300 are realized using a single information processing device. However, at least some of the functions of the learning device 300 may be realized using multiple information processing devices, for example, on the cloud. For example, the learning device 300 may be composed of a processing device that determines how to divide the classes, and a learning device that selects and learns learning data in accordance with the determination by the processing device. Furthermore, the learning device 300 may not include some of the components illustrated above, such as the operation input unit 310 or the screen display unit 320, or may have components other than those illustrated above.

[0028] Operation input unit 310 is made up of operation input devices such as a keyboard, a mouse, etc. Operation input unit 310 detects operations of the operator operating learning device 300 and outputs the operations to calculation processing unit 350.

[0029] The screen display unit 320 is composed of a screen display device such as a liquid crystal display, an organic electroluminescence (EL) display, etc. The screen display unit 320 can display various information stored in the storage unit 340 on the screen in response to instructions from the arithmetic processing unit 350.

[0030] The communication I / F unit 330 includes a data communication circuit, etc. The communication I / F unit 330 performs data communication with an external device connected via a communication line.

[0031] The storage unit 340 is a storage device such as a hard disk or memory. The storage unit 340 stores processing information and programs 344 required for various processes in the arithmetic processing unit 350. The programs 344 are read into the arithmetic processing unit 350 and executed to realize various processing units. The programs 344 are read in advance from an external device or recording medium via a data input / output function such as the communication I / F unit 330, and are stored in the storage unit 340. Main information stored in the storage unit 340 includes, for example, definition information 341, a class division policy 342, and service level requirement information 343.

[0032] The definition information 341 includes information about how to divide classes in a default state. The definition information 341 is acquired in advance by receiving it from an external device via the communication I / F unit 330, inputting information via the operation input unit 310, or the like, and is stored in the storage unit 340.

[0033] For example, the definition information 341 includes information indicating how to divide classes in a default state, such as dividing the range of OSNR from 10 dB to less than 40 dB into five steps, selecting 1,000 pieces of training data for each divided class, etc. The definition information 341 may include information indicating the range of classes to be divided, such as OSNR from 10 dB to less than 15 dB, OSNR from 15 dB to less than 20 dB, or OSNR from 20 dB to less than 25 dB, or may include information indicating the width of division, such as 5 dB, or any other information other than the examples.

[0034] The class division policy 342 includes information indicating a method for determining how to divide classes according to a region of interest for each guideline such as cost priority, accuracy priority, etc. The class division policy 342 is acquired in advance by receiving it from an external device via the communication I / F unit 330, inputting information via the operation input unit 310, or the like, and is stored in the storage unit 340.

[0035] FIG. 5 shows an example of information included in the class division policy 342. As illustrated in FIG. 5, the class division policy 342 can include information indicating a method for determining how to divide classes for each guideline, such as cost priority or accuracy priority. Furthermore, the class division policy 342 can include information indicating a region of interest, a region with a value higher than the region of interest, a region with a value lower than the region of interest, and a method for determining how to divide classes in each region, for each guideline. For example, in the case of the learning system 100 described in the present disclosure, as described above, the service level requirement (SLR) value is used as the region of interest information. Therefore, the class division policy 342 can include information indicating a region where the OSNR is the SLR value, a region where the OSNR is greater than the SLR value, and a region where the OSNR is smaller than the SLR value, and a method for determining how to divide classes in each region.

[0036] Specifically, in the case of a cost-first approach, the class division policy 342 can indicate that the class division method should remain the default state indicated by the definition information 341 between a region where the OSNR is equal to the SLR value and a region where the OSNR is greater than the SLR value. Furthermore, in the case of a cost-first approach, the class division policy 342 can indicate that in a region where the OSNR is smaller than the SLR value, classes that are a predetermined value n or more away from the class corresponding to the region of interest should be merged. In other words, the class division policy 342 can indicate that classes that are a predetermined value n or more away from the class corresponding to the region of interest can be determined as not being of interest. Similarly, in the case of an accuracy-first approach, the class division policy 342 can indicate that the width of the class corresponding to the region where the OSNR is equal to the SLR value should be set to 1, or that the width of the class within the predetermined value n should be determined sequentially. In other words, the class division policy 342 can indicate that the region determined to be of interest should be densely divided. Furthermore, the class division policy 342 can indicate that, when prioritizing accuracy while maintaining costs, the width of classes within a predetermined value n is determined numerically and the number of classes is maintained.

[0037] For example, as described above, the class division policy 342 may include, for each guideline, information indicating a method for determining how to divide classes according to the region of interest. The value of n may be determined arbitrarily. The value of n may be the same for each guideline or may be different for each guideline. Although FIG. 5 shows an example of the class division policy 342, the information included in the class division policy 342 may be other than that exemplified in FIG. For example, the degree to which the region determined to be of interest is to be divided may be determined arbitrarily.

[0038] The service level requirement information 343 includes a service level requirement (SLR) value indicating a region of interest. For example, the service level requirement information 343 may include information such as an OSNR of 30 dB or higher. When the service level requirement information 343 includes the above information, the region of interest is 30 dB. The service level requirement information 343 is updated in response to the division requirement information acquisition unit 351 acquiring an SLR value as region of interest information, etc.

[0039] The arithmetic processing unit 350 has an arithmetic device such as a CPU (Central Processing Unit) and its peripheral circuits. The arithmetic processing unit 350 reads and executes a program 344 from the storage unit 340, thereby realizing various processing functions by causing the above hardware and the program 344 to work together. Major processing units realized by the arithmetic processing unit 350 include, for example, a division requirement information acquisition unit 351, a class division calculation unit 352, a learning data selection unit 353, a model learning unit 354, and an output unit 355.

[0040] In addition, instead of the above-mentioned CPU, the arithmetic processing unit 350 may have a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), an MPU (Micro Processing Unit), an FPU (Floating point number Processing Unit), a PPU (Physics Processing Unit), a TPU (Tensor Processing Unit), a quantum processor, a microcontroller, or a combination of these.

[0041] The division requirement information acquisition unit 351 acquires division requirement information that indicates requirements for calculating and determining how to divide classes. The division requirement information acquisition unit 351 acquires the division requirement information in response to receiving the information from an external device via the communication I / F unit 330 or receiving information input via the operation input unit 310. For example, the division requirement information acquisition unit 351 may acquire the division requirement information from an external device such as a network management device connected via the communication I / F unit 330.

[0042] In the present disclosure, the division requirement information acquisition unit 351 acquires, as the division requirement information, area-of-interest information indicating the SLR. Then, the division requirement information acquisition unit 351 stores the acquired area-of-interest information in the storage unit 340 as service level requirement information 343. Furthermore, the division requirement information acquisition unit 351 can acquire guideline information as the division requirement information.

[0043] The class division calculation unit 352 calculates the range of classes to be divided, etc., according to the division requirement information acquired by the division requirement information acquisition unit 351. In other words, the class division calculation unit 352 calculates and determines the method of class division by calculating the range of classes to be divided, etc., according to the division requirement information. For example, the class division calculation unit 352 references the definition information 341, the class division policy 342, etc., according to the division requirement information acquired by the division requirement information acquisition unit 351. Then, the class division calculation unit 352 can calculate and determine the method of class division using the acquired division requirement information and each piece of referenced information.

[0044] FIG. 6 shows an example of how the class division calculation unit 352 calculates class division when the SLR indicates an OSNR of 30 dB or more. For example, in the above case, assume that the class division calculation unit 352 determines that the region with an OSNR of less than 25 dB is not of interest. In this case, the class division calculation unit 352 can determine to merge classes in the region with an OSNR of less than 25 dB, as shown in FIG. 6 . In the above case, the region of interest is 30 dB. Therefore, the class division calculation unit 352 can determine to perform denser class division in the region of interest, around 30 dB, as shown in FIG.

[0045] 7 shows a specific example of calculation when making a decision in accordance with the class division policy 342 illustrated in Fig. 5 according to the guideline information. For example, referring to Fig. 7, the class division calculation unit 352 determines the range of classes to be divided into, such as OSNR 10 dB or more and less than 15 dB, OSNR 15 dB or more and less than 20 dB, and OSNR 20 dB or more and less than 25 dB, in a default state.

[0046] Furthermore, when the guideline is cost-first, the class division policy 342 indicates that the default state is to be maintained in the region where the OSNR is equal to the SLR value and in the region where the OSNR is greater than the SLR value. Furthermore, the class division policy 342 indicates that in the region where the OSNR is smaller than the SLR value, classes that are a predetermined value n or more away from the class to which the region of interest belongs are to be merged. Therefore, the class division calculation unit 352 determines that the default state is to be maintained in the region where the OSNR is equal to or greater than the SLR value. Furthermore, when n = 2, for example, the class division calculation unit 352 merges classes that are two or more steps below the class in the region where the OSNR is 30. That is, the class division calculation unit 352 merges the classes corresponding to the regions where the OSNR is 10 dB or more but less than 15 dB, the OSNR is 15 dB or more but less than 20 dB, and the OSNR is 20 dB or more but less than 25 dB. As a result, the class division calculation unit 352 calculates division into four classes: OSNR 10 dB or more and less than 25 dB, OSNR 25 dB or more and less than 30 dB, OSNR 30 dB or more and less than 35 dB, and OSNR 35 dB or more and less than 40 dB.

[0047] Furthermore, when the guideline prioritizes accuracy, the class division policy 342 indicates that the class width to which the region where the OSNR is equal to the SLR belongs is set to 1, and that the width of the classes within a predetermined value n is determined in a numerical progression. Therefore, the class division calculation unit 352 sets the class width for the region where the OSNR is 30 dB indicated by the SLR to 1, which is narrower than the class width in the default state. In other words, the class division calculation unit 352 determines that the region where the OSNR is equal to or greater than 30 dB and less than 31 dB is to be divided into one class. Furthermore, for example, assume that the class division policy 342 specifies n=4 and the Fibonacci sequence as the numerical progression. In this case, the class division calculation unit 352 calculates a range of four classes above and below the determined class so that the class width satisfies the Fibonacci sequence, and classifies the region outside the above range into one class. As a result, the class division calculation unit 352 calculates division into classes of OSNR 19 dB or more but less than 24 dB, OSNR 24 dB or more but less than 27 dB, OSNR 27 dB or more but less than 29 dB, OSNR 29 dB or more but less than 30 dB, OSNR 31 dB or more but less than 32 dB, OSNR 32 dB or more but less than 34 dB, OSNR 34 dB or more but less than 37 dB, and OSNR 37 dB or more but less than 40 dB. Note that this disclosure illustrates an example of a class division method targeting OSNRs of 10 dB or more but less than 40 dB. Therefore, the class division calculation unit 352 calculates an OSNR of 37 dB or more but less than 40 dB, excluding 40 dB or more. Furthermore, the class division calculation unit 352 classifies any range outside the above ranges into a single class. In other words, the class division calculation unit 352 classifies values ​​equal to or greater than 10 dB and less than 19 dB as one class. For example, as described above, when the guideline is to prioritize accuracy, the class division calculation unit 352 narrows the class width corresponding to the region of interest more narrowly than the class width corresponding to the region of no interest, thereby dividing the region of interest into classes more densely.

[0048] Furthermore, when the guideline is to prioritize accuracy and maintain cost, the class division calculation unit 352 can perform merging processing to maintain the number of classes in addition to the processing when the guideline is to prioritize accuracy.

[0049] For example, as described above, the class division calculation unit 352 can determine how to divide classes according to the region of interest information and guideline information included in the division requirement information. Note that the class division calculation unit 352 may make a determination according to the region of interest information and guideline information using a method other than the above example. For example, in addition to or instead of the above example processing, the class division calculation unit 352 may be configured to adjust the number of training data to be selected, such as by increasing the number of training data to be selected in a class corresponding to a region of interest or by decreasing the number of training data to be selected in a class determined not to be of interest.

[0050] The learning data selection unit 353 selects and acquires learning data from the WDM terminal device 200 according to the result determined by the class division calculation unit 352. For example, the learning data selection unit 353 selects the same number of learning data, such as 1000 pieces, for each class determined by the class division calculation unit 352. Note that if the class division calculation unit 352 determines the number of learning data to select for each class, the learning data selection unit 353 may select learning data according to the number of learning data determined for each class.

[0051] The model learning unit 354 performs supervised machine learning using the learning data selected by the learning data selection unit 353. Note that the model learning unit 354 may perform supervised machine learning using the learning data selected by the learning data selection unit 353 using any method.

[0052] The output unit 355 transmits the model learned by the model learning unit 354 to the WDM terminal device 200. In other words, the output unit 355 can transmit parameters such as weight values ​​of the model learned by the model learning unit 354 to the WDM terminal device 200. The output unit 355 may also transmit information stored in the storage unit 340 to an external device such as the WDM terminal device 200.

[0053] The above is an example of the configuration of the learning device 300.

[0054] Next, an example of the operation of the learning device 300 will be described with reference to Fig. 8. Fig. 8 is a flowchart showing an example of the operation of the learning device 300. Referring to Fig. 8, the division requirement information acquisition unit 351 acquires division requirement information such as region of interest information and guideline information (step S101).

[0055] The class division calculation unit 352 calculates the range of classes to be divided, etc., in accordance with the division requirement information acquired by the division requirement information acquisition unit 351 (step S102). For example, the class division calculation unit 352 refers to the definition information 341, the class division policy 342, etc., in accordance with the division requirement information acquired by the division requirement information acquisition unit 351. Then, the class division calculation unit 352 calculates and determines how to divide classes using the acquired division requirement information and each piece of information that it has referred to.

[0056] The learning data selection unit 353 selects and acquires learning data from the WDM terminal device 200 in accordance with the result determined by the class division calculation unit 352 (step S103). For example, the learning data selection unit 353 selects the same number of learning data, such as 1000 pieces, for each class determined by the class division calculation unit 352.

[0057] The model learning unit 354 performs supervised machine learning using the learning data selected by the learning data selection unit 353 (step S104). Note that the model learning unit 354 may perform supervised machine learning using the learning data selected by the learning data selection unit 353 using any method.

[0058] The above is an example of the operation of the learning device 300. The output unit 355 of the learning device 300 can transmit the model learned by the model learning unit 354 to the WDM terminal device 200.

[0059] As described above, the learning device 300 includes a division requirement information acquisition unit 351 and a class division calculation unit 352. With this configuration, the class division calculation unit 352 can determine how to divide classes based on the region of interest information acquired by the division requirement information acquisition unit 351. As a result, it becomes possible to differentiate the cost during learning depending on whether or not the region is of interest. This allows for more appropriate machine learning to be performed depending on whether or not the region is of interest when classifying classes with an order.

[0060] Furthermore, the class division calculation unit 352 is configured to determine the class division method according to the guideline information. With this configuration, it is possible to determine the class division method according to the user's wishes. As a result, machine learning can be performed more appropriately.

[0061] In this disclosure, an example configuration of the learning system 100 has been described with reference to FIG. 2 . However, the configuration of the learning system 100 is not limited to the example shown in FIG. 2 . For example, as shown in FIG. 9 , functions similar to those of the processing units of the learning device 300 may be realized by a calculation device such as a CPU of the WDM terminal device 200 executing a program. In this case, the learning system 100 may not include the learning device 300 as a standalone device, but may be composed only of the WDM terminal device 200. Configuring the learning system 100 in this manner can reduce, for example, the transfer costs of training data. Furthermore, as shown in FIG. 10 , the learning system 100 may be composed of the WDM terminal device 200, the learning device 300, and a learning instruction device 400. For example, the learning instruction device 400 realizes the function of the class division calculation unit 352, one of the processing units of the learning device 300, by a calculation device of the learning instruction device 400 executing a program. For example, configuring the learning system 100 as shown in Figure 10 allows for improved configuration flexibility, such as arranging the learning device 300 at each base station or local station and arranging the learning instruction device 400 on the cloud. The configuration shown in Figure 10 also makes it easier to make batch changes to guidelines, for example. The learning system 100 may have any of the configurations illustrated in Figures 2, 9, and 10.

[0062] In addition, in the present disclosure, the division requirement information acquisition unit 351 is described as acquiring an SLR as the region of interest information. However, the division requirement information acquisition unit 351 may acquire, as the region of interest information, information that can calculate or infer an OSNR value, such as the required communication encoding method or required bandwidth. For example, in the above case, the division requirement information acquisition unit 351 may be configured to determine the region of interest by determining the required OSNR to realize the encoding method indicated by the acquired information. The division requirement information acquisition unit 351 may acquire information that can identify the region of interest other than the above examples.

[0063] Furthermore, in the present disclosure, the case where the WDM terminal device 200 stores the learning data in advance has been exemplified. However, the WDM terminal device 200 may be configured to acquire the learning data required for each class in accordance with the determination by the class division calculation unit 352. For example, the received light data collection unit 230 may be configured to collect a constellation waveform or the like in accordance with the determination by the class division calculation unit 352. In this way, the learning data does not necessarily have to be stored in advance in the WDM terminal device 200.

[0064] Furthermore, the learning device 300 may be configured to determine how to divide classes according to region-of-interest information without using guideline information. For example, the learning device 300 may be configured to acquire only region-of-interest information as division requirement information and perform predetermined processing such as class merging or dense class division according to the acquired region-of-interest information. In such a configuration, the learning device 300 does not need to acquire guideline information.

[0065] Furthermore, the present invention may be applied to cases other than inferring quality values ​​such as OSNR according to input of a constellation waveform. The present invention may be applied to various situations where ordered class classification is performed. For example, the present invention may be applied to the field of healthcare.

[0066] 11 and 12 show an example of applying the present invention to the healthcare field. For example, the present invention may be applied to a situation in which a risk level of dementia or the like is inferred based on input of gait information, such as walking speed, walking cycle, stride length, and bilateral symmetry, acquired using various sensors, as shown in FIG. 11 . For example, in this configuration, a model is trained by performing machine learning using training data in which gait information (data) is associated with risk levels (labels). In this case, the learning system 100 can perform the above learning after determining how to classify the training data used for machine learning. For example, among the risk levels shown in FIG. 12 , assume that a risk level of 60 or higher indicates a high probability of dementia. In this case, the learning system 100 can determine how to classify the data according to the region of interest, with a risk level of 60 or higher as the region of interest. As a result, it is possible to focus on inferring risk levels according to the region of interest, thereby achieving a balance between cost and accuracy. The present invention is not limited to the above-mentioned examples, and may be applied to inferring health conditions such as disease risk and mental state based on gait information, or to other situations where health conditions are inferred based on arbitrary data. As described above, the present invention may be applied to various situations where ordered class classification is performed, other than in fields such as optical communications and healthcare.

[0067] Second Embodiment Next, a configuration example of a processing device 500 will be described with reference to Fig. 13 to Fig. 15. Fig. 13 is a diagram showing a hardware configuration example of the processing device 500. Fig. 14 is a block diagram showing a configuration example of the processing device 500. Fig. 15 is a flowchart showing an operation example of the processing device 500.

[0068] In a second embodiment of the present disclosure, a processing device 500 will be described, which is an information processing device that determines a method of class division in learning data used in machine learning. The processing device 500 determines a method of class division for machine learning that is capable of classifying into ordered classes. FIG. 13 shows an example of the hardware configuration of the processing device 500. Referring to FIG. 13, the processing device 500 has, as an example, the following hardware configuration. CPU (Central Processing Unit) 501 (arithmetic device) ROM (Read Only Memory) 502 (storage device) RAM (Random Access Memory) 503 (storage device) Programs 504 loaded into RAM 503 Storage device 505 for storing the programs 504 Drive device 506 for reading and writing data from and to a storage medium 510 external to the information processing device Communication interface 507 for connecting to a communication network 511 external to the information processing device Input / output interface 508 for inputting and outputting data Bus 509 for connecting the various components

[0069] 14 by the CPU 501 acquiring the program group 504 and executing it. The program group 504 is stored in advance in the storage device 505 or the ROM 502, for example, and is loaded into the RAM 503 or the like by the CPU 501 for execution as needed. The program group 504 may be supplied to the CPU 501 via the communication network 511, or may be stored in advance in the recording medium 510, and the drive device 506 may read out the program and supply it to the CPU 501.

[0070] 13 shows an example of the hardware configuration of the processing device 500. The hardware configuration of the processing device 500 is not limited to the above-described case. For example, the processing device 500 may be configured with only a part of the above-described configuration, such as excluding the drive device 506. Furthermore, the CPU 501 may be a GPU or the like exemplified in the first embodiment.

[0071] The acquisition unit 521 acquires region-of-interest information indicating a region of interest in the learning data used when performing machine learning to classify classes with an order. For example, the acquisition unit 521 can acquire, as the region-of-interest information, information indicating a value or region of interest among values ​​inferred by the learned model.

[0072] The class division calculation unit 522 determines how to divide the training data into classes, depending on the region of interest information acquired by the acquisition unit 521 .

[0073] The above is an example of the configuration of the processing device 500. Next, an example of the operation of the processing device 500 will be described with reference to FIG.

[0074] 15 shows an example of the operation of the processing device 500. Referring to Fig. 15, the acquisition unit 521 acquires region-of-interest information indicating a region of interest in learning data used when performing machine learning to classify ordered classes (step S201).

[0075] The class division calculation unit 522 determines how to divide the training data into classes, depending on the region of interest information acquired by the acquisition unit 521 (step S202).

[0076] The above is an example of the operation of the processing device 500.

[0077] As described above, the processing device 500 includes an acquisition unit 521 and a class division calculation unit 522. With this configuration, the class division calculation unit 522 can determine how to perform class division according to the region of interest information acquired by the acquisition unit 521. As a result, it becomes possible to differentiate the cost during learning depending on whether or not the region is a region of interest. This allows for more appropriate machine learning to be performed depending on whether or not the region is a region of interest when classifying into ordered classes.

[0078] The processing device 500 described above can be realized by incorporating a predetermined program into an information processing device such as the processing device 500. Specifically, a program according to another embodiment of the present disclosure is a program for realizing processing in which an information processing device acquires region-of-interest information indicating a region of interest in training data used when performing machine learning to classify ordered classes, and determines how to divide the training data into classes according to the acquired region-of-interest information.

[0079] Furthermore, a processing method executed by an information processing device such as the processing device 500 described above is a method of acquiring region of interest information indicating a region of interest in training data used when performing machine learning to classify ordered classes, and determining how to divide the training data into classes according to the acquired region of interest information.

[0080] Any program having the above-described configuration, or a computer-readable recording medium having the program recorded thereon, or a processing method, etc., can achieve the same functions and effects as the above-described processing device 500, and therefore can achieve the above-described objective of the present disclosure.

[0081] <Supplementary Notes> Part or all of the above-described embodiments can also be described as follows: Below, an outline of the processing device and the like in the present disclosure will be described. However, the present disclosure is not limited to the following configuration.

[0082] (Supplementary Note 1) A processing device comprising: an acquisition unit that acquires region-of-interest information indicating a region of interest in training data used when performing machine learning to classify data into ordered classes; and a class division calculation unit that determines how to divide the training data into classes in accordance with the region-of-interest information acquired by the acquisition unit. (Supplementary Note 2) The processing device according to Supplementary Note 1, wherein the class division calculation unit determines, in accordance with the region-of-interest information, to merge multiple classes corresponding to regions determined to be of no interest. (Supplementary Note 3) The processing device according to Supplementary Note 1 or Supplementary Note 2, wherein the class division calculation unit determines, in accordance with the region-of-interest information, to make the width of classes corresponding to regions determined to be of interest narrower than the width of classes corresponding to regions determined to be of no interest. (Supplementary Note 4) The processing device according to any one of Supplementary Notes 1 to 3, wherein the acquisition unit acquires guideline information that indicates a guideline for performing class division together with the region-of-interest information, and the class division calculation unit determines how to divide the training data into classes in accordance with the guideline information and the region-of-interest information. (Supplementary Note 5) The processing device according to Supplementary Note 4, wherein the class division calculation unit determines, in a case where the guideline information indicates a cost priority, to merge multiple classes corresponding to regions determined not to be of interest, in accordance with the region of interest information. (Supplementary Note 6) The processing device according to Supplementary Note 4 or Supplementary Note 5, wherein, in a case where the guideline information indicates an accuracy priority, the class division calculation unit determines, in a case where the guideline information indicates an accuracy priority, to make a width of classes corresponding to regions determined to be of interest narrower than a width of classes corresponding to regions determined not to be of interest, in accordance with the region of interest information. (Supplementary Note 7) The processing device according to any one of Supplementary Notes 1 to 6, wherein the class division calculation unit determines the number of learning data to select for each class, in accordance with the region of interest information.(Supplementary Note 8) The processing device according to any one of Supplementary Notes 1 to 7, comprising: a selection unit that selects training data to be used for machine learning in accordance with a decision made by the class division calculation unit; and a training unit that performs machine learning using the training data selected by the selection unit. (Supplementary Note 9) The processing device according to any one of Supplementary Notes 1 to 8, wherein the acquisition unit acquires, as the region of interest information, information indicating a service level requirement that specifies, for example, that an optical signal-to-noise ratio value be equal to or greater than a predetermined value. (Supplementary Note 10) A processing method in which an information processing device acquires region of interest information indicating a region of interest in training data used when performing machine learning to classify into ordered classes, and determines a method of class division in the training data in accordance with the acquired region of interest information. (Supplementary Note 11) A computer-readable recording medium having recorded thereon a program for implementing in an information processing device a process of acquiring region of interest information indicating a region of interest in training data used when performing machine learning to classify into ordered classes, and determining a method of class division in the training data in accordance with the acquired region of interest information.

[0083] Note that some or all of the configurations described in Supplementary Notes 2 to 9 that are dependent on the processing device described in Supplementary Note 1 may also be dependent in a similar dependent relationship on the processing method described in Supplementary Note 10 and the program (recording medium) described in Process 11. Furthermore, not limited to Supplementary Notes 10 and 11, some or all of the configurations described as Supplements may be similarly made dependent on various hardware, software, various recording means for recording software, or systems within the scope of the above-described embodiments.

[0084] The programs described in the above embodiments and appendices may be stored in a storage device or a computer-readable recording medium, such as a portable medium such as a flexible disk, an optical disk, a magneto-optical disk, or a semiconductor memory.

[0085] Although the present disclosure has been described above with reference to the above-described embodiments, the present disclosure is not limited to the above-described embodiments. Various modifications that can be understood by those skilled in the art can be made to the configuration and details of the present disclosure within the scope of the present disclosure. Furthermore, each embodiment can be combined with other embodiments as appropriate.

[0086] 100 Learning system 200 WDM terminal device 210 Demultiplexer 220 Receiver 221 OE conversion unit 222 AD conversion unit 223 DSP 230 Received light data collection unit 240 Learning data information 250 Trained model 300 Learning device 310 Operation input unit 320 Screen display unit 330 Communication I / F unit 140 Memory unit 341 Definition information 342 Class division policy 343 Service level requirement information 344 Program 350 Arithmetic processing unit 351 Division requirement information acquisition unit 352 Class division calculation unit 353 Learning data selection unit 354 Model learning unit 355 Output unit 400 Learning instruction device 500 Processing device 501 CPU 502 ROM 503 RAM 504 Program group 505 Storage device 506 Drive device 507 Communication interface 508 Input / output interface 509 Bus 510 Recording medium 511 Communication network 521 Acquisition unit 522 Class division calculation unit

Claims

1. A processing device having: an acquisition unit that acquires area of ​​interest information indicating an area of ​​interest in learning data used when performing machine learning to classify classes with order; and a class division calculation unit that determines how to divide the learning data into classes in accordance with the area of ​​interest information acquired by the acquisition unit.

2. A processing device according to claim 1, wherein the class division calculation unit determines, in accordance with the area of ​​interest information, to merge a plurality of classes corresponding to areas determined to be uninterested.

3. A processing device according to claim 1, wherein the class division calculation unit determines, in accordance with the area of ​​interest information, that the width of the class corresponding to an area determined to be of interest is narrower than the width of the class corresponding to an area determined not to be of interest.

4. A processing device as described in claim 1, wherein the acquisition unit acquires, together with the area of ​​interest information, guideline information indicating a guideline for performing class division, and the class division calculation unit determines how to perform class division in the learning data according to the guideline information and the area of ​​interest information.

5. A processing device according to claim 4, wherein the class division calculation unit determines, when the guideline information indicates cost priority, to merge multiple classes corresponding to areas determined to be unfocused according to the area of ​​interest information.

6. A processing device according to claim 4, wherein, when the guideline information indicates that priority is given to accuracy, the class division calculation unit determines, in accordance with the area of ​​interest information, that the width of the class corresponding to an area determined to be of interest is to be narrower than the width of the class corresponding to an area determined not to be of interest.

7. A processing device according to claim 1, wherein the class division calculation unit determines the number of learning data to be selected for each class according to the region of interest information.

8. A processing device according to claim 1, comprising: a selection unit that selects learning data to be used for machine learning in accordance with a decision made by the class division calculation unit; and a learning unit that performs machine learning using the learning data selected by the selection unit.

9. A processing device according to claim 1, wherein the acquisition unit acquires, as the area of ​​interest information, information indicating a service level requirement, which specifies, for example, that an optical signal-to-noise ratio value is equal to or greater than a predetermined value.

10. A processing method in which an information processing device acquires area of ​​interest information indicating an area of ​​interest in learning data used when performing machine learning to classify classes with order, and determines how to divide the learning data into classes according to the acquired area of ​​interest information.

11. A computer-readable recording medium having recorded thereon a program for implementing a process in an information processing device, which comprises: acquiring area of ​​interest information indicating an area of ​​interest in training data used in machine learning to classify classes with order; and determining how to divide the training data into classes according to the acquired area of ​​interest information.

Images