WO2019176087A1 - 学習装置および学習方法 - Google Patents
学習装置および学習方法 Download PDFInfo
- Publication number
- WO2019176087A1 WO2019176087A1 PCT/JP2018/010446 JP2018010446W WO2019176087A1 WO 2019176087 A1 WO2019176087 A1 WO 2019176087A1 JP 2018010446 W JP2018010446 W JP 2018010446W WO 2019176087 A1 WO2019176087 A1 WO 2019176087A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- learning
- neural network
- data
- unit
- reference label
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/217—Validation; Performance evaluation; Active pattern learning techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/2431—Multiple classes
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Definitions
- the present invention relates to a learning apparatus and a learning method for learning a discriminator using a neural network.
- learning image data used for detecting an abnormality is generally determined in units of images as normal (hereinafter referred to as “OK”) or abnormal (hereinafter referred to as “NG”).
- OK normal
- NG abnormal
- an image region having an NG element is often local, and most of the regions are not locally different from the learning image data determined to be OK.
- Patent Document 1 describes a technique in which an input image is divided into a plurality of image areas, and a label for distinguishing the foreground and the background is assigned to each divided image area. In this technique, based on continuity from an image region to which a label with a fixed foreground or background is assigned, it is estimated whether a pixel with an indeterminate label is a foreground or a background.
- This invention solves the said subject, and aims at obtaining the learning apparatus and learning method which can provide the discriminator which performs highly accurate discrimination.
- the learning device includes a learning unit and a dynamic update unit.
- the learning unit inputs learning data composed of a plurality of elements and group information defined to determine that at least one of the data belonging to the group is determined to be NG, and an OK label is provided for each element of the learning data Alternatively, by using a learning reference label to which an NG label is assigned and learning a neural network using a plurality of elements of learning data as input, at least one of the data belonging to the group is determined to be NG using the learning result neural network. Is output.
- the dynamic update unit dynamically updates the learning reference label during learning of the neural network by the learning unit.
- the learning unit generates an initial value of the learning reference label for each element of the learning data using the learning data and the group information, and the learning updated from the initial value of the learning reference label by the dynamic update unit
- the neural network is iteratively learned using the reference labels sequentially, and an NG index value that can be NG as an element of the learning data is calculated using the output value of the neural network at each learning iteration.
- the dynamic updating unit updates the learning reference label based on a result of statistically testing the change tendency of the NG index value obtained for each transition of the number of learning iterations.
- the learning device can provide a discriminator that performs high-accuracy discrimination by determining at least one of the data belonging to the group as NG.
- FIG. 2 is a block diagram illustrating a detailed configuration of a learning device according to Embodiment 1.
- FIG. It is a figure which shows an example of the content of data DB for learning. It is a figure which shows an example of the content of NG determination group DB. It is a figure which shows an example of the content of the learning reference label buffer. It is a figure which shows an example of a structure of a neural network. It is a figure which shows an example of the content of the NG index value buffer. It is a figure which shows an example of the content of the log
- FIG. 10A is a block diagram illustrating a hardware configuration that implements the function of the learning device according to the first embodiment.
- FIG. 10B is a block diagram illustrating a hardware configuration for executing software that implements the functions of the learning device according to Embodiment 1.
- 3 is a flowchart illustrating a learning method according to the first embodiment. It is a flowchart which shows a learning reference
- FIG. 1 is a block diagram showing a configuration of a learning device 1 according to Embodiment 1 of the present invention.
- the learning device 1 includes learning data input from a learning data database (hereinafter referred to as learning data DB) 2 and an NG determination group input from an NG determination group database (hereinafter referred to as NG determination group DB) 3.
- learning data DB learning data database
- NG determination group DB NG determination group database
- the learning data stored in the learning data DB 2 is data composed of a plurality of elements. For example, when the learning data is image data, each image region obtained by dividing the image indicated by the image data into a plurality of elements is an element.
- the NG determination group information stored in the NG determination group DB 3 is information that defines that at least one of the data belonging to the NG determination group is determined to be NG.
- the learning device 1 includes a learning unit 10 and a dynamic update unit 11.
- the learning unit 10 inputs learning data from the learning data DB 2, inputs NG determination group information from the NG determination group DB 3, and uses a learning reference label for each element of the learning data, to Learning a neural network using elements as inputs.
- the learning reference label is information in which an OK label or an NG label is assigned to each element of the learning data.
- the learning unit 10 outputs a neural network parameter that provides a discriminator that determines that at least one of the data belonging to the NG determination group is NG.
- the learning unit 10 generates an initial value of the learning reference label for each element of the learning data using the learning data and the NG determination group information, and the dynamic update unit 11 operates the dynamic reference unit from the initial value of the learning reference label.
- the learning network is repetitively learned by sequentially using the updated learning reference labels.
- the learning unit 10 uses the output value of the neural network for each learning iteration, calculates an index value (hereinafter referred to as an NG index value) in which the element of the learning data can be NG.
- the dynamic update unit 11 dynamically updates the learning reference label during the learning of the neural network by the learning unit 10. Specifically, the dynamic update unit 11 statistically tests the change tendency of the NG index value obtained for each transition of the number of learning iterations, and updates the learning reference label based on the statistical test result. . The dynamic updating unit 11 updates the NG label assigned to the element that is originally OK in the learning reference label to the OK label by using the statistical test result of the change tendency of the NG index value.
- FIG. 2 is a block diagram showing a detailed configuration of the learning device 1 according to the first embodiment.
- the learning unit 10 includes a learning reference label buffer 10a, a learning reference label generation unit 10b, a neural network management unit 10c, a neural network update unit 10d, and an NG index value buffer 10e.
- the dynamic update unit 11 includes a history buffer 11a, an NG candidate buffer 11b, a history management unit 11c, a test unit 11d, and a learning reference label update unit 11e.
- the neural network initial parameter is a parameter that defines the initial state of the neural network.
- the neural network initial parameters include the number of hidden unit nodes in the neural network, weight parameters assigned between the nodes, bias parameters, and initial values of learning rate parameters. These parameters are those described in Reference Document 1 below.
- the learning device 1 updates the neural network initial parameter to generate a neural network parameter that provides a discriminator that determines that at least one of the data belonging to the NG determination group is abnormal. (Reference 1) C.I. M.M. Bishop, Hiroshi Motoda (supervised), “Pattern Recognition and Machine Learning”, Maruzen Publishing, pp. 225-247.
- the learning data DB 2 stores learning data composed of a plurality of elements.
- FIG. 3 is a diagram illustrating an example of the contents of the learning data DB2.
- the learning data is composed of a plurality of elements identified by “data ID”, and each of the plurality of elements is composed of “data value series”.
- FIG. 3 shows an element with data ID D0001, an element with data ID D0002, an element with data ID D0003, an element with data ID D0004,. .
- the NG determination group DB 3 stores NG determination group information including identification information of an NG determination group and identification information of elements of learning data belonging to the NG determination group.
- FIG. 4 is a diagram illustrating an example of the contents of the NG determination group DB3.
- Each of the plurality of elements constituting the learning data is identified by “data ID” as in FIG.
- the NG determination group is identified by “NG determination group ID”.
- the element with the data ID D0001, the element with the data ID D0002, and the element with the data ID D0004 belong to the NG determination group with the NG determination group ID G001, and at least one of these elements is determined to be NG.
- the learning standard label buffer 10a stores a learning standard label for each element of the learning data.
- FIG. 5 is a diagram illustrating an example of the contents of the learning reference label buffer 10a.
- the learning reference label information includes a data ID of an element of learning data, an OK flag, and an NG flag.
- Each of the plurality of elements constituting the learning data is identified by “data ID” as in FIG.
- a value 1 is set in the OK flag corresponding to this element, and a value 0 is set in the NG flag corresponding to this element.
- a value 0 is set in the OK flag corresponding to this element, and a value 1 is set in the NG flag corresponding to this element.
- the learning reference label generation unit 10b generates an initial value of the learning reference label based on the determination result of whether or not the learning data belongs to the NG determination group. For example, the learning reference label generation unit 10b determines whether the same data ID as the element belonging to the NG determination group in the NG determination group DB3 exists among the data IDs of the elements registered in the learning data DB2. . When it is determined that the data ID is the same as the element belonging to the NG determination group, the learning reference label generation unit 10b sets the value 0 to the OK flag corresponding to this element and sets the value 1 to the NG flag. A label is generated and stored in the learning reference label buffer 10a.
- the neural network management unit 10c converts the element corresponding to the data ID stored in the learning data DB2 into a flag having a value of 1 among the OK flag and the NG flag corresponding to the data ID in the learning reference label buffer 10a. It manages the learning of the neural network that distributes to the corresponding class (OK class or NG class).
- FIG. 6 is a diagram illustrating an example of the configuration of a neural network. As shown in FIG. 6, the neural network includes an input layer, a hidden layer, and an output layer.
- the sequence of data values x 1, x 2 which constitutes an element of learning data, ..., the x N is input to the input layer, the weight parameter and x 1 between the nodes of the node of the input layer and the hidden layer , x 2, ⁇ , z 1 , z 2 which is produced by the non-linear converting the multiplication sum of the x N, ⁇ , z M is calculated. Subsequently, y 1 and y 2 obtained by nonlinearly transforming the multiplication sum of the weight parameter between the hidden layer node and the output layer node and z 1 , z 2 ,..., Z M are output from the output layer.
- the output values y 1 and y 2 are NG evaluation values for each class, the output value y 1 represents the degree to which the element belongs to the OK class, and the output value y 2 represents the degree to which the element belongs to the NG class. Yes.
- An element (data value series) input to the neural network is determined to belong to the OK class when y 1 > y 2 , and is determined to belong to the NG class when y 1 ⁇ y 2 .
- the NG index value may be a ratio of y 2 and y 1 corrected so that the denominator does not become 0.
- y 2 ⁇ y 1 is set as the NG index value.
- FIG. 6 shows a neural network having one hidden layer
- the neural network managed by the neural network management unit 10c may be a neural network having a plurality of hidden layers.
- the neural network update unit 10d updates the neural network parameter based on the NG index value obtained by the neural network management unit 10c, thereby outputting a neural network parameter that gives a discriminator of the learning result.
- the neural network update unit 10d updates the neural network parameters using the parameter optimization method described in Reference Document 1.
- the neural network updating unit 10d outputs the neural network parameters optimized and updated to the outside of the learning device 1 when the number of learning iterations reaches a threshold value or when the learning end condition is satisfied. To do.
- the NG index value buffer 10e stores an NG index value for each element of learning data obtained in the process of learning iteration of the neural network.
- FIG. 7 is a diagram illustrating an example of the contents of the NG index value buffer 10e.
- the data ID of the element of the learning data and the NG index value at the time point calculated by the neural network management unit 10c among the NG index values of the element are set.
- the history buffer 11a stores the number of learning iterations and the corresponding NG index value for each element of the learning data.
- FIG. 8 is a diagram illustrating an example of the contents of the history buffer 11a.
- NG index values when the number of learning iterations is 1, 100 and 200 are set.
- the NG index value obtained when the number of learning iterations is 1 is 0.55
- the NG index value obtained when the number of learning iterations is 100 is It is 0.35
- the NG index value obtained when the number of learning iterations is 200 is 0.12.
- FIG. 9 is a diagram illustrating an example of the contents of the NG candidate buffer 11b.
- a data ID of an element of learning data, an NG candidate level, an OK determination flag, and an NG determination flag are set.
- the NG candidate level changes according to the value of the NG index value taken in the past.
- the NG candidate is an element to which the NG label is assigned as the learning reference label, and the NG candidate level is a level corresponding to the NG index value corresponding to the element of the NG candidate.
- the OK confirmation flag is a flag indicating that the element is confirmed to be OK. For example, if the value 0 is set in the OK determination flag, this indicates that the element corresponding to the OK determination flag is not determined OK, and if the value 1 is set, the element corresponds to this OK determination flag. This indicates that the element to be confirmed is OK.
- the NG determination flag is a flag indicating that the element is determined to be NG. For example, if the value 0 is set in the NG determination flag, it indicates that the element corresponding to the NG determination flag is not determined as NG. If the value 1 is set, the element corresponds to the NG determination flag. This indicates that the element to be determined is NG.
- the history management unit 11c stores the NG index value for each element of the learning data obtained for each learning iteration in the history buffer 11a.
- the history management unit 11c stores the NG index value stored in the NG index value buffer 10e in the history buffer 11a together with the number of learning iterations at that time.
- the history buffer 11a stores the NG index value obtained at the preset learning repetition count (1, 100, 200 in the example of FIG. 8).
- the test unit 11d statistically tests the change tendency of the NG index value stored in the history buffer 11a, and updates the contents of the NG candidate buffer 11b for each element of the learning data based on the statistical test result. . For example, the test unit 11d extracts, from the history buffer 11a, a time series of NG index values for each transition of the number of learning iterations corresponding to the processing target element. Next, the test unit 11d statistically tests the time series of NG index values extracted from the history buffer 11a, and confirms the change tendency of the NG index values.
- the test unit 11d performs irregular samples on the time series of NG index values. When it is confirmed by the irregularity test that the change in the NG index value is not regular, the testing unit 11d leaves this element as an NG candidate. When it is confirmed that there is regularity, the testing unit 11d performs a trend test on a time series of NG index values. If it is confirmed by the tendency test that the NG index value is increasing, the test unit 11d keeps this element as an NG candidate.
- the test unit 11d determines that this element is NG confirmed, and the NG confirmation flag corresponding to this element in the NG candidate buffer 11b. Set the value 1 to.
- the test unit 11d determines that this element is OK and determines that this element in the NG candidate buffer 11b corresponds to this element. Set the value 1 to the confirmation flag.
- the test method described in Reference 2 below may be used. (Reference 2) Yusuke Muto, “Statistical Analysis Handbook”, Asakura Shoten, pp. 398-399, 402-403.
- the learning reference label update unit 11e updates the learning reference label of the element of the learning data that has been determined OK by the test unit 11d from the NG label to the OK label. For example, when the element of the learning data is confirmed OK by the test unit 11d, the learning reference label update unit 11e sets the value 1 to the OK flag in the learning reference label corresponding to this element, and sets the value NG to 0. Set.
- the learning device 1 includes a processing circuit for executing processing from step ST1 to step ST5 in FIG.
- This processing circuit may be dedicated hardware, or may be a CPU (Central Processing Unit) that executes a program stored in a memory.
- CPU Central Processing Unit
- FIG. 10A is a block diagram showing a hardware configuration for realizing the function of the learning device 1.
- FIG. 10B is a block diagram illustrating a hardware configuration that executes software that implements the functions of the learning device 1.
- an information input interface 100 is an interface that relays input of information from an external device (not shown) to the learning device 1 shown in FIGS.
- the neural network management unit 10 c illustrated in FIG. 2 inputs the neural network initial parameters from the external device via the information input interface 100.
- the external device may be a storage device provided independently of the learning device 1.
- the learning device 1 may use a storage device that exists on the cloud.
- the DB input / output interface 101 is an interface that relays data exchange between the learning device 1 and the database 102.
- the database 102 is the learning data DB2 and the NG determination group DB3 shown in FIGS.
- the learning reference label generation unit 10 b included in the learning unit 10 inputs data from the learning data DB 2 and the NG determination group DB 3 via the DB input / output interface 101.
- the database 102 may be constructed in a storage device provided independently of the learning device 1.
- the learning device 1 inputs data from the database 102 constructed in the storage device existing in the cloud via the DB input / output interface 101.
- the learning reference label buffer 10a, the NG index value buffer 10e, the history buffer 11a, and the NG candidate buffer 11b shown in FIG. 2 are constructed in the memory built in the processing circuit 104 shown in FIG. 10A, or as shown in FIG. 10B. It is constructed in the memory 106 shown.
- the information output interface 103 is an interface that relays the output of information from the learning device 1 to an external device (not shown).
- the neural network update unit 10 d illustrated in FIG. 2 outputs the neural network parameters to the external device via the information output interface 103.
- An example of the external device is an abnormality detection device that uses a discriminator given by a neural network parameter.
- the processing circuit 104 may be, for example, a single circuit, a composite circuit, a programmed processor, a parallel programmed processor, or an ASIC (Application Specific Integrated). Circuit), FPGA (Field-Programmable Gate Array), or a combination thereof.
- the functions of the learning unit 10 and the dynamic update unit 11 in the learning device 1 may be realized by separate processing circuits, or these functions may be realized by a single processing circuit.
- the processing circuit is the processor 105 shown in FIG. 10B
- the functions of the learning unit 10 and the dynamic update unit 11 in the learning device 1 are realized by software, firmware, or a combination of software and firmware.
- the software or firmware is described as a program and stored in the memory 106.
- the processor 105 reads out and executes the program stored in the memory 106, thereby realizing the functions of the learning unit 10 and the dynamic update unit 11 in the learning device 1. That is, the learning device 1 includes a memory 106 for storing a program that, when executed by the processor 105, results in the processing from step ST1 to step ST5 shown in FIG. These programs cause the computer to execute the procedures or methods of the learning unit 10 and the dynamic update unit 11.
- the memory 106 may be a computer-readable storage medium that stores a program for causing a computer to function as the learning unit 10 and the dynamic update unit 11.
- the memory 106 includes, for example, a nonvolatile or volatile semiconductor such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically-EPROM), or the like. Magnetic disks, flexible disks, optical disks, compact disks, mini disks, DVDs, and the like are applicable.
- RAM Random Access Memory
- ROM Read Only Memory
- flash memory flash memory
- EPROM Erasable Programmable Read Only Memory
- EEPROM Electrically-EPROM
- Part of the functions of the learning unit 10 and the dynamic update unit 11 may be realized by dedicated hardware, and part of the functions may be realized by software or firmware.
- the learning unit 10 realizes the function by a processing circuit that is dedicated hardware
- the dynamic update unit 11 realizes the function by the processor 105 reading and executing a program stored in the memory 106.
- the processing circuit can realize the above functions by hardware, software, firmware, or a combination thereof.
- FIG. 11 is a flowchart illustrating the learning method according to the first embodiment.
- the learning reference label generation unit 10b sets a value 0 to an OK flag corresponding to an element belonging to the NG determination group among a plurality of elements of the learning data, and sets a value 1 to the NG flag. Then, a learning reference label is generated for each element of the learning data.
- the neural network management unit 10c constructs an initial state neural network using the neural network initial parameters.
- step ST2 the neural network management unit 10c inputs learning data elements to the neural network, and calculates an NG index value for each element based on the output result of the neural network.
- the neural network management unit 10c determines whether the target element belongs to the OK class or the NG class based on the NG index value.
- the neural network management unit 10c evaluates the error of the initial value of the learning reference label corresponding to the target element and the learning reference label updated by the dynamic update unit 11 and the determined class using an error function. To do.
- the function described in Reference 1 may be used as the error function.
- the history management unit 11c stores the NG index value obtained each time learning is repeated in the history buffer 11a.
- the test unit 11d evaluates whether or not the NG candidate element should be confirmed as OK or NG based on the result of statistically testing the change tendency of the NG index value stored in the history buffer 11a.
- the learning reference label updating unit 11e updates the learning reference label based on the evaluation result by the testing unit 11d. For example, the learning reference label update unit 11e updates the learning reference label corresponding to the element determined to be OK by the test unit 11d from the NG label to the OK label.
- step ST4 the neural network update unit 10d updates the neural network parameters based on the NG index value obtained by the neural network management unit 10c.
- a series of learning processes from step ST2 to step ST4 are repeatedly performed for each of a plurality of elements constituting the learning data.
- the neural network update unit 10d outputs the neural network parameters updated by the series of learning described above as a learning result (step ST5).
- FIG. 12 is a flowchart showing learning reference label generation processing, and shows processing for generating an initial value of a learning reference label.
- the learning reference label generation unit 10b sequentially reads out the data ID for each element of the learning data registered in the learning data DB 2 and executes the subsequent processing.
- the learning reference label generation unit 10b determines whether there is the same data ID as the data ID of the element belonging to the NG determination group registered in the NG determination group DB3 among the data IDs of the elements read from the learning data DB2. Determine (step ST1a).
- the learning reference label generation unit 10b sets the value 0 to the OK flag corresponding to the target element. And a value 1 is set in the NG flag (step ST2a).
- the learning reference label generation unit 10b sets a value in the OK flag corresponding to the target element. 1 is set, and a value 0 is set in the NG flag (step ST3a).
- Step ST4a When the learning reference label generation unit 10b sets values in the OK flag and the NG flag corresponding to the target element, the learning reference label generation unit 10b stores the data ID of the target element and the values of the OK flag and the NG flag in the learning reference label buffer 10a.
- the learning reference label generation unit 10b repeatedly executes a series of processes from step ST1a to step ST4a for all elements of the learning data registered in the learning data DB2. By this process, an initial value of the learning reference label is generated.
- FIG. 13 is a flowchart showing a neural network management process, and shows a process for calculating an NG index value for each element.
- the neural network management unit 10c constructs a neural network in an initial state based on the neural network initial parameters (step ST1b).
- the neural network management unit 10c inputs the learning data elements (data value series D i ) read from the learning data DB2 to the neural network, and outputs the output values y 1 — i calculated in the output layer and y2_i is obtained (step ST2b).
- Output value y 1_I represents the degree to which elements belong to the OK class
- the output value y 2_I represents the degree to which elements belong to NG class.
- An element (data value series D i ) input to the neural network is determined to belong to the OK class when y 1 — i > y 2 — i , and is determined to belong to the NG class when y 1 — i ⁇ y 2 — i.
- the neural network management unit 10c evaluates the error between the determined class and the learning reference label.
- step ST3b the neural network management unit 10c calculates y 2 — i ⁇ y 1 — i as the NG index value NGE i , and stores the calculated NG index value NGE i in the NG index value buffer 10e together with the data ID of the corresponding element.
- the y 2_i -y 1_i and NGE i may be the ratio of y 1_I and y 2_I as NGE i, the distance in consideration of distribution of y 1_I and y 2_I may NGE i.
- the neural network management unit 10c repeatedly executes a series of processes from step ST2b to step ST3b for all elements of the learning data registered in the learning data DB2.
- FIG. 14 is a flowchart showing the history management process of the NG index value.
- the history management unit 11c determines whether or not the number of learning iterations is divisible by a certain number (for example, 100) (step ST1c).
- the number of learning iterations is the number of times that a series of processing by the learning unit 10 from step ST2 to step ST4 shown in FIG. 11 is repeated.
- step ST1c When it is determined that the number of learning iterations is divisible by a certain number (step ST1c; YES), the history management unit 11c extracts a data ID and a series of NG index values of elements corresponding thereto from the NG index value buffer 10e (step). ST2c). Next, the history management unit 11c adds the current number of learning iterations to the extracted data ID and NG index value series, and stores it in the history buffer 11a (step ST3c). Thereafter, the history management unit 11c ends the process of FIG. 14 and proceeds to the process of the verification unit 11d.
- step ST1c when it is determined that the number of learning iterations is not divisible by a certain number (step ST1c; NO), the history management unit 11c ends the process of FIG. 14 and proceeds to the process of the neural network update unit 10d.
- FIG. 15 is a flowchart showing the NG index value statistical test process.
- the test unit 11d extracts, from the history buffer 11a, an NG index value series V j for each number of learning iterations corresponding to the data ID of the j-th element (target element), and the extracted NG index value series V j is converted into a series L j of NG candidate levels (step ST1d).
- the sequence L j is determined depending on what number range obtained by equally dividing a range of values taken by the sequence V j in the past by a certain number to which the current NG index value belongs.
- NG candidate level 0 corresponds to an intermediate range of NG index values taken in the past.
- the NG candidate level + 1 corresponds to a range where the NG index value is larger than the intermediate range
- the NG candidate level + 2 corresponds to a range where the NG index value is larger than the range corresponding to the NG candidate level + 1.
- NG candidate level-1 corresponds to a range having an NG index value smaller than the intermediate range
- NG candidate level-2 corresponds to a range having an NG index value smaller than the range corresponding to NG candidate level-1. Note that since there is only one NG index value for the first time, the range of values that the series V j has taken in the past is zero. The NG candidate level at this time is set to 0.
- the verification unit 11d determines whether or not the value 1 is set in the OK confirmation flag corresponding to all the elements other than the target element among the elements belonging to the NG determination group to which the target element (NG candidate) belongs. Determination is made (step ST2d). At this time, when the value 0 is set in the OK confirmation flag corresponding to any element (step ST2d; NO), the test unit 11d refers to the NG candidate buffer 11b and determines NG corresponding to the target element. It is determined whether or not the value 0 is set in both the confirmation flag and the OK confirmation flag (step ST3d).
- the test unit 11d uses the current number of learning iterations and the data corresponding to the target element in the NG candidate buffer 11b. Is updated with the current NG candidate level of the element, the current OK confirmation flag and the NG confirmation flag of the target element (step ST4d). Thereafter, the test unit 11d continuously extracts the sequence V j + 1 of NG index values corresponding to the data ID of the (j + 1) th element from the history buffer 11a, and repeats the processing from step ST1d.
- step ST2d When the value 1 is set in the OK confirmation flag corresponding to all the elements other than the target element among the elements belonging to the NG determination group to which the target element belongs (step ST2d; YES), the test unit 11d The value to be set in the NG confirmation flag corresponding to the target element is determined to be “1” (step ST5d). Thereafter, the process proceeds to step ST4d, and the verification unit 11d updates the NG determination flag corresponding to the target element in the NG candidate buffer 11b by setting the value 1 to the value.
- Step ST3d If the value 0 in any of the NG determination flag and OK confirmation flag corresponding to the elements of the target is set (Step ST3d; YES), the assay unit 11d, a certain confidence against sequence L j of NG candidate level An irregularity test based on a coefficient (for example, 0.95) is performed (step ST6d). Test unit 11d, based on the non-regularity of the test results for the sequence L j of NG candidate levels, determines whether there is regularity in the sequence L j of NG candidate level (step ST7d).
- step ST7d When the regularity is not confirmed in the NG candidate level sequence L j and it is determined that the change tendency of the NG index value changes randomly each time learning is repeated (step ST7d; NO), the test unit 11d The target element is not confirmed OK and NG is not confirmed. Thereafter, the process proceeds to step ST4d.
- test unit 11d An upward tendency test is performed on the level series L j (step ST8d). Assay portion 11d is based on the tendency of the test results for the sequence L j of NG candidate levels, it determines whether there is an upward trend in the series L j of NG candidate level (step ST9d).
- step ST9d When the NG candidate level series L j tends to increase (step ST9d; YES), the test unit 11d determines that the NG index value increases every time learning is repeated, and proceeds to the processing of step ST5d. As a result, the value to be set in the NG determination flag corresponding to the target element is determined to be “1”, and NG is determined.
- test unit 11d performs a downward trend test for the NG candidate level series L j .
- Assay portion 11d is based on the tendency of the test results for the sequence L j of NG candidate levels, it determines whether there is a downward trend in the series L j of NG candidate level (step ST11d).
- test unit 11d proceeds to the process of step ST4d and determines the value of the series L j as the current NG candidate level. And reflected in the NG candidate buffer 11b.
- step ST11d determines that the NG index value decreases each time learning is repeated, and corresponds to the target element.
- the value to be set in the OK confirmation flag is determined as “1” (step ST12d). Thereafter, the process proceeds to step ST4d, and the testing unit 11d updates the OK determination flag corresponding to the target element in the NG candidate buffer 11b by setting a value of 1.
- FIG. 16 is a flowchart showing the learning reference label update process.
- the learning reference label update unit 11e extracts, from the NG candidate buffer 11b, the data ID of the element whose value is set to 1 in the OK confirmation flag as a series U having no overlap (step ST1e).
- the series U includes the data ID of the element whose value is set to 1 in the OK confirmation flag.
- the learning reference label update unit 11e specifies a record having the same data ID as the data ID included in the series U in the learning reference label buffer 10a, and the OK flag in the specified record has the value 1 and the NG flag has the value.
- the learning reference label is updated to be 0 (step ST2e). Since this processing is executed during learning of the neural network by the learning unit 10, the learning reference label is dynamically updated.
- FIG. 17 is a flowchart showing neural network parameter update processing.
- the neural network update unit 10d extracts, from the NG index value buffer 10e, the NG index value calculated from the output result of the neural network constructed with the current neural network parameters by the neural network management unit 10c.
- the neural network update unit 10d Based on the NG index value extracted from the NG index value buffer 10e, the neural network update unit 10d optimizes the neural network parameters (step ST1f). For example, the neural network update unit 10d optimizes the current neural network parameters using the gradient descent optimization method described in Reference 1 based on the NG index value extracted from the NG index value buffer 10e. This optimization is performed sequentially each time learning is repeated.
- the neural network updating unit 10d determines whether or not the number of learning iterations has reached a predetermined value (for example, 10000) or more (step ST2f). When the number of learning iterations is less than the predetermined value (step ST2f; NO), the neural network update unit 10d ends the process of FIG. 17 and proceeds to the process from step ST2 of FIG. 11 by the neural network management unit 10c.
- a predetermined value for example, 10000
- the neural network update unit 10d outputs the neural network parameters at that time as a learning result (step ST3f).
- the neural network constructed using the neural network parameters of the learning result functions as a discriminator that determines at least one of the data belonging to the NG determination group as NG.
- NG determination group (1) ⁇ E11 (OK), E12 (NG), E13 (OK) ⁇
- NG determination group (2) ⁇ E21 (NG ), E22 (OK), E23 (OK) ⁇
- OK OK determination group (3) ⁇ E31 (OK), E32 (OK), E33 (OK) ⁇ .
- the NG determination group (1) only the second element is NG, and the remaining elements are OK.
- the NG determination group (2) only the first element is NG, and the remaining elements are OK.
- the entire group is determined to be NG.
- the element unit OK or NG is shown in () of the element E, but OK and NG for each element are unknown at the stage where the three groups are treated as learning data.
- OK and NG of the group (4) ⁇ E41 (OK), E42 (OK), E43 (OK) ⁇ and the group (5) ⁇ E51 (OK), E52 (OK), E53 (NG) ⁇
- the conventional classifier an element that is originally OK is mistakenly learned as NG, and therefore, any element belonging to the group (4) may be determined as NG. If any of the elements belonging to the group (4) is determined as NG, the entire group (4) is also determined as NG.
- the conventional discriminator has low identification accuracy in element units and low discrimination accuracy in group units.
- the learning device 1 updates the learning reference label in which the OK label or the NG label is assigned for each element based on the result of statistically testing the change tendency of the NG index value. .
- the learning unit 10 can learn a discriminator that performs high-accuracy discrimination with almost no erroneous learning of elements that are originally OK as NG.
- the learning result classifier determines that at least one of the data belonging to the NG determination group is NG, if the group (5) is the NG determination group, the group (5) is appropriately determined as NG, Group (4) is determined to be OK.
- the learning device 1 includes the learning unit 10 and the dynamic update unit 11.
- the learning unit 10 learns the neural network using a plurality of elements of the learning data as an input using the learning reference label, and uses at least one of the data belonging to the group as NG using the neural network of the learning result.
- the discriminator to be judged is output.
- the dynamic update unit 11 dynamically updates the learning reference label during the learning of the neural network by the learning unit 10.
- the learning unit 10 repeatedly learns the neural network using the learning reference label updated by the dynamic update unit 11 from the initial value of the learning reference label, and outputs the neural network for each learning iteration.
- the NG index value corresponding to the element of the learning data is calculated using the value.
- the dynamic update unit 11 updates the learning reference label based on the result of statistically testing the change tendency of the NG index value obtained for each transition of the number of learning iterations. Since the learning result discriminator determines that at least one of the data belonging to the group is NG, the learning device 1 can provide a discriminator that performs high-accuracy discrimination.
- the learning device can provide a discriminator that performs high-precision discrimination, it can be used, for example, in an abnormality detection system.
- 1 learning device 2 learning data DB, 3 NG judgment group DB, 10 learning unit, 10a learning reference label buffer, 10b learning reference label generation unit, 10c neural network management unit, 10d neural network update unit, 10e NG index value buffer 11 dynamic update unit, 11a history buffer, 11b NG candidate buffer, 11c history management unit, 11d verification unit, 11e learning reference label update unit, 100 information input interface, 101 DB input / output interface, 102 database, 103 information output interface , 104 processing circuit, 105 processor, 106 memory.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Molecular Biology (AREA)
- General Health & Medical Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Neurology (AREA)
- Image Analysis (AREA)
Abstract
Description
一方、NGと判定された学習用画像データにおいて、NG要素のある画像領域は局所的であることが多く、大半の領域は、OKと判定された学習用画像データと局所的には変わらない。
例えば、特許文献1には、入力画像を複数の画像領域に分割し、分割した画像領域ごとに前景と背景とを区別するラベルを付与する技術が記載されている。この技術では、前景または背景が確定しているラベルが付与された画像領域からの連続性に基づいて、不確定のラベルが付与されたピクセルが前景であるか背景であるかが推定される。
実施の形態1.
図1は、この発明の実施の形態1に係る学習装置1の構成を示すブロック図である。
学習装置1は、学習用データデータベース(以下、学習用データDBと記載する)2から入力した学習用データとNG判定グループデータベース(以下、NG判定グループDBと記載する)3から入力したNG判定グループ情報とを用いて、ニューラルネットワーク初期パラメータを更新していくことにより、学習結果の識別器を与えるニューラルネットワークパラメータを生成する。
(参考文献1)C.M.ビショップ, 元田浩(監訳), “パターン認識と機械学習 上”, 丸善出版, pp. 225-247.
図6は、ニューラルネットワークの構成の一例を示す図である。図6に示すように、ニューラルネットワークは、入力層、隠れ層および出力層から構成される。学習用データの要素を構成するデータ値系列x1,x2,・・・,xNが入力層に入力されると、入力層のノードと隠れ層のノードとの間の重みパラメータとx1,x2,・・・,xNとの乗算和を非線形変換したz1,z2,・・・,zMが算出される。続いて、隠れ層のノードと出力層のノードとの間の重みパラメータとz1,z2,・・・,zMとの乗算和を非線形変換したy1,y2が出力層から出力される。
なお、図6において、隠れ層が1層のニューラルネットワークを示したが、ニューラルネットワーク管理部10cによって管理されるニューラルネットワークは、隠れ層が複数層のニューラルネットワークであってもよい。
NG候補は、学習基準ラベルでNGラベルが割り当てられた要素であり、NG候補レベルは、NG候補の要素に対応するNG指標値に応じたレベルとなる。
無規則性の検定および傾向性の検定の方法として、下記の参考文献2に記載された検定方法を使用してもよい。
(参考文献2)武藤眞介著,“統計解析ハンドブック”, 朝倉書店, pp. 398-399, 402-403.
すなわち、学習装置1は、プロセッサ105によって実行されるときに、図11に示すステップST1からステップST5までの処理が結果的に実行されるプログラムを記憶するためのメモリ106を備える。これらのプログラムは、学習部10および動的更新部11の手順または方法をコンピュータに実行させる。メモリ106は、コンピュータを、学習部10および動的更新部11として機能させるためのプログラムが記憶されたコンピュータ可読記憶媒体であってもよい。
図11は、実施の形態1に係る学習方法を示すフローチャートである。
ステップST1において、学習基準ラベル生成部10bが、学習用データの複数の要素のうち、NG判定グループに属する要素に対応するOKフラグに値0を設定し、NGフラグに値1を設定することにより、学習用データの要素ごとの学習基準ラベルを生成する。また、ニューラルネットワーク管理部10cは、ニューラルネットワーク初期パラメータを用いて、初期状態のニューラルネットワークを構築する。
学習の反復が終了すると、ニューラルネットワーク更新部10dは、前述した一連の学習で更新したニューラルネットワークパラメータを、学習結果として出力する(ステップST5)。
図12は、学習基準ラベル生成処理を示すフローチャートであって、学習基準ラベルの初期値を生成する処理を示している。
学習基準ラベル生成部10bは、学習用データDB2に登録されている学習用データの要素ごとのデータIDを順次読み出して以降の処理を実行する。
学習基準ラベル生成部10bは、学習用データDB2から読み出した要素のデータIDのうち、NG判定グループDB3に登録されているNG判定グループに属する要素のデータIDと同じデータIDがあるか否かを判定する(ステップST1a)。
学習基準ラベル生成部10bは、ステップST1aからステップST4aまでの一連の処理を、学習用データDB2に登録されている学習用データの全ての要素について繰り返し実行する。この処理により学習基準ラベルの初期値が生成される。
図13は、ニューラルネットワーク管理処理を示すフローチャートであり、要素ごとのNG指標値を算出する処理を示している。
初回処理において、ニューラルネットワーク管理部10cは、ニューラルネットワーク初期パラメータに基づいて、初期状態のニューラルネットワークを構築する(ステップST1b)。
なお、y2_i-y1_iをNGEiとしたが、y1_iおよびy2_iの比率をNGEiとしてもよく、y1_iおよびy2_iの分布を考慮した距離をNGEiとしてもよい。
ニューラルネットワーク管理部10cは、ステップST2bからステップST3bまでの一連の処理を、学習用データDB2に登録されている学習用データの全ての要素について繰り返し実行する。
図14は、NG指標値の履歴管理処理を示すフローチャートである。
まず、履歴管理部11cは、学習反復回数が一定数(例えば、100)で割り切れるか否かを判定する(ステップST1c)。学習反復回数は、図11に示したステップST2からステップST4までの学習部10による一連の処理が繰り返し行われる回数である。
次に、履歴管理部11cは、抽出したデータIDおよびNG指標値の系列に対して現在の学習反復回数を追加して履歴バッファ11aに格納する(ステップST3c)。
この後、履歴管理部11cは図14の処理を終了し、検定部11dの処理に移行する。
図15は、NG指標値の統計的検定処理を示すフローチャートである。
検定部11dは、履歴バッファ11aから、第j番目の要素(対象の要素)のデータIDに対応する、学習反復回数ごとのNG指標値の系列Vjを抽出し、抽出したNG指標値の系列Vjを、NG候補レベルの系列Ljに換算する(ステップST1d)。系列Ljは、系列Vjが過去にとった値の範囲を一定数で等分割した何番目の範囲に、現在のNG指標値が属しているかによって決定される。
NG候補レベル0は、NG指標値が過去にとった中間の範囲に対応する。NG候補レベル+1は、中間の範囲よりもNG指標値が大きい範囲に対応し、NG候補レベル+2は、NG候補レベル+1に対応する範囲よりもNG指標値が大きい範囲に対応する。NG候補レベル-1は、中間の範囲よりもNG指標値が小さい範囲に対応し、NG候補レベル-2は、NG候補レベル-1に対応する範囲よりもNG指標値が小さい範囲に対応する。
なお、初回はNG指標値が1つしかないため、系列Vjが過去にとった値の範囲は0である。このときのNG候補レベルを0とする。
この後、検定部11dは、履歴バッファ11aから、引き続き、第j+1番目の要素のデータIDに対応するNG指標値の系列Vj+1を抽出してステップST1dからの処理を繰り返す。
図16は、学習基準ラベル更新処理を示すフローチャートである。
学習基準ラベル更新部11eは、NG候補バッファ11bから、OK確定フラグに値1が設定されている要素のデータIDを、重複のない系列Uとして抽出する(ステップST1e)。系列Uには、OK確定フラグに値1が設定されている要素のデータIDが含まれている。
図17は、ニューラルネットワークパラメータ更新処理を示すフローチャートである。
ニューラルネットワーク更新部10dは、NG指標値バッファ10eから、ニューラルネットワーク管理部10cによって、現時点のニューラルネットワークパラメータで構築されたニューラルネットワークの出力結果から算出されたNG指標値を抽出する。
学習反復回数が既定値未満である場合(ステップST2f;NO)、ニューラルネットワーク更新部10dは、図17の処理を終了し、ニューラルネットワーク管理部10cによる図11のステップST2からの処理に移行する。
ここで、グループ(事例群)全体に対してNGと判定された履歴はあるが、グループ内のどの要素がNGであるかがわからない状況、例えば、3つのグループがあり、そのうちの2つのグループの全体に対してNGと判定され、残りの1つのグループの全体に対してOKと判定された状況を考える。NGと判定されたグループの1つを、NG判定グループ(1){E11(OK),E12(NG),E13(OK)}とし、もう1つを、NG判定グループ(2){E21(NG),E22(OK),E23(OK)}とする。OKと判定されたグループを、OK判定グループ(3){E31(OK),E32(OK),E33(OK)}とする。
また、学習結果の識別器は、NG判定グループに属するデータの少なくとも一つをNGと判定するので、グループ(5)がNG判定グループであれば、グループ(5)は適切にNGと判定され、グループ(4)はOKと判定される。
Claims (4)
- 複数の要素から構成された学習用データと、グループに属するデータの少なくとも一つがNGと判定されることが定義されたグループ情報とを入力し、前記学習用データの要素ごとにOKラベルまたはNGラベルが割り当てられた学習基準ラベルを用いて、前記学習用データの複数の要素を入力としてニューラルネットワークを学習することにより、学習結果のニューラルネットワークを用いて前記グループに属するデータの少なくとも一つをNGと判定する識別器を出力する学習部と、
前記学習部によるニューラルネットワークの学習の途中で前記学習基準ラベルを動的に更新する動的更新部とを備え、
前記学習部は、
前記学習用データおよび前記グループ情報を用いて前記学習用データの要素ごとの前記学習基準ラベルの初期値を生成し、
前記学習基準ラベルの初期値から前記動的更新部によって更新された前記学習基準ラベルを逐次用いてニューラルネットワークを反復して学習し、
学習の反復ごとのニューラルネットワークの出力値を用いて、前記学習用データの要素がNGとなり得るNG指標値を算出し、
前記動的更新部は、
学習の反復回数の推移ごとに得られたNG指標値の変化傾向を統計的に検定した結果に基づいて、前記学習基準ラベルを更新すること
を特徴とする学習装置。 - 前記学習部は、
前記学習用データの要素ごとの前記学習基準ラベルが格納された学習基準ラベルバッファと、
前記学習用データの要素が前記グループに属さない場合にOKラベルを割り当て、前記グループに属する場合にNGラベルを割り当てた前記学習基準ラベルの初期値を生成する学習基準ラベル生成部と、
ニューラルネットワークの出力値を用いて前記学習用データの要素のNG指標値を算出し、OKクラスおよびNGクラスのうち、NG指標値に基づいて前記学習用データの要素が属するクラスを判定し、前記学習基準ラベルの初期値および前記動的更新部によって更新された前記学習基準ラベルと判定したクラスとを誤差評価するニューラルネットワーク管理部と、
前記ニューラルネットワーク管理部によって算出されたNG指標値に基づいて、ニューラルネットワークパラメータを更新するニューラルネットワーク更新部とを備えたこと
を特徴とする請求項1記載の学習装置。 - 前記動的更新部は、
学習の反復回数ごとに得られたNG指標値が格納された履歴バッファと、
学習の反復回数、NG指標値に応じたNG候補レベル、前記学習用データの要素がOKに確定されたことを示すOK確定フラグ、および前記学習用データの要素がNGに確定されたことを示すNG確定フラグが、前記学習用データの要素ごとに格納されたNG候補バッファと、
学習の反復ごとに得られたNG指標値を前記履歴バッファに格納する履歴管理部と、
前記履歴バッファに格納されたNG指標値の変化傾向を、前記NG候補レベルの推移を用いて統計的に検定し、検定結果に基づいてNG候補の要素のOKまたはNGを確定して前記NG候補バッファにおける前記OK確定フラグまたは前記NG確定フラグを更新する検定部と、
NG候補の要素のうち、前記検定部によってOKに確定された要素に対応する前記学習基準ラベルをNGラベルからOKラベルに更新する学習基準ラベル更新部とを備えたこと
を特徴とする請求項1または請求項2記載の学習装置。 - 学習部が、複数の要素から構成された学習用データと、グループに属するデータの少なくとも一つがNGと判定されることが定義されたグループ情報とを入力し、前記学習用データの要素ごとにOKラベルまたはNGラベルが割り当てられた学習基準ラベルを用いて、前記学習用データの複数の要素を入力としてニューラルネットワークを学習することにより、学習結果のニューラルネットワークを用いて前記グループに属するデータの少なくとも一つをNGと判定する識別器を出力するステップと、
動的更新部が、前記学習部によるニューラルネットワークの学習の途中で前記学習基準ラベルを動的に更新するステップとを備え、
前記学習部は、
前記学習用データおよび前記グループ情報を用いて前記学習用データの要素ごとの前記学習基準ラベルの初期値を生成し、
前記学習基準ラベルの初期値から前記動的更新部によって更新された前記学習基準ラベルを逐次用いてニューラルネットワークを反復して学習し、
学習の反復ごとのニューラルネットワークの出力値を用いて、前記学習用データの要素がNGとなり得るNG指標値を算出し、
前記動的更新部は、
学習の反復回数の推移ごとに得られたNG指標値の変化傾向を統計的に検定した結果に基づいて、前記学習基準ラベルを更新すること
を特徴とする学習方法。
Priority Applications (7)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP18909932.8A EP3748549B1 (en) | 2018-03-16 | 2018-03-16 | Learning device and learning method |
US16/971,482 US20200387792A1 (en) | 2018-03-16 | 2018-03-16 | Learning device and learning method |
JP2020506078A JP6701467B2 (ja) | 2018-03-16 | 2018-03-16 | 学習装置および学習方法 |
KR1020207025875A KR20200108912A (ko) | 2018-03-16 | 2018-03-16 | 학습 장치 및 학습 방법 |
CN201880091109.7A CN111837143A (zh) | 2018-03-16 | 2018-03-16 | 学习装置及学习方法 |
PCT/JP2018/010446 WO2019176087A1 (ja) | 2018-03-16 | 2018-03-16 | 学習装置および学習方法 |
TW107121878A TW201939363A (zh) | 2018-03-16 | 2018-06-26 | 學習裝置以及學習方法 |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/JP2018/010446 WO2019176087A1 (ja) | 2018-03-16 | 2018-03-16 | 学習装置および学習方法 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2019176087A1 true WO2019176087A1 (ja) | 2019-09-19 |
Family
ID=67908184
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2018/010446 WO2019176087A1 (ja) | 2018-03-16 | 2018-03-16 | 学習装置および学習方法 |
Country Status (7)
Country | Link |
---|---|
US (1) | US20200387792A1 (ja) |
EP (1) | EP3748549B1 (ja) |
JP (1) | JP6701467B2 (ja) |
KR (1) | KR20200108912A (ja) |
CN (1) | CN111837143A (ja) |
TW (1) | TW201939363A (ja) |
WO (1) | WO2019176087A1 (ja) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200387826A1 (en) * | 2019-06-04 | 2020-12-10 | Fuji Xerox Co., Ltd. | Information processing apparatus and non-transitory computer readable medium |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH07168799A (ja) * | 1993-09-22 | 1995-07-04 | Fuji Electric Co Ltd | ニューラルネットワークの学習装置 |
JP2006226708A (ja) * | 2005-02-15 | 2006-08-31 | Nagoya Electric Works Co Ltd | 良否判定装置、良否判定方法および良否判定プログラム |
JP2012208913A (ja) | 2011-03-28 | 2012-10-25 | Nec (China) Co Ltd | ハイブリッドラベルに基づいてオブジェクトを分割する方法とシステム |
JP2017102906A (ja) * | 2015-11-25 | 2017-06-08 | キヤノン株式会社 | 情報処理装置、情報処理方法及びプログラム |
Family Cites Families (14)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
EP1515270A1 (en) * | 2003-09-09 | 2005-03-16 | Semeion | An artificial neural network |
JP4352980B2 (ja) * | 2004-04-23 | 2009-10-28 | オムロン株式会社 | 拡大表示装置及び拡大画像制御装置 |
US20060045312A1 (en) * | 2004-08-27 | 2006-03-02 | Bernstein Daniel B | Image comparison device for providing real-time feedback |
JP4661413B2 (ja) * | 2005-07-11 | 2011-03-30 | 富士フイルム株式会社 | 撮像装置、撮影枚数管理方法及び撮影枚数管理プログラム |
JP2015087903A (ja) * | 2013-10-30 | 2015-05-07 | ソニー株式会社 | 情報処理装置及び情報処理方法 |
JP6291844B2 (ja) * | 2014-01-06 | 2018-03-14 | 日本電気株式会社 | データ処理装置 |
JP6450858B2 (ja) * | 2015-11-25 | 2019-01-09 | 株式会社日立製作所 | 設備管理装置および方法 |
WO2017130699A1 (ja) * | 2016-01-26 | 2017-08-03 | 富士フイルム株式会社 | ひび割れ情報検出装置、ひび割れ情報検出方法およびひび割れ情報検出プログラム |
JP6182242B1 (ja) * | 2016-06-13 | 2017-08-16 | 三菱電機インフォメーションシステムズ株式会社 | データのラベリングモデルに係る機械学習方法、コンピュータおよびプログラム |
US10339471B2 (en) * | 2017-01-17 | 2019-07-02 | International Business Machines Corporation | Ensemble based labeling |
JP2018142097A (ja) * | 2017-02-27 | 2018-09-13 | キヤノン株式会社 | 情報処理装置、情報処理方法及びプログラム |
US11436428B2 (en) * | 2017-06-06 | 2022-09-06 | Sightline Innovation Inc. | System and method for increasing data quality in a machine learning process |
JP6568175B2 (ja) * | 2017-10-20 | 2019-08-28 | ヤフー株式会社 | 学習装置、生成装置、分類装置、学習方法、学習プログラム、および動作プログラム |
US11382601B2 (en) * | 2018-03-01 | 2022-07-12 | Fujifilm Sonosite, Inc. | Method and apparatus for annotating ultrasound examinations |
-
2018
- 2018-03-16 US US16/971,482 patent/US20200387792A1/en active Pending
- 2018-03-16 JP JP2020506078A patent/JP6701467B2/ja active Active
- 2018-03-16 WO PCT/JP2018/010446 patent/WO2019176087A1/ja unknown
- 2018-03-16 KR KR1020207025875A patent/KR20200108912A/ko active IP Right Grant
- 2018-03-16 CN CN201880091109.7A patent/CN111837143A/zh active Pending
- 2018-03-16 EP EP18909932.8A patent/EP3748549B1/en active Active
- 2018-06-26 TW TW107121878A patent/TW201939363A/zh unknown
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH07168799A (ja) * | 1993-09-22 | 1995-07-04 | Fuji Electric Co Ltd | ニューラルネットワークの学習装置 |
JP2006226708A (ja) * | 2005-02-15 | 2006-08-31 | Nagoya Electric Works Co Ltd | 良否判定装置、良否判定方法および良否判定プログラム |
JP2012208913A (ja) | 2011-03-28 | 2012-10-25 | Nec (China) Co Ltd | ハイブリッドラベルに基づいてオブジェクトを分割する方法とシステム |
JP2017102906A (ja) * | 2015-11-25 | 2017-06-08 | キヤノン株式会社 | 情報処理装置、情報処理方法及びプログラム |
Non-Patent Citations (1)
Title |
---|
C. M. BISHOPHIROSHI MOTODA: "Pattern Ninshiki To Kikai Gakusyu, Jyou", ASAKURA PUBLISHING CO. LTD., pages: 398 - 399,402-403 |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20200387826A1 (en) * | 2019-06-04 | 2020-12-10 | Fuji Xerox Co., Ltd. | Information processing apparatus and non-transitory computer readable medium |
Also Published As
Publication number | Publication date |
---|---|
KR20200108912A (ko) | 2020-09-21 |
CN111837143A (zh) | 2020-10-27 |
EP3748549A1 (en) | 2020-12-09 |
JPWO2019176087A1 (ja) | 2020-05-28 |
US20200387792A1 (en) | 2020-12-10 |
JP6701467B2 (ja) | 2020-05-27 |
EP3748549A4 (en) | 2021-05-05 |
TW201939363A (zh) | 2019-10-01 |
EP3748549B1 (en) | 2023-06-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
US10867244B2 (en) | Method and apparatus for machine learning | |
JP7117934B2 (ja) | 画像分類装置及びプログラム | |
CN113610747B (zh) | 检验半导体样本的***和方法 | |
CN113487223B (zh) | 一种基于信息融合的风险评估方法和评估*** | |
WO2019176087A1 (ja) | 学習装置および学習方法 | |
CN113269255A (zh) | 用于检测缺陷的方法、设备和计算机可读存储介质 | |
JP2019158684A (ja) | 検査システム、識別システム、及び識別器評価装置 | |
CN111723010A (zh) | 一种基于稀疏代价矩阵的软件bug分类方法 | |
CN110717037A (zh) | 对用户分类的方法和装置 | |
US11995150B2 (en) | Information processing method and information processing system | |
CN115398442A (zh) | 用于评估传感器测量值的设备和自动化方法和该设备的应用 | |
CN116610806B (zh) | 基于ai的rpa数字化业务处理方法及计算机设备 | |
CN113822532B (zh) | 信息***资产风险评估方法及其装置和存储介质 | |
US20230267321A1 (en) | Viability determination with self-attention for process optimization | |
CN113449746B (zh) | 尾气遥测数据修正方法和装置 | |
CN112529038B (zh) | 一种主板物料的识别方法、装置及存储介质 | |
US20230315769A1 (en) | A method for address matching and related electronic device | |
CN117952717A (zh) | 一种基于大数据的机票订单处理方法及*** | |
CN116340740A (zh) | 一种数据质检方法、装置、设备及存储介质 | |
CN117541912A (zh) | 目标检测算法的失效原因分析方法、装置及设备 | |
CN116702014A (zh) | 人口识别方法、装置、终端设备及存储介质 | |
JP2022070744A (ja) | 学習処理装置、方法及びプログラム | |
CN115759265A (zh) | 一种基于模糊偏好关系的故障诊断方法及*** | |
JP2024009787A (ja) | 人工ニューラルネットワークの最適なアーキテクチャを特定するための方法 | |
CN118071006A (zh) | 一种装备体系能力需求满足度评估方法 |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18909932 Country of ref document: EP Kind code of ref document: A1 |
|
ENP | Entry into the national phase |
Ref document number: 2020506078 Country of ref document: JP Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 20207025875 Country of ref document: KR Kind code of ref document: A |
|
ENP | Entry into the national phase |
Ref document number: 2018909932 Country of ref document: EP Effective date: 20200904 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |