Patent Application
20210209468
The present disclosure relates to a training device, an inference device, a method, and a program.
Carrying out deep learning, which is one method of machine learning, needs setting of training parameters in accordance with a purpose, the characteristics of the training data, and the like. However, appropriately setting the training parameters, including selecting a learning model, determining the scale of the neural network, and the like, is not easy for a user that is not knowledgeable about neural networks, artificial intelligence (AI), and the like. Therefore, it is difficult for such a user to perform deep learning.
In an authentication device described in Patent Literature 1 that performs individual authentication on the basis of writing information, the individual authentication is performed using a neural network assigned to a category of the writing information that is the recognition subject.
Patent Literature 1: Unexamined Japanese Patent Application Publication No. 2002-175515
The authentication device described in Patent Literature 1 simply uses, among a plurality of neural networks, a neural network that is assigned to the category of the recognition subject. Furthermore, the number of layers of the plurality of neural networks is the same as the number of nodes or the like of each layer. In other words, each neural network has the same scale. Consequently, when, for example, changing the scale of the neural network, the user needs to determine the scale on their own. As such, it is difficult for a user that is not knowledgeable about neural networks, AI, and the like to appropriately operate the authentication device described in Patent Literature 1.
The present disclosure is made with the view of the above situation, and an objective of the present disclosure is to enable the setting of appropriate training parameters without the user being aware of the settings of the training parameters.
To achieve the above objective, a training device of the present disclosure performs training using a neural network. Training condition acquisition means acquire training conditions that indicate prerequisites of the training. Learning model selection means select, in accordance with the training conditions, a learning model that serves as a framework of a structure of the neural network. Learning model scale determination means determine, in accordance with the training conditions, a scale of the neural network for the selected learning model. Training means perform training by inputting training data into the neural network in which the learning model is configured to the scale.
In accordance with training conditions, the training device of the present disclosure selects a learning model that serves as the framework of the structure of a neural network and determines the scale of the neural network for the selected learning model. Providing the training device of the present disclosure with such a configuration enables the setting of appropriate training parameters without the user being aware of the setting of the training parameters.
Hereinafter, a training inference device 1000 according to embodiments of the present disclosure is described with reference to the drawings.
A training inference device 1000 according to the present embodiment automatically determines appropriate training parameters on the basis of information that indicates prerequisites and restrictions related to training that are specified by a user. In this case, the training parameters include a learning model representing the structure of the neural network, the scale of the neural network, a training rate, an activation function, a bias value, and the like.
Specifically, in the present embodiment, the training inference device 1000 automatically determines, from among the training parameters and on the basis of the information that indicates prerequisites and restrictions related to training that are specified by the user, the learning model representing the structure of the neural network and the scale of the neural network.
The training inference device 1000 selects the learning model and expands or shrinks the scale of the neural network for the selected learning model to obtain a deep learning neural network changed to an optimal configuration, and uses this neural network to execute deep learning. The training inference device 1000 performs inference based on training results of the deep learning and data to be inferred.
Here, the term “deep learning” refers to a learning method that uses a multi-layer neural network. The term “multi-layer neural network” refers to a neural network that includes a plurality of intermediate layers positioned between an input layer and an output layer. Hereinafter, the multi-layer neural network is sometimes referred to as a deep neural network. In deep learning, a learning model is presumed, training data is input into a neural network that realizes the presumed learning model, and weighting of the nodes of the intermediate layers of the neural network are modified such that the output of the neural network approaches a true value obtained in advance. Thus, the deep neural network is trained with the relationship between the input and the output.
The deep neural network that has been trained is used in inference. The term “inference” refers to estimating using a trained deep neural network. In the inference, data to be inferred is input into the trained network, and a value output by the trained deep neural network is set as an inference value with respect to the input.
The training inference device 1000 performs training and inference in a production system, a control system, or the like for quality inspection, abnormality cause estimation, device failure prediction, and the like. In one example, the training data provided to the training inference device 1000 is data collected over a given period in the past from various devices such as programmable logic controllers and intelligent functional units that operate in production systems, control systems, and the like, and sensors provided in facilities.
Furthermore, the training inference device 1000 performs inference by the trained deep neural network for quality inspection, abnormality cause estimation, device failure prediction, and the like. In one example, the data to be inferred that is provided to the training inference device 1000 is data collected from various devices such as programmable logic controllers, intelligent functional units, and sensors provided in facilities.
As illustrated in
The storage 1 includes volatile memory and non-volatile memory, and stores programs and various types of data. The storage 1 is used as the working memory of the operator 4. The programs stored in the storage 1 include a training processing program 11 for realizing the various functions of a training device 100 (described later), and an inference processing program 12 for realizing the various functions of an inference device 200 (described later).
The inputter 2 includes a keyboard, a mouse, a touch panel, or the like. The inputter 2 detects input operations performed by the user and outputs, to the operator 4, signals representing the detected input operations performed by the user.
The display 3 includes a display, a touch panel, or the like. The display 3 displays images based on signals supplied from the operator 4.
The operator 4 includes a central processing unit (CPU). The operator 4 executes the various programs stored in the storage 1 to realize the various functions of the training inference device 1000. The operator 4 may include a processor dedicated to AI use.
As illustrated in
In the present embodiment, the training device 100 selects, on the basis of information that indicates prerequisites and restrictions related to training that are input by the user, a learning model that serves as the framework of the pre-modified deep neural network, changes the selected learning model to a configuration that satisfies the prerequisites and restrictions of the training that are input by the user, and generates a deep neural network. Prior to the inference by the inference device 200, the training device 100 modifies the deep neural network by training using the training data.
As illustrated in
The training condition acquirer 110 acquires, from the input of the user received by the inputter 2, content of the training conditions representing the prerequisites and restrictions related to the training, and outputs the acquired content of the training conditions to the model selector 150. The prerequisites and restrictions input by the user include an inference purpose, restrictions on hardware resources, information representing the characteristics of the training data, and a target to be achieved in the training.
Next, the information that the training condition acquirer 110 receives from the user is described in detail.
The training condition acquirer 110 receives an input about the inference purpose from the user, and outputs, to the model selector 150, information indicating the purpose selected by the user. The inference purpose indicates the purpose of the inference to be performed by the inference device 200 (described later). The inference device 200 uses the deep neural network that is modified by the training device 100 and, as such, the training device 100 performs training that corresponds to the inference purpose specified by the user.
The training condition acquirer 110 displays an input screen such as illustrated in
The training condition acquirer 110 receives inputs about restrictions on hardware resources from the user. The restrictions on hardware resources indicate restrictions on hardware resource usable by the training inference device 1000 for the training of the training device 100.
The training condition acquirer 110 displays an input screen such as illustrated in
The training condition acquirer 110 illustrated in
In the case of the present embodiment, the training data includes simple numerical data and data that is labeled. The data that is labeled (hereinafter referred to as “labeled data”) is data for which meanings represented by the possible values are defined.
The labeled data includes data defined for each value. For example, in order to indicate whether a switch is ON or OFF, “1” is associated with ON and “0” is associated with OFF. This definition is stored in advance in the storage 1. When defined in this manner, the value of the labeled data related to a switch in the training data is 1 or 0. In another example, in order to indicate ranges of air temperatures, “1” is associated with 1° C. to 20° C., “2” is associated with 20.1° C. to 30° C., and “3” is associated with 30.1° C. to 40° C. When defined in this manner, the value of the labeled data related to air temperature in the training data is 1, 2, or 3. The preprocessor 130, the model selector 150, and the trainer 170 handle each of the labeled data related to the switch and the labeled data related to the air temperature on the basis of information about the definitions stored in the storage 1.
The label may represent a characteristic of that value. For example, a “rotation speed” label may be attached to data obtained by measuring rotation speed. In this case, the value in the training data is any value obtained by measuring rotation speed. The preprocessor 130, the model selector 150, and the trainer 170 handle the data labeled with “rotation speed” as data obtained by measuring rotation speed.
As described above, the training data includes simple numerical data and labeled data. As such, the type of training data acquired by the training condition acquirer 110 includes information indicating if the training data is simple numerical data or is labeled data. Furthermore, when the training data is labeled data, the training condition acquirer 110 acquires a label name. For example, the label name is “switch”, “air temperature”, or “rotation speed.”
The training condition acquirer 110 displays an input screen such as illustrated in
In
Additionally, the range of possible values of the training data acquired by the training condition acquirer 110 is indicated by the maximum value and the minimum value of the training data. The maximum value of each column is the maximum value of the set of data of that dimension, and the minimum value of each column is the minimum value of the set of data of that matter. In one example, the maximum value and the minimum value are used when preprocessing. In the example illustrated in
Information indicating whether the training data acquired by the training condition acquirer 110 is time series data is also input via the screen illustrated in
The training condition acquirer 110 illustrated in
The training data storage section 120 illustrated in
Correct answer data used in training for the purpose of quality inspection is, for example, data collected at the time of manufacture of a part, and includes information indicating if the quality of that part passed or failed.
Correct answer data used in training for the purpose of abnormality cause estimation is, for example, data collected from a device that is operated at the time of occurrence of an abnormality, from a sensor provided on that device, or the like; and includes information indicating the cause of the occurrence of the abnormality.
Correct answer data used in training for the purpose of failure sign sensing is, for example, data collected from a device that operates, from a sensor provided on that device, or the like; and includes information indicating if an operating state of that device is normal or abnormal.
Alternatively, the correct answer data used in training for the purpose of failure sign sensing may, for example, consist only of data collected at the time of failure occurrence from the device that operates, the sensor provided on that device, or the like. In this case, the correct answer data includes information indicating a level, among a number of predefined levels representing degrees of failure, of the operating state of that device.
Prior to the training, the preprocessor 130 carries out preprocessing on the training data, and outputs the preprocessed data to the trainer 170. In one example, the preprocessing includes fast Fourier transformation, difference processing, logarithmic conversion, and differential processing. The preprocessor 130 carries out preprocessing corresponding to each individual piece of training data. For example, when the training data is a measured value of rotation speed, and is labeled data labeled with “rotation speed”, that data is subjected to frequency analysis by fast Fourier transformation. The preprocessor 130 stores information identifying the content of the preprocessing and the preprocessed data in the training results storage section 180. This is done to use the same preprocessing method in the inference device 200 (described later).
The learning model storage section 140 stores information related to a plurality of learning models. Specifically, the learning model storage section 140 includes a model definition region 1401 that stores equations expressing each of the learning models selectable by the model selector 150. The learning model storage section 140 further includes an initial parameter region 1402 that stores initial parameters of each of the learning models. The initial parameter region 1402 stores, for each of the pre-modified learning models, an initial value for the number of intermediate layers, an initial value for the number of nodes in each intermediate layer, an initial value for the number of nodes of the output layer, an initial value of a weighting applied to the input value of each node, and a training rate that indicates the updatable range of the weighting of each node. These initial values and training rates that are stored in the learning model storage section 140 may be defined for each of the plurality of learning models to be selected by the model selector 150 (described later). Note that, fundamentally, the number of nodes of the input layer of the deep neural network is set so as to be equivalent to the number of dimensions of the training data.
Furthermore, the learning model storage section 140 includes a selection table 1403 that the model selector 150 uses when selecting the learning model. As illustrated in
The model selector 150 illustrated in
In the present embodiment, the model selector 150 selects the learning model on the basis of the inference purpose, the characteristics of the training data, and the selection table 1403 illustrated in
Furthermore, the model selector 150 changes the configuration of the learning model in accordance with the type of training data input by the user. For example, as illustrated in
The model scale determiner 160 determines the scale of the learning model in accordance with the training conditions acquired by the training condition acquirer 110. In the present embodiment, for the learning model selected by the model selector 150, the model scale determiner 160 increases or decreases the number of intermediate layers, increases or decreases the number of nodes in each intermediate layer, and determines whether to provide a connection between nodes on the basis of the restrictions on hardware resources specified by the user. For example, when the scale of the intermediate layers increases, the model scale determiner 160 sets the connections between a portion of the nodes to null. The speed of computation can be increased by setting the connections between a portion of the nodes to null in this manner.
In one example, when the target correct answer rate input by the user in the screen illustrated in
The model scale determiner 160 outputs the learning model, which is changed to the determined scale, to the trainer 170. Additionally, the model scale determiner 160 stores, in the training results storage section 180, the changed number of intermediate layers and the changed number of nodes in each intermediate layer as information indicating the determined scale of the learning model.
The trainer 170 inputs the preprocessed training data supplied from the preprocessor 130 into the deep neural network that uses the learning model output by the model scale determiner 160 to carry out the training. The trainer 170 inputs the training data into the deep neural network and appropriately updates the weighting of each of the nodes by back propagation so that the output value approaches the correct answer data stored in the training data storage section 120.
Additionally, the trainer 170 consecutively calculates the correct answer rate from the difference between the output of the deep neural network and the correct answer information in order to determine the training end condition. The trainer 170 ends the training when the calculated correct answer rate reaches the correct answer rate specified by the user. The trainer 170 stores the modified weighting of each node of the deep neural network in the training results storage section 180 as training results. Additionally, the trainer 170 carries out training processing while monitoring the load on the operator 4 so as not to exceed the processor utilization specified by the user in the screen illustrated in
The trainer 170 displays a screen indicating progress such as illustrated in
The trainer 170 starts, interrupts, and restarts the training in accordance with commands from the user.
The training results storage section 180 stores the final weighting of each node of the deep neural network as the training results of the trainer 170. The configuration of the training device 100 is described above.
Next the inference device 200 illustrated in
The inference data storage section 210 stores the inference subject data.
Prior to the inference, the inferer 220 reads, from the training results storage section 180, the preprocessing method that the preprocessor 130 performs on the training data, and preprocesses the inference subject data.
After the preprocessing, the inferer 220 inputs, on the basis of the information stored in the training results storage section 180, the inference subject data into the modified deep neural network, and outputs an output value to the inference results storage section 230. During the execution of the inference as well, the inferer 220 displays a screen indicating progress on the display 3, similar to the progress when training illustrated in
The inference results storage section 230 stores the inference results of the inferer 220. Specifically, the inference results storage section 230 stores inference results based on the output of the deep neural network. The configuration of the inference device 200 is described above.
Next, the flow of training processing of the training device 100 is described while referencing
The preprocessor 130 selects the preprocessing method according to the training conditions supplied from the training condition acquirer 110 and the training data stored in the training data storage section 120 (step S12). The preprocessor 130 uses the selected preprocessing method to preprocess the training data stored in the training data storage section 120 (step S13), and supplies the preprocessed training data to the trainer 170. Additionally, the preprocessor 130 stores the preprocessing method that is used in the training results storage section 180.
The model selector 150 selects the learning model from the learning model storage section 140 in accordance with the training conditions supplied from the training condition acquirer 110 and the training data stored in the training data storage section 120 (step S14). Furthermore, the model selector 150 changes the configuration of the selected learning model in accordance with the type of the training data, and supplies information identifying that learning model to the model scale determiner 160.
The model scale determiner 160 determines, in accordance with the training condition supplied from the training condition acquirer 110, the scale of the learning model selected by the model selector 150 (step S15), and supplies the determined content to the trainer 170.
Until the target correct answer rate specified by the user is reached (step S16; No), the trainer 170 performs the training processing (step S17). Specifically, the trainer 170 inputs the training data into the deep neural network that uses the configuration determined by the model selector 150 and the model scale determiner 160, and calculates the correct answer rate from the correct answer data and the output of the deep neural network. The trainer 170 updates the display of the screen with the current training progress and the newest correct answer rate (step S18).
When the target correct answer rate specified by the user is reached (step S16; Yes), the trainer 170 ends the training and outputs the training results that include the weighting of each node (step S19). The flow of the training processing of the training device 100 is described above.
Next, inference processing of the inference device 200 using the trained deep neural network is described while referencing
The inferer 220 reads, from the training results storage section 180, the preprocessing method that the preprocessor 130 performed on the training data, and performs preprocessing on the inference subject data stored in the inference data storage section 210 (step S21).
After the preprocessing, the inferer 220 reads, from the training results storage section 180, the information identifying the learning model selected by the model selector 150, the information identifying the scale determined by the model scale determiner 160, and the weightings of the deep neural network updated by the trainer 170. The inferer 220 inputs the inference subject data into the deep neural network that uses the read content, and executes the inference (step S22). The inferer 220 stores the inference results in the inference results storage section 230. The inference processing is described above.
As described above, in the present embodiment, the training device 100 automatically optimizes the learning model by selecting an appropriate learning model and determining the scale of the selected learning model in accordance with prerequisites and restrictions related to training that are specified by the user. As a result, the need for the user to select the learning model and determine of the scale of the learning model, which are tasks conventionally performed by the user, is eliminated. Therefore, deep learning can be easily performed even when the user is not especially knowledgeable.
The model scale determiner 160 modifies the scale of the learning model in accordance with the restrictions on hardware resources specified by the user. As such, when, for example, another application is running on the training device 100, the training can be executed without interfering with the running of the other application.
Since the model scale determiner 160 appropriately modifies the scale of the learning model, training in which a large-scale neural network is used for uncomplicated training data is avoided. Additionally, the training device 100 does not perform training using a small-scale neural network for complex training data. Due to this configuration. disadvantages such as unnecessary time use and unnecessary increases in processing loads on the processor, resulting from performing training using large-scale neural networks on uncomplicated training data without modifying the scale, are avoided. Additionally, disadvantages such as not obtaining satisfactory training results, resulting from performing training using small-scale neural networks on complex training data without modifying the scale, are avoided.
Furthermore, the model selector 150 may change the configuration of the learning model in accordance with the type of training data input by the user. For example, the model selector 150 may change the learning model so that labeled data is input directly into the intermediate layers without being input into the input layer. This is because the input training data is standardized in the input layer and, in this case, since the definitions of the various values are defined in advance in the labeled data, it is possible to omit standardization processing.
In the present embodiment, an example is described in which the model scale determiner 160 expands or shrinks the scale of the learning model in accordance with the memory capacity specified by the user as the restriction on hardware resources. However, the method of expanding or shrinking the scale of the learning model is not limited thereto.
For example, the model scale determiner 160 may expand or shrink the scale of the learning model in accordance with the number of dimensions of the input training data. Additionally, the model scale determiner 160 may expand or shrink the scale of the learning model in accordance with the degree of complexity of the training data. For example, when the training data is complex data, the scale of the learning model may be expanded and, when the training data is not complex, the scale of the learning model may be shrunk. The degree of complexity of the training data can be calculated, for example, by acquiring the average, variation, or other statistical quantity of the training data.
Additionally, the model scale determiner 160 can expand or shrink the scale of the learning model in accordance with the characteristics of the training data. For example, the scale of the learning model can be expanded or shrunk in accordance with whether the training data is temporally continuous data, or whether the training data has relevancy in a time series. For example, when the training data is temporally continuous data or has relevancy in a time series, one cycle of the data must be collectively input into the neural network. In this case, the number of input dimensions of the neural network increases. As a result, the scale of the neural network expands.
Additionally, the model scale determiner 160 can expand or shrink the scale of the learning model in accordance with the data type of the training data. This is because the structure of the neural network differs depending on the data type of the training data, which results in the scale of the neural network expanding or shrinking. Here, the types of data include numerical values, labeled data, and the like.
In the embodiment, the selection of the learning model and the determination of the scale are performed in accordance with the information indicating the inference purpose, the restrictions on hardware resources, the characteristics of the training data, and the target to be achieved that are input as the training conditions. However, it is possible to use only a portion of these as the training conditions. For example, the user may input only the inference purpose as the training condition, and the training device 100 may select the model and determine the scale in accordance with the input inference purpose.
The method of selecting the model is not limited to the method described in the embodiment. In one example, the learning model storage section 140 stores an evaluation value obtained by evaluating, in advance, the performance of each learning model. In a case in which, based on the inference purpose and the characteristics of the training data input by the user, there are a plurality of matching learning models in the selection table 1403, the model selector 150 selects the learning model on the basis of a target value to be achieved input by the user and the evaluation values indicating the performance of each of the matching learning models. When the target value to be achieved, namely the target correct answer rate, is greater than or equal to a predetermined value, the model selector 150 may select the learning model for which the evaluation value that represents performance is high.
A configuration is possible in which the training device 100 does not use the training conditions related to the selection and the scale of the model that are input by the user via the training condition input screens. For example, a file indicating conditions specified by the user may be stored in advance in the storage 1, and this file may be read out to perform the selection of the model and the determination of the scale in accordance with the training conditions.
In the embodiment, an example is described in which the training inference device 1000 includes the training device 100 and the inference device 200. However, a configuration is possible in which the training device 100 and the inference device 200 are separate devices.
In the embodiment, an example is described in which the training data is stored in advance in the training data storage section 120. However, the location where the training data is stored is not limited thereto. For example, a configuration is possible in which the training device 100 is provided with a network interface that enables communication with other devices, and the training data is provided from another device that is connected to the training device 100 via a network.
Likewise, a configuration is possible in which the inference subject data is provided to the inference device 200 from another device via a network. Additionally, a configuration is possible in which the inference device 200 processes inference subject data supplied in real-time and outputs inference results in real-time.
A computer-readable non-transitory recording medium such as a magnetic disk, an optical disk, a magneto-optical disk, a flash memory, a semiconductor memory. and a magnetic tape can be used as a recording medium on which the programs for the training processing and the inference processing in accordance with the embodiment described above.
The foregoing describes some example embodiments for explanatory purposes. Although the foregoing discussion has presented specific embodiments, persons skilled in the art will recognize that changes may be made in form and detail without departing from the broader spirit and scope of the invention. Accordingly, the specification and drawings are to be regarded in an illustrative rather than a restrictive sense. This detailed description, therefore, is not to be taken in a limiting sense, and the scope of the invention is defined only by the included claims, along with the full range of equivalents to which such claims are entitled.
| Filing Document | Filing Date | Country | Kind |
|---|---|---|---|
| PCT/JP2018/021488 | 6/5/2018 | WO | 00 |
