Individualized learning models

Abstract

To efficiently individualize an existing learning model.SOLUTION: In an information processing method executed by an information processing device having a processor, the processor performs: obtaining first learning data associated with first correct answer data, out of a predetermined data set including each piece of learning data associated with each piece of correct answer data; obtaining second correct answer data given to the first learning data; calculating prediction data by inputting the first learning data to a learning model corresponding to the predetermined data set; adjusting a weight given to an error between the second correct answer data and the prediction data that are used for a loss function set for the learning model, according to a result of comparison between the first correct answer data and the second correct answer data; and learning the learning model by error backpropagation using an output value of the loss function.SELECTED DRAWING: Figure 7

Description

[Technical Field]

[0001] The present invention relates to an information processing method, an information processing device, and a program for individualizing a learning model. Regarding. [Background technology]

[0002] A general learning model is generated by inputting learning data based on information from multiple users. There are known techniques for personalizing general learning models based on information from a single user. (See, for example, Patent Document 1). [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Patent No. 6432859 Summary of the Invention [Problem to be solved by the invention]

[0004] For example, for a general learning model trained by supervised learning, When individualizing data by adding correct answer data to the data, the correct answer data for each user is added. Learning is performed using training data.

[0005] However, there is a discrepancy between the correct answer data provided by the user and the correct answer data provided in advance. If there is a discrepancy, this learning data should be the focus from the perspective of individualization. The current situation is that no particular action has been taken.

[0006] Therefore, the present invention focuses on the discrepancy between existing correct answer data and newly assigned correct answer data. , an information processing method and an information processing device that enable efficient individualization of existing learning models We provide facilities and programs. [Means for solving the problem]

[0007] An information processing method according to one aspect of the present invention is executed by an information processing device having a processor. The information processing method includes: First training data associated with the first correct answer data from a predetermined dataset containing the first correct answer data. and acquiring second correct answer data assigned to the first training data. The first training data is input to a training model corresponding to the predetermined data set to generate a prediction data. and calculating data according to a comparison result between the first correct answer data and the second correct answer data. The second correct answer data and the predicted data used in the loss function set in the learning model and adjusting the weight given to the error between the loss function and the error backpropagation method using the output value of the loss function. The learning model is learned by the above. [Effects of the Invention]

[0008] According to the present invention, it is possible to efficiently individualize an existing learning model. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram illustrating an example of a system configuration according to an embodiment. [Figure 2] FIG. 1 is a diagram illustrating an example of a physical configuration of an information processing apparatus according to an embodiment. [Figure 3] FIG. 2 is a diagram illustrating an example of a processing block of the information processing apparatus according to the embodiment. [Figure 4] FIG. 2 is a diagram illustrating an example of a data set according to the embodiment. [Figure 5] FIG. 10 is a diagram illustrating an example of an annotation according to an embodiment. [Figure 6] FIG. 10 is a diagram illustrating an example of selecting a group in a dataset according to an embodiment. [Figure 7] FIG. 10 is a sequence diagram illustrating an example of processing by the information processing device according to the embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] An embodiment of the present invention will be described with reference to the accompanying drawings. Those marked with the same or similar symbols have the same or similar configurations.

[0011] <System configuration> FIG. 1 is a diagram illustrating an example of a system configuration according to an embodiment. In the example illustrated in FIG. The server 10 and each of the information processing devices 20A, 20B, 20C, and 20D are connected via a network. When the information processing devices are not individually distinguished, the information processing devices are connected so that data can be transmitted and received. It is also referred to as device 20.

[0012] The server 10 is an information processing device capable of collecting and analyzing data, and includes one or more information processing devices. The information processing device 20 may be a smartphone, a personal computer, or the like. Information on which machine learning can be performed, such as computers, tablets, servers, and connected cars The information processing device 20 is an invasive or non-invasive device that senses brain waves. It is also a device that is directly or indirectly connected to the electrodes and can analyze, send and receive EEG data. good.

[0013] In the system shown in FIG. 1, a server 10 stores one or more databases used for supervised learning. and a dataset trained on at least one of the datasets. The server 10 receives requests from the information processing devices 20 and has a trained learning model. Depending on the test, a predetermined dataset and a learning model trained using this dataset A set of the predetermined data set or a predetermined learning model may be transmitted to each information processing device 20. In addition, the server 10 may provide the learned information in response to a request from each information processing device 20. Send a program to personalize the learning model or record this program for non-transient purposes. A recording medium may be provided.

[0014] Each information processing device 20 executes a program stored in a memory or acquired from the server 10. Each information processing device 20 executes the learning model to individualize the learning model. When personalizing a model, the model is based on one or more training data contained in a given dataset. For example, each information processing device 20 may The data is output to a screen or the like, and correct answer data (also called correct answer value or correct answer label) is accepted.

[0015] In this case, the individual learning models are created based on the learning data to which the correct answer data is added by the user. At this time, the relationship between the existing correct answer data and the correct answer data assigned by the user is The training data with deviations has more characteristics than the training data without deviations. The weight assigned to the error in the loss function is adjusted so that it is more strongly reflected in the learning.

[0016] This allows the user's characteristics to be reflected when personalization is performed on existing learning models. By assigning weights to the prediction errors of the training data, it becomes possible to train efficiently. For example, there is a discrepancy between the existing correct answer data and the correct answer data assigned by the user. By increasing the prediction error of the training data and performing the backpropagation algorithm, the characteristics of this user can be The prediction error of the training data with the user's characteristics is reflected in the output value of the loss function. As learning is repeated using backpropagation based on output values ​​that reflect the characteristics, The hyperparameters are set so that the prediction error of the training data with this user's characteristics becomes smaller. The data is adjusted, and the learning process can more strongly reflect the user's characteristics. The configuration of each device in this embodiment will be described below.

[0017] <Hardware configuration> 2 is a diagram showing an example of the physical configuration of the information processing device 10 according to the embodiment. The processing device 10 includes a CPU (Central Processing Unit) 10a, which corresponds to a calculation unit, and a storage unit RAM (Random Access Memory) 10b corresponding to the memory unit, and ROM (Read Only Memory) 10c corresponding to the memory unit. The device has a memory (only memory) 10c, a communication unit 10d, an input unit 10e, and a display unit 10f. These components are connected via a bus so that they can transmit and receive data to and from each other.

[0018] In this embodiment, the information processing device 10 is configured by one computer. As will be explained, the information processing device 10 is a combination of multiple computers or multiple calculation units. 1 is an example, and the information processing device 10 may be realized in other ways. The above configuration may be included, or some of these configurations may be omitted.

[0019] The CPU 10a is an example of a processor, and the RAM 10b or the ROM 10c stores It is a control unit that controls the execution of programs and performs data calculations and processing. a is, for example, a program (learning program) that performs learning using a predetermined learning model. The CPU 10a receives various data from the input unit 10e and the communication unit 10d. The data is then displayed on the display unit 10f and stored in the RAM 10b. do.

[0020] The RAM 10b is a memory unit in which data can be rewritten, and is, for example, a semiconductor memory. The RAM 10b may be configured with a memory element. Models, data on the parameters of each learning model, data on the features of the learning data These are just examples, and the RAM 10b may store other data. Other data may be stored, or some of these may not be stored.

[0021] The ROM 10c is a memory from which data can be read, for example, a semiconductor memory. The ROM 10c may be configured with a memory element. For example, the ROM 10c may store a learning program or It may store data that is not

[0022] The communication unit 10d is an interface that connects the information processing device 10 to other devices. The receiving unit 10d may be connected to a communication network such as the Internet.

[0023] The input unit 10e receives data input from the user, for example, a keyboard. It may include a keyboard and a touch panel.

[0024] The display unit 10f visually displays the results of calculations performed by the CPU 10a. For example, The display unit 10f may be configured by an LCD (Liquid Crystal Display). Displaying the information can contribute to XAI (eXplainable AI). f may display, for example, learning results and data related to learning.

[0025] The learning program can be read by a computer such as RAM 10b or ROM 10c. The information may be stored in a non-transitory recording medium, or may be provided by a device connected via the communication unit 10d. In the information processing device 10, the CPU 10a By executing the learning program, various operations described below are realized. The physical configuration is an example and does not necessarily have to be an independent configuration. The processing unit 10 is an LSI (Large-Scale Integrated Circuit) in which a CPU 10a, a RAM 10b, and a ROM 10c are integrated. The information processing device 10 may also include a GPU (Graphics Processing Unit). Equipped with a 3D processing unit (ASIC) and an ASIC (Application Specific Integrated Circuit) It may be possible.

[0026] The configuration of the information processing device 20 is the same as the configuration of the information processing device 10 shown in FIG. The information processing device 10 and the information processing device 20 are data processing devices. The input unit 1 only needs to have a CPU 10a, a RAM 10b, etc., which are basic components for performing the above operations. The input unit 10e and the display unit 10f may not be provided. The unit may be connected to the network using an interface.

[0027] <Processing configuration> 3 is a diagram showing an example of a processing block of the information processing device 20 according to the embodiment. The processing device 20 includes an acquisition unit 21, a learning unit 22, an output unit 23, and a storage unit 24. The device 20 may be configured as a general-purpose computer, a quantum computer, or the like.

[0028] The acquisition unit 21 acquires a predetermined data set including each learning data item associated with each correct answer data item. The first learning data associated with the first correct answer data is acquired from the acquired data. For example, the acquisition unit 2 1 is at least one first learning data set of a data set 24a stored in a storage unit 24. The first training data is associated with the first correct answer data. The set 24a may be a data set received from the server 10 and stored in the storage unit 24. .

[0029] The data set 24a may include, for example, at least image data, sequence data, and text data. Here, the image data includes either still image data or video data. Series data includes voice data, stock price data, etc.

[0030] The acquiring unit 21 also acquires the second supervised data assigned to the first learning data. For example, the acquisition unit 21 may acquire a user's input for the first learning data. The acquisition unit 21 acquires correct answer data annotated by the user. It may also have a function as a reception unit that receives solution data.

[0031] The learning unit 22 performs, for example, supervised learning, and performs supervised learning on a predetermined problem. The learning model 22a includes a predetermined learning model, a calculation unit 22b, and an adjustment unit 22c. 2a is a system that has not yet been trained using the training model or training data received from the server 10. It may also be an untrained model.

[0032] The predetermined learning model 22a is, for example, a prediction model, an image recognition model, a sequence data model, Analysis models, robot control models, reinforcement learning models, speech recognition models, speech generation models , image generation model, natural language processing model, etc. Also, a predetermined learning model Examples of Del22a include CNN (Convolutional Neural Network) and RNN (Recursive Neural Network). rrent Neural Network), DNN (Deep Neural Network), LSTM (Long Short-Term Memory), bidirectional LSTM, DQN (Deep Q-Network), VAE (Variational AutoEnc order), GANs (Generative Adversarial Networks), flow-based generative models Dell or the like.

[0033] In addition, the learning model 22a is a trained model that has undergone pruning and quantization. This includes models obtained by ion, distillation, or transfer. These are merely examples, and the learning unit 22 may use the machine learning of the learning model for problems other than these. The learning unit 22 may perform learning on the learning model 22 according to the characteristics of the data set to be learned. a may be selected and supervised learning may be performed using this learning model. The loss function used in is the squared error between the output of the learning model 22a and the correct data. The learning unit 22 may use a loss function such as a sigma function or a cross-entropy function. The hyperparameter is calculated using backpropagation until a certain condition is met so that the value of Repeat the learning process while adjusting the parameters.

[0034] The calculation unit 22b calculates the first learning data in the learning model 22a corresponding to a predetermined data set. For example, the calculation unit 22b inputs a learning model 2 that solves a regression problem and calculates prediction data. In the case of 2a, the calculation unit 22b calculates a predicted value as predicted data. In the case of the solving learning model 22a, a classification value is calculated as prediction data.

[0035] The adjustment unit 22c adjusts the first correct answer data set in advance in the learning data and the first correct answer data set given by the user. Depending on the result of the comparison with the second correct answer data, a loss function set in the learning model 22a is used. The weight assigned to the error between the second correct answer data and the predicted data is adjusted. For example, If the first correct answer data and the second correct answer data are different, the unit 22c The prediction error of the given training data is more strongly reflected in the training. Increase the weight of the

[0036] The learning unit 22 learns the learning model 22a by backpropagation using the output value of the loss function. For example, the learning unit 22 may generate learning data in which the first supervised answer data and the second supervised answer data are different from each other. The output value of the loss function is adjusted so that the error for In other words, the hyperparameters are adjusted by using the training data that has the user's characteristics. The output value of the loss function is reduced based on the output value adjusted to increase the prediction error. The hyperparameters are adjusted to reflect the user's characteristics more strongly in the learning. do.

[0037] By the above process, it is possible to efficiently individualize existing learning models. For example, Even if a user has few annotations, the prediction error of the training data that represents the user's characteristics can be reduced. By increasing the weighting on the differences, the learning model can be personalized efficiently. In addition, for users who do not want to provide their annotation results, It will also be possible to provide personalized learning models without providing application results.

[0038] The output unit 23 outputs information about the learning result. For example, the output unit 23 outputs Information about the learning model and information about the learning accuracy may be output to the server 10.

[0039] In addition, the adjustment unit 22c adjusts the first correct answer data and the second correct answer data at the time of the second or subsequent repeated learning. If the predicted data of the training data that is different from the solution data matches the second correct answer data, There is no need to adjust the weights after that. This is because the learning model at that point is able to accurately represent the user's characteristics. By reflecting this, the predicted data now matches the second correct answer data. This is because there is no longer any need to adjust the prediction error of the training data.

[0040] For example, when learning is repeated for the second time or later, the first correct answer data and the second correct answer data may differ. The learning data is input to the learning model. The adjustment unit 22c calculates the If the error between the predicted data and the second correct answer data is within a predetermined value, the predicted data is the second correct answer data. It is determined that the data matches or is similar to the solution data, and the adjustment of the weight for this learning data is stopped. This prevents prediction errors from becoming larger than necessary, allowing for appropriate learning to be repeated. You will be able to return it.

[0041] Furthermore, the adjustment unit 22c determines whether the comparison result between the first correct answer data and the second correct answer data is greater than the first correct answer data. If the data indicates that the second correct answer data is different from the first correct answer data, the weight used in the loss function is set to a predetermined value. For example, the adjustment unit 22c may use a squared error as the loss function. When the difference function is used, the weight wn in the following formula (1) is adjusted.

number

[0042] In this case, the adjuster 22c adjusts s When it is shown that the first correct answer data and the second correct answer data of the training data are different, ws Set to a value greater than 1. This will give you correct data that is different from the existing correct data. By increasing the error for the training data, the features of this training data are strengthened. This can be easily reflected in the learning model.

[0043] The adjustment unit 22c selects a loss function according to the type of predetermined problem to be solved by the learning unit 22. For example, in the case of a regression problem, the adjustment unit 22c may use the loss function of the above formula (1) In the case of a binary classification problem, the loss function is the following formula (2), and in the case of a multi-class classification problem, the loss function is the following formula (2). Select the loss function of formula (3). Note that the loss function described above is an example and is not limited to this. It is not something that can be done.

number

number

[0044] In addition, the adjustment unit 22c adjusts the loss as the error between the first correct answer data and the second correct answer data increases. For example, the adjustment unit 22c may adjust the weight wn used in the function by: By setting the weight wn based on the absolute value of the error value, the magnitude of the error value is reflected in the weight wn. You may do so.

[0045] This means that the more the correct answer data provided by the user differs from the existing correct answer data, the more likely it is that the correct answer data will be used. Since the training data contains the user's characteristics, the error for this training data is increased. As a result, the output value of the loss function becomes larger, and the hyperparameter -When adjusting parameters, the user characteristics contained in this training data are more strongly incorporated into the learning model. can be reflected in.

[0046] The acquiring unit 21 also acquires a predetermined number of first learning data from a predetermined data set. In this case, the acquiring unit 21 may acquire correct answer data for each of the first learning data. and acquiring each second correct answer data selected by a predetermined user from the options. For example, the acquisition unit 21 may ask the user to select the learning data items displayed on the screen in order. The annotation result (second correct answer data) may be received and acquired.

[0047] The calculation unit 22b calculates predicted data and an error for each of the first learning data, and the adjustment unit 22c multiplies the error by a weight that reflects the comparison result between the first correct answer data and the second correct answer data. The output value of the loss function is calculated.

[0048] This allows users to annotate a set number of training data in order, Individualize existing learning models while more strongly reflecting user characteristics based on annotation results. It becomes possible to

[0049] The acquisition unit 21 also acquires a predetermined data set from another information processing device (for example, the server 10). The output unit 23 may also acquire the user identification information of the predetermined user and the The second correct answer data assigned by the first learning data is associated with the first correct answer data and updated. The data set is output to another information processing device. The output destination of the data set does not necessarily have to be the same, but may be a different information processing device. It is also possible.

[0050] As a result, the server 10 allows each user to perform annotation and creates annotations for each user. The server 10 can generate a data set that associates the results of the collaboration. By specifying a user, it is possible to generate a learning model tailored to that user. do.

[0051] The acquisition unit 21 also acquires predetermined data acquired from another information processing device (for example, the server 10). In this case, the acquisition unit may acquire first training data included in the dataset. 21 is a block diagram showing a method for identifying a specific user (user ID) and a method for identifying a specific user (user ID) For example, the acquiring unit 21 may acquire second supervised data corresponding to the learning data. The data set in which each correct answer data of each user ID is associated is acquired from the server 10. You may do so.

[0052] As a result, the information processing device 20 selects a user ID according to the purpose or use, and It is possible to individualize existing learning models by using correct answer data corresponding to the user ID. It becomes Noh.

[0053] Each processing unit included in the information processing device 20 may be provided in the server 10. For example, Each user of each information processing device 20 performs annotation, and the individual learning model is The learning may be performed by the learning unit of the server 10.

[0054] <Data example> 4 is a diagram illustrating an example of a data set according to the embodiment. A specific user can add additional correct answer data to the training data that already has correct answer data. Represents a dataset that can be provided.

[0055] In the example shown in FIG. 4, the data set includes each first learning data. The data is then combined with the first correct answer data set in advance and the second correct answer data given by the user indicated by the user identification information. For example, if the user identification information is different, the answer data provided by different users will be This indicates that this is the second correct answer data.

[0056] The dataset shown in Figure 4 is an example, and only specific users provide correct data. If so, the data set may not include user identification information. The dataset contains information indicating the category of training data (images, sequence data, etc.) may be included in the

[0057] <User interface example> 5 is a diagram illustrating an example of an annotation according to the embodiment. The training data included in the dataset is music data, and the correct answer data is the user's perception of music. It's how you feel.

[0058] In the example shown in Figure 5, the title of each piece of music is displayed on the screen, and when you operate the title, the sound changes. For example, the user selects "calming" for "Music A." This indicates that you have selected "Uplifting" for "Music B." When the user operates the The second correct answer data is stored.

[0059] When the user completes the annotation as described above, the learning unit 22 performs learning. However, the learning model is personalized using the learning data used for annotation. At this time, the adjustment unit 22c compares the already assigned correct answer data (first correct answer data) with the In the case where the assigned correct data (second correct data) is different from the example shown in Figure 5 In order to more strongly reflect the characteristics of the music data in the learning, we adjusted the weights of the loss function. Adjust.

[0060] For example, in the example shown in FIG. 5, the second correct answer data assigned to music A and music C has already If the predicted data of music A is different from the first correct answer data, the predicted data of music A and the second correct answer data are compared. By increasing the weight of the error and the error between the predicted data of music C and the second correct answer data, The characteristics of Music A and Music B are more strongly reflected in the learning model.

[0061] FIG. 6 is a diagram illustrating an example of selecting a group in a data set according to an embodiment. In the example shown in Figure 1, the training data included in the dataset is divided into groups according to age. The dataset shown in Figure 6 is assigned a second correct answer data for each generation. In this case, the user identification information in the data set shown in Figure 4 includes a group ID. In this example, the category "Teenagers" is selected for dataset A.

[0062] A learning model for teenagers is created using learning data containing correct answers given by teenage users. When the rule is generated, the second correct answer data for teenagers is compared with the first correct answer data, which is the general correct answer. If they are different, the characteristics of the training data are strongly reflected in the individualization learning.

[0063] <Processing example> FIG. 7 is a sequence diagram showing an example of processing by the information processing device 20 according to the embodiment. In step S102, the acquisition unit 21 acquires a set of training data including the training data associated with each of the correct answer data. Acquire first training data associated with the first correct answer data from a predetermined data set. .

[0064] In step S104, the acquiring unit 21 acquires the second correct value assigned to the first learning data. Obtain the solution data.

[0065] In step S106, the calculation unit 22b calculates a learning model corresponding to a predetermined data set. The first learning data is input to the rule 22a to calculate the prediction data.

[0066] In step S108, the adjustment unit 22c adjusts the first correct answer data set in advance in the learning data. The learning model 22a is generated according to the comparison result between the data and the second correct answer data provided by the user. The weight assigned to the error between the second correct answer data and the predicted data used in the loss function set Adjust the brightness.

[0067] In step S110, the learning unit 22 performs backpropagation using the output value of the loss function. Learning is performed using learning model 22a.

[0068] By the above process, it is possible to efficiently individualize existing learning models. For example, Even if a user has few annotations, the prediction error of the training data that represents the user's characteristics can be reduced. By increasing the weighting on the differences, the learning model can be personalized efficiently. do.

[0069] The above-described embodiments are provided to facilitate understanding of the present invention and should not be construed as limiting the present invention. The elements included in the embodiment and their arrangement, materials, conditions, and shapes are not intended to be interpreted as examples. The dimensions and size are not limited to those shown as examples and can be changed as appropriate. In addition, the configurations shown in different embodiments may be partially substituted or combined with each other. do.

[0070] In the above embodiment, the learning unit 22 of the information processing device 20 may be implemented in another device. In this case, the information processing device 20 may instruct the other device to individualize the learning model. [Explanation of symbols]

[0071] 10, 20...information processing device, 10a...CPU, 10b...RAM, 10c...ROM, 10d ...communication unit, 10e...input unit, 10f...display unit, 21...acquisition unit, 22...learning unit, 22a...learning unit Learning model, 22b... calculation unit, 22c... adjustment unit, 23... output unit, 24... storage unit

Claims

1. An information processing method executed by an information processing device having a processor, The processor: Obtaining correct answer data annotated from training data used in supervised learning; inputting the learning data into a learning model used in the supervised learning to calculate predicted data; adjusting a weight assigned to an error between the annotated correct answer data and the predicted data, which is used in a loss function set in the learning model, according to a comparison result between the correct answer data associated with the learning data and the annotated correct answer data; training the learning model by backpropagation using the output value of the loss function in which the weight is used; outputting the trained learning model; An information processing method that performs the above.

2. An information processing device having a processor, The processor: Obtaining correct answer data annotated from training data used in supervised learning; inputting the learning data into a learning model used in the supervised learning to calculate predicted data; adjusting a weight assigned to an error between the annotated correct answer data and the predicted data, which is used in a loss function set in the learning model, according to a comparison result between the correct answer data associated with the learning data and the annotated correct answer data; training the learning model by backpropagation using the output value of the loss function in which the weight is used; outputting the trained learning model; An information processing device that executes the above.

3. A processor provided in an information processing device, Obtaining correct answer data annotated from training data used in supervised learning; inputting the learning data into a learning model used in the supervised learning to calculate predicted data; adjusting a weight assigned to an error between the annotated correct answer data and the predicted data, which is used in a loss function set in the learning model, according to a comparison result between the correct answer data associated with the learning data and the annotated correct answer data; training the learning model by backpropagation using the output value of the loss function in which the weight is used; outputting the trained learning model; A program that executes the following.

Images