Machine Learning Systems

Abstract

To provide a machine learning system that can easily improve performance at the time of transfer learning.SOLUTION: A machine learning system that, using a given dataset and a given prior learning model, performs learning using a parameter in the prior learning model to perform transfer learning that outputs a learnt model, comprises: a dataset storage part that stores one or more datasets; and a first learning part that performs learning using the datasets stored in the dataset storage part to generate the prior learning model, then stores the generated prior learning model in a prior learning model database. The dataset storage part stores tag information that is configured with either one or more of domain information indicating an object of data included in the dataset to be stored, class information indicating a class included in the data, or data acquisition condition information regarding the acquisition condition of the data in association with the dataset.SELECTED DRAWING: Figure 3

Description

[Technical field]

[0001] The present invention relates to a machine learning system that performs data processing by machine learning, and in particular to a machine learning system that is effective for constructing a machine learning model for a new data set at high speed and with high accuracy. [Background technology]

[0002] In machine learning technology for data processing, particularly in the multi-layered neural networks known as deep learning, a technique called transfer learning is frequently used, which uses past learning models and training data to improve performance on new data sets.

[0003] For example, Patent Document 1 describes a method for efficiently performing transfer learning by evaluating the relationship between the values ​​of the output layer of a trained neural network in a new dataset and the classes of the new dataset, and preferentially allocating output units that have a higher frequency of obtaining evaluation values ​​with the smallest difference from the classes of the evaluation items. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2016-143094 A Summary of the Invention [Problem to be solved by the invention]

[0005] Patent Document 1 provides a system suitable for transfer learning by appropriately assigning classes of a new data set to the output units of a trained neural network. However, since transfer learning is highly dependent on the properties of the data set used in advance for learning, there are cases in which it is difficult to obtain the effects of transfer learning.

[0006] First, when the number of classes in a data set to be newly trained is larger than the number of classes previously trained, the method provided in Patent Document 1 cannot be used.

[0007] Secondly, it is difficult to obtain accurate performance for classes that have not been trained in advance. For example, if only tool classes such as hammers and nails have been trained, it is difficult to extract enough information necessary to recognize animal classes such as dogs and cats.

[0008] Third, even if the class granularity of the new dataset is more finely divided than that of the trained model, it is difficult to extract the information necessary to recognize the new class. For example, in a case where animal species such as dogs and cats are recognized in pre-training and dog species are subdivided in the new dataset, the trained model is likely to extract similar features because many of the new datasets are classified into a single class, dog, in the original dataset, making it difficult to obtain the benefits of transfer learning.

[0009] Fourth, when a task is different from the dataset used for prior training, most of the machine learning model needs to be initialized and trained. For example, when performing object detection on a new dataset after performing image classification in prior training, many layers of a neural network need to be added and trained in addition to the prior training model during transfer learning, which may result in prolonged training times.

[0010] As described above, when performing transfer learning, the performance on a new dataset varies greatly depending on the dataset and learning task used in pre-learning. Therefore, the object of the present invention is to provide a machine learning system that can easily perform an appropriate pre-learning model and improve performance during transfer learning. [Means for solving the problem]

[0011] A machine learning system that performs transfer learning using a given dataset and a given pre-learning model, learning using parameters of the pre-learning model, and outputting a trained model, comprising a dataset memory unit that stores one or more datasets, and a learning unit that performs learning using data stored in the dataset memory unit and generates a pre-learning model database, and the machine learning system is characterized in that the pre-learning model is created using data from the pre-learning model database, and the dataset memory unit associates and stores the dataset with tag information consisting of one or more of domain information indicating the object of the data contained in the stored dataset, class information indicating the class contained in the data, and data acquisition condition information regarding the conditions for acquiring the data. Effect of the Invention

[0012] According to the present invention, in a machine learning system that obtains a machine learning model through transfer learning, it is possible to easily configure an appropriate pre-learning model, and to achieve high accuracy and high speed in learning with a new dataset.

[0013] Problems, configurations and effects other than those described above will become apparent from the following description of the embodiments. [Brief description of the drawings]

[0014] [Figure 1] FIG. 1 is a diagram showing an example of the configuration of a multilayer neural network. [Diagram 2] FIG. 1 is a diagram showing an example of the configuration of transfer learning. [Diagram 3] FIG. 1 is a diagram showing an example of the configuration of a machine learning system according to a first embodiment of the present invention. [Figure 4] 4 is a flowchart showing the behavior of the machine learning system according to the first embodiment of the present invention. [Diagram 5] FIG. 4 is a diagram showing the configuration of a dataset storage unit. [Figure 6] FIG. 1 shows the configuration of a pre-training model database. [Figure 7] FIG. 4 is a diagram showing the configuration of a dataset relevance evaluation unit. [Figure 8] 11 is a flowchart showing the operation of a data set relevance evaluation unit. [Figure 9] FIG. 1 is a block diagram showing the configuration of a machine learning system including a user interface according to a first embodiment of the present invention. [Figure 10] FIG. 13 is a diagram showing a user interface of a data set integration unit. [Figure 11] FIG. 4 is a diagram showing a user interface for a learning unit. [Figure 12] 10 is a flowchart showing the operation of a learning unit in the second embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0015] Hereinafter, an embodiment of the present invention will be described with reference to the drawings, and detailed description of overlapping parts will be omitted.

[0016] Before describing the embodiments of the present invention, the configuration of a transfer learning function, which is a prerequisite for applying the present invention, will be described with reference to FIGS. 1 and 2. FIG.

[0017] First, Figure 1 shows an example of the configuration of a multi-layer neural network called deep learning. As shown in Figure 1, a deep learning model is composed of multiple layers. These are an input layer that receives data, a weighting layer that extracts features from the input data and the output of other weighting layers, and an output layer that uses the extracted features to output the processing results. In this invention, one or more layers before the output layer are referred to as feature extraction layers.

[0018] 2 shows an example of the configuration of transfer learning. In transfer learning, a learning unit 203 uses a given data set 201 and a given pre-trained model 202 to perform learning using parameters of the pre-trained model, and outputs a trained model 204.

[0019] In order to maximize the effect of transfer learning in such a configuration, it is necessary to train the pre-training model 202 with a dataset that has a larger number of classes than the dataset 201 to be learned, or that contains classes similar to those contained in the dataset 201, or that has finer class granularity.

[0020] However, constructing such a dataset for the pre-training model 202 is generally difficult and requires a lot of man-hours. Even when combining multiple datasets, if the tasks for each dataset are different, they cannot be simply combined. For example, image classification and object detection have different annotation formats, so it is difficult to simply integrate datasets.

[0021] In addition, even if the tasks are the same, simply combining them will not provide the benefits of transfer learning. For example, when training an image classification model to classify dog ​​breeds, the benefits of transfer learning are small even if you first classify images by observing samples under a microscope.

[0022] The present invention has been made in consideration of these circumstances, and provides a machine learning system that efficiently manages multiple data sets and easily performs appropriate pre-learning using appropriate data sets. EXAMPLES

[0023] Below, we will explain a machine learning system that can easily secure a dataset for the pre-training model 202 in the transfer learning configuration shown in Figures 1 and 2.

[0024] An example of the configuration of a machine learning system according to the first embodiment of the present invention is shown in Fig. 3. The machine learning system in Fig. 3 is composed of a dataset storage unit 301, a dataset integration unit 302, a self-instruction learning unit 303, and a pre-learning model database 304, and data accumulated in the pre-learning model database 304 is used in pre-learning as a dataset for the pre-learning model in Fig. 2.

[0025] In FIG. 3, a data set storage unit 301 stores a plurality of data sets in association with tag information based on the domain, data class, data acquisition conditions, and the like of each data set.

[0026] Here, the domain refers to the subject of the data, such as animals, natural objects, man-made objects, etc. For example, in the case of product inspection, this refers to inspection object information, such as industrial products, medical images, and semiconductor images, as well as information on lower-level inspection processes, such as cell inspection and etching processes.

[0027] The data class is class information included in each data. For example, this is information about a class included in the domain, such as dogs or cats, and in the case of an inspection process, this may be the type of cell or bacteria to be inspected, or the type of defect.

[0028] The data acquisition conditions are information about the acquisition conditions of the data included in the dataset, such as information about the data acquisition device, such as an optical camera, optical microscope, electron microscope, radar, infrared, X-ray, or acoustic, or acquisition conditions such as magnification, exposure time, resolution, number of integrations, and sampling frequency.

[0029] Fig. 4 is a flowchart showing the behavior of the machine learning system according to the first embodiment of the present invention shown in Fig. 3. In the series of processes in Fig. 4, first, in processing step S401 corresponding to the processing of the dataset integration unit 302, a specified dataset or a dataset corresponding to specified tag information is acquired from the dataset storage unit 301. Then, in processing step S402 corresponding to the processing of the dataset integration unit 302, the datasets acquired from the dataset storage unit 301 are integrated to create an integrated dataset.

[0030] Here, integration of datasets refers to extracting only input data from multiple datasets acquired in processing step S401 to create a larger dataset. As a result, for example, when two datasets A and B are acquired in processing step S401, the number of data included in the integrated dataset after integration is the sum of the number of data included in dataset A and the number of data included in dataset B. At this time, when there is data included in both dataset A and dataset B, duplication may be avoided and the number of unique data may be set to the number of data included in the integrated dataset.

[0031] In this case, by using a method in the self-instruction learning unit 303 that does not use the teaching information contained in the dataset acquired in processing step S401, it is possible to easily create an integrated dataset that integrates the multiple datasets acquired in processing step S401 even if the format or class granularity of the teaching information contained in the multiple datasets acquired in processing step S401 is different.

[0032] Note that the format of the instruction information here may differ for each task of each dataset. For example, in the case of image classification, the instruction information may be composed of class information such as a class name and class number for the image; in the case of object detection, the instruction information may be composed of information linking the position information of an object in the image with class information; and in the case of segmentation, the instruction information may be composed of class information and instance information associated with each pixel.

[0033] In addition, in the case of image conversion tasks such as noise removal, and 3D reconstruction tasks such as depth estimation from images, class information is not given. In addition, even if there is no teaching information for some data sets, the present invention can be used because it does not use teaching information.

[0034] In addition, the method of integrating the datasets may be changed using the teaching information of the multiple datasets obtained. For example, in the case of classification of abnormal data, data imbalance occurs frequently in which the number of data for each class differs within a dataset due to the existence of anomalies that occur easily and anomalies that occur rarely. In such a case, an integrated dataset may be constructed so that each class is used with as equal a probability as possible during pre-learning. In other words, the sampling probability of each data in the integrated dataset may be changed, or the data may be adjusted so that the number of data for each class is approximately the same, and then the data may be integrated into an integrated dataset.

[0035] Next, in process step S403, which corresponds to the process of the self-instruction learning unit 303, self-instruction learning is performed using the integrated data set created in process step S402. Here, the network structure and learning conditions of the neural network to be subjected to self-instruction learning are set in advance.

[0036] Here, self-instruction learning refers to learning using, for example, SimCLR (Non-patent literature; Tien Chen, Simon Kornblith, Mohammad Norouzi, Geoffrey Hinton, A Simple Framework for Contrastive Learning of Visual Represetations, International Conference on Machine Learning, 2020).

[0037] SimCLR is a technique that applies two different random transformations to the data contained in the training dataset, and then trains the model so that images obtained from the same image and the two random transformations have the same features, and images obtained from different images and random transformations have different features.

[0038] SimCLR can be interpreted as a method of learning to identify all data contained in a dataset. In other words, it can be interpreted as learning a dataset with the finest class granularity, regardless of the class granularity of the dataset used. Therefore, pre-training with SimCLR requires solving a task that is more difficult than expected, so a detailed feature extraction layer can be learned, and it is possible to obtain an appropriate pre-trained model when training with a dataset similar to a new dataset.

[0039] However, it is difficult to construct an appropriate pre-trained model by simply integrating multiple datasets and performing self-teaching learning such as SimCLR. Generally, neural networks are trained by stochastic gradient descent or its variants. In stochastic gradient descent, a portion of data is obtained from the entire dataset as a mini-batch, and the mini-batch is used to calculate and train the neural network parameters.

[0040] In contrast, self-instruction learning such as SimCLR learns to identify each piece of data contained in the acquired mini-batch. If datasets are integrated while ignoring tag information such as the domain of the dataset, the mini-batch will be composed of datasets from various domains, and self-instruction learning will be able to identify each piece of data by identifying the domain, which will reduce the effect of integrating the datasets.

[0041] For example, if a dataset of images of general objects such as dogs and cats captured by an optical camera is combined with a dataset of sample observations by a microscope to form a mini-batch consisting of general object images and microscope images, the image features of the two are significantly different, so that the two can be identified with simple information. In such a case, the value of the loss function during learning will be small if only domain information can be identified. Therefore, it is desirable that the dataset acquired in processing step S401 is data from a similar domain.

[0042] Conversely, performing self-instruction learning using only a single dataset without integrating datasets from multiple domains also reduces the effectiveness of pre-learning through self-instruction learning. For example, even if the number of data included in the dataset used for pre-learning is small, it becomes possible to identify individual data using only relatively simple features.

[0043] From the above viewpoints, it is desirable that the data set used for the pre-learning performed by the self-instruction learning unit 303 is configured with limited tag information and contains as much data as possible. Note that although the configuration using SimCLR has been described, other self-instruction learning methods may also be used.

[0044] In addition, in processing step S404, which corresponds to the processing of the self-instruction learning unit 303, the parameters of the pre-training model obtained by self-instruction learning in processing step S403 are stored in the pre-training model database 304 together with information regarding the network structure, learning conditions, and the datasets used in the integrated dataset created in processing step S402.

[0045] With this configuration, the present invention can provide a machine learning system that efficiently manages multiple data sets and easily performs appropriate pre-learning using appropriate data sets. That is, the system stores each data set in association with tag information in the data set storage unit 301, configures an integrated data set by acquiring and integrating data sets based on one or more specified tag information, and performs pre-learning of a neural network capable of detailed feature extraction by self-instruction learning that is assumed to be more difficult than a new task.

[0046] In this case, a dataset suitable for self-instruction learning can be constructed by managing the dataset in association with tag information. In other words, by constructing the largest possible dataset with limited domain information, data classes, and data acquisition conditions using tag information, it becomes possible to learn a more detailed feature extraction layer through self-instruction learning.

[0047] According to this configuration, when used for multiple inspection processes in a manufacturing process, data sets from the multiple inspection processes are integrated and pre-learning is performed, making it possible to efficiently perform transfer learning in a new inspection process.

[0048] Fig. 5 shows an example of the data configuration of the dataset storage unit 301 used in the present invention. A plurality of datasets 502 (502A, 502B in the illustrated example) are stored in the dataset storage unit 301. Dataset-specific information is assigned to each of the datasets 502A, 502B. This is assigned individually to each dataset, as shown by 502A and 502B in Fig. 5. This may be the name of the dataset, a unique character string or number string, or information consisting of the name of the dataset, a unique character string, creation date and time, registration date and time, version information, etc.

[0049] Each data set 502A, 502B includes input data 503, instruction information 504, data set format information 505, tag information 506, and creation information 507, which are stored in association with each other. These are also stored for each data set. That is, as shown in Fig. 5, input data 503A, 503B, instruction information 504A, 504B, data set format information 505A, 505B, tag information 506A, 506B, and creation information 507A, 507B are created for each data set for data set A and data set B, respectively.

[0050] First, input data 503 is input data included in each data set 502. For example, this corresponds to an image obtained by photographing an object to be observed, or an image after processing.

[0051] The instruction information 504 is instruction information corresponding to the input data included in each data set 502. This is configured so that the correspondence with each input data can be understood according to a specific rule. In addition, in the case of a data set that does not have instruction information, the instruction information 504 may not be necessary.

[0052] The dataset format information 505 stores information about the format of the dataset 502. This is, for example, information about the task of the dataset, such as whether the dataset is in an image classification format or an object detection format, or information about the storage format of the dataset depending on the task. This dataset format information 505 may be configured to know the correspondence between the input data 503 and the instruction information 504. Furthermore, this dataset format information 504 may be used when referring to class information of each data when integrating the integrated datasets in step S402.

[0053] Tag information 506 is tag information based on the domain, data class, data acquisition conditions, etc. of each data set 502. Domain information indicates the object of the data included in the data set, class information indicates the class included in the data, and data acquisition condition information relates to the data acquisition conditions.

[0054] The creation information 507 is information that is added when the dataset 502 is created. For example, it is information about the date or time when the dataset was created, or it is version information of the dataset. Here, the version information of the dataset 502 may include information that specifies a specific process performed on the dataset. For example, this may be information that a specific preprocessing has been applied, or information on the teachings has been confirmed by multiple people.

[0055] When the dataset storage unit 301 contains multiple datasets that are the same but have different creation dates and times or version information, the dataset storage unit 301 outputs the data to the dataset integration unit 302 based on a preset rule using the creation information 507. The preset rule here is to output the latest dataset, to output one whose version information contains a specific character string, etc.

[0056] In this way, various information is stored in the dataset storage unit 301, but in short, the present invention stores a dataset and tag information 506 in association with each other in the dataset storage unit 301, in which case the dataset is composed of input data 503 and instruction information 504, and the tag information 506 can be said to be composed of one or more of domain information indicating the object of the data contained in the input data 503, class information indicating the class contained in the data, and data acquisition condition information relating to the data acquisition conditions. A further difference is that the dataset is information obtained from the observed object, while the tag information is human information formed by a user's appropriate artificial input according to the situation of the observed object.

[0057] Fig. 6 is a diagram showing an example of the data configuration in the pre-training model database 304 in Fig. 3. The pre-training model database 304 is composed of model information 601, used dataset information 602, network structure 603, parameters 604, and learning conditions 605.

[0058] Among these, model information 601 (601A, 601B) is information assigned to each model. This may be the name of the model, a unique character string or a number string, or information composed of the model name, unique character string, creation date and time, registration date and time, version information, etc. As shown in Fig. 6, it is assigned individually to multiple models, such as model A 601A and model B 601B.

[0059] Furthermore, the used data set information 602, the network structure 603, the parameters 604, and the learning conditions 605 are managed for each model information 601. That is, the used data set information 602A, 602B, the network structures 603A, 603B, the parameters 604A, 604B, and the learning conditions 605A, 605B are assigned to each of the model A (601A9) and the model B (601B), and are stored.

[0060] The dataset usage information 602 stores information about the datasets and integration conditions integrated by the dataset integration unit 302 when training each pre-trained model 202. This may be, for example, dataset-specific information about the integrated datasets or specified tag information 506. When performing transfer learning with a new dataset, the user can check the tag information learned by each model by referring to the dataset usage information 602 linked to each model, and select a pre-trained model appropriate for training the new dataset.

[0061] Information about the network structure of each model is stored in the network structure 603. Since the network structure affects the recognition accuracy and processing time, the user can select a model having a network structure that meets the requirements of a new dataset by checking the network structure 603.

[0062] The parameters of the pre-learning model learned by the self-instruction learning unit 303 are stored in the parameters 604. The user can realize efficient transfer learning by using these parameters.

[0063] The learning conditions 605 store the learning conditions used in the pre-learning in the self-teaching learning unit 303. This is, for example, the number of parameter updates, or information on the batch size, the learning rate and learning rate schedule, the stochastic gradient descent method used, and the like.

[0064] In addition, in order to efficiently select datasets to be integrated in the dataset integration unit 302, the relevance between the datasets stored in the dataset storage unit 301 may be evaluated, and datasets with similar characteristics may be integrated.

[0065] FIG. 7 shows the configuration of a dataset relevance evaluation unit 701 that evaluates the relevance between datasets, and FIG. 8 shows the processing flow of the dataset relevance evaluation unit 701.

[0066] The dataset relevance evaluation unit 701 in Fig. 7 evaluates the degree of similarity of the data included in each dataset. Since it is important to integrate many similar datasets in the pre-learning in the self-instruction learning unit 303 in Fig. 3, it is possible to perform efficient pre-learning by evaluating the similarity between the datasets in advance.

[0067] 8, first, in processing step S801, the dataset relevance assessment unit 701 acquires datasets stored in the dataset storage unit 301. Here, tag information 506 associated with each dataset may be used to acquire only datasets having specific tag information, or conversely, datasets may be acquired without any restriction based on tag information.

[0068] In processing step S802, the data set relevance assessment unit 701 performs self-instruction learning using the data set acquired in processing step S801. The self-instruction learning performed here may be the same as the self-instruction learning used by the self-instruction learning unit 303, or may be a different method.

[0069] In process step S803, the data set relevance assessment unit 701 extracts features from the data included in the data set acquired in process step S801 using the machine learning model obtained by self-instruction learning in process step S802.

[0070] In processing step S804, the dataset relevance assessment unit 701 assesses the similarity of the features for each dataset using the features obtained in processing step S803. Here, the assessment of the similarity of the features for each dataset may be a value calculated from a distance obtained by a distance function such as Euclidean distance or cosine distance between the average values ​​of the features for each dataset, or a value calculated from an amount of information described by an amount of information such as the Kullback-Leibler information between the distributions of the features for each dataset.

[0071] In this way, the dataset relevance assessment unit 701 can assess the relevance of the datasets, such as the degree of similarity between the datasets, by evaluating the relationship between the feature quantities of each dataset obtained when multiple datasets are integrated and self-instruction learning is performed. Using this information, the user can appropriately select the datasets to be integrated by the dataset integration unit 302.

[0072] Fig. 9 shows an example of the configuration of the present invention including a user interface 901. In Fig. 9, a component consisting of a dataset 201, a pre-trained model 202, a learning unit 203, and a trained model 204 is a transfer learning mechanism of a general configuration to which the present invention shown in Fig. 1 can be applied.

[0073] On the other hand, the components consisting of the dataset storage unit 301, the dataset integration unit 302, the self-instruction learning unit 303, and the pre-learning model database 304 constitute the dataset creation mechanism for the pre-learning model according to the present invention shown in Fig. 3. The components consisting of the dataset storage unit 301 and the dataset relevance evaluation unit 701 constitute the dataset relevance evaluation mechanism shown in Fig. 7.

[0074] These mechanisms are organically linked by the user interface 901, and the user can make various settings, give operational instructions, check results, etc. Specifically, for example, the user can use the user interface 901 to select datasets to be integrated in the dataset integration unit 302 and to select the pre-training model 202 to be used in the training unit 203.

[0075] An example of a display screen serving as a user interface for setting up pre-learning is shown in Fig. 10. An example of a screen configuration used for setting up pre-learning includes a dataset filter unit 1001, a dataset selection unit 1002, and a dataset relevance display unit 1003 as information display units, and a learning method selection unit 1004, a network structure selection unit 1005, and a learning start unit 1006 as operation instruction units such as button operations.

[0076] This screen enables the user to select a data set to be used for pre-learning, and to select the learning method and network structure for self-instruction learning to be used in the self-instruction learning unit 203.

[0077] In the dataset filter section 1001, the user can operate the datasets displayed in the dataset selection section 1002 and the dataset relevance display section 1003 by inputting tag information such as the domain and data class for which pre-learning is desired, and the data acquisition conditions. This makes it possible to efficiently select the dataset to be used without displaying unrelated datasets.

[0078] In the dataset selection unit 1002, each dataset is displayed together with the associated tag information 506, and the user can select a dataset to be used for pre-learning. In addition, a brief description of the dataset may be displayed in the dataset selection unit 1002 by storing the description in association with the dataset in the dataset storage unit 301.

[0079] The dataset relevance display unit 1003 displays values ​​obtained by mapping the feature amounts of each dataset evaluated by the dataset relevance evaluation unit 701 in two or three dimensions. Here, mapping the feature amounts of the datasets in two or three dimensions is realized by, for example, manifold learning.

[0080] By changing the mark displayed for each dataset, it becomes easier for the user to check the similarity of the features between the datasets. For example, in the example of Fig. 10, the features of three datasets are displayed in the dataset relevance display section 1003-a, and the features related to dataset 1 indicated by circles and the features related to dataset 3 indicated by triangles are in a close area, but the features related to dataset 2 indicated by diamonds are in an area distant from dataset 1 and dataset 3. In such a case, if the user has decided to use dataset 1, he or she can easily identify that the similar dataset is dataset 3, and can easily consider whether to integrate them.

[0081] A thumbnail of each data may be displayed together with the mark of each feature amount. Also, the similarity between the data sets evaluated by the data set relevance evaluation unit 701 may be displayed in a table format like the data set relevance display unit 1003-b to present similar data sets to the user.

[0082] In the learning method selection section 1004, the user can select the learning method and learning conditions for the self-instruction learning to be used in the self-instruction learning section 203.

[0083] In the network structure selection section 1005 , the user can select the network structure of the pre-learning model to be used in the self-instruction learning section 203 .

[0084] In the learning start unit 1006, when the user indicates that learning should be started, pre-learning is performed using the dataset selected in the dataset selection unit 1002, the learning method and learning conditions selected in the learning method selection unit 1003, and the network structure selected in the network structure selection unit 1005, and the pre-learning model is stored in the pre-learning model database 304.

[0085] Figure 11 shows an example of a display screen that serves as a user interface when performing transfer learning.

[0086] An example of a screen configuration used in transfer learning includes a learning data selection unit 1101, a learning dataset tag display unit 1102, and a recommended model display unit 1103 as information display units, and a pre-learning model selection unit 1104, a learning condition setting unit 1105, and a learning start unit 1106 as operation instruction units such as button operations.

[0087] In the learning data selection unit 1101, the user selects a learning target data set 201. This may be newly registered, or may be stored in the data set storage unit 301 in advance.

[0088] The learning dataset tag display unit 1102 displays tag information related to the learning target dataset 201. If the learning target dataset 201 selected by the learning data selection unit 1101 is new, this may be input by the user, or if it has been stored in advance in the dataset storage unit 301, the tag information 506 stored in the dataset storage unit 301 may be displayed.

[0089] The recommended model display unit 1103 displays models recommended based on the evaluation results of the dataset relevance evaluation unit 701. This may not be displayed if there is no applicable model, may display multiple models if there are multiple models, may limit the models to be displayed under specific conditions such as the most recommended model, or may display them in order of recommendation level.

[0090] Here, the recommendation level of each model is the degree of similarity between the feature amount obtained by the model through self-instruction learning in the dataset relevance evaluation unit 701 for the dataset selected by the learning data selection unit 1101 and the feature amount obtained from the dataset 602 used for learning each pre-learning model. A dataset learned from a dataset with high similarity is also effective in transfer learning.

[0091] In the pre-training model selection unit 1104, the user can select the pre-training model 202 to be used for transfer learning from the pre-training model database 304. Here, model information 601, data set information 602 used, network structure 603, etc. stored in the pre-training model database 202 are displayed. In addition, if evaluation by transfer learning has already been performed, evaluation indicators such as the accuracy rate may also be displayed.

[0092] In the learning condition setting unit 1105, the user can set the learning conditions for transfer learning.

[0093] When the learning start unit 1106 is selected by the user, transfer learning is executed using the selected data set 201, pre-training model 202, and learning conditions. EXAMPLES

[0094] According to the present invention shown in the first embodiment, if the dataset storage unit 301 contains a dataset that sufficiently covers the data contained in the new dataset 201, the feature extraction layer may have been sufficiently trained in advance. In such a case, a high-precision trained model 204 can be obtained quickly by newly training only the output layer.

[0095] In this embodiment, a configuration for further speeding up learning during transfer learning in such a case will be described.

[0096] Stochastic gradient descent is usually used for training neural networks, but it requires many parameter updates before convergence. Therefore, it is possible to speed up training by using optimization methods such as the quasi-Newton method, the natural gradient method, or their variants, which converge with fewer parameter updates.

[0097] A flow for learning only the output layer at a higher speed is shown in Fig. 12. In the first processing step S1201 in Fig. 12, the learning unit 203 reads the network structure 603 and the parameters 604 from the pre-learning model 202.

[0098] In processing step S 1202 , the learning unit 203 acquires data from the data set 201 .

[0099] In process step S1203, the learning unit 203 uses the feature extraction layer of the read pre-learning model 202 to extract features from the data acquired in process step S1202.

[0100] In processing step S1204, the learning unit 203 temporarily stores the acquired data and feature amounts.

[0101] In processing step S1205, the learning unit 203 checks whether feature extraction has been completed for all data in the dataset 201, and if not, returns to processing step S1202, and if completed, proceeds to processing step S1206.

[0102] In processing step S1206, parameters of the output layer are determined using the data and features temporarily stored in the learning unit 203. This is done by learning the output layer using a quasi-Newton method such as the L-BFGS method or a natural gradient method, using teaching information contained in the extracted features and related data.

[0103] In processing step S1207, the learning unit 203 outputs the output layer parameters determined in processing step S1206 together with the network structure 603 and the feature extraction layer parameters 604 of the pre-training model 202 as the trained model 204.

[0104] The present invention is not limited to the above-mentioned embodiment, but includes various modified examples. For example, the above-mentioned embodiment has been described in detail to aid in understanding the present invention, and is not necessarily limited to those including all of the configurations described. It is also possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, and it is also possible to add the configuration of another embodiment to the configuration of one embodiment. It is also possible to add, delete, or replace a part of the configuration of each embodiment with another configuration. [Explanation of symbols]

[0105] 201…Dataset 202…Pre-learning model 203…Learning Department 204…Trained model 301...Data set storage unit 302…Dataset Integration Department 303…Self-instruction learning section 304…Pre-training model database

Claims

1. A machine learning system that performs transfer learning by using a given dataset and a given pre-trained model to learn using parameters of the pre-trained model and output a trained model, A data set storage unit that stores one or more data sets, and a first learning unit that performs learning using the data sets stored in the data set storage unit to generate the pre-trained model, and stores the generated pre-trained model in a pre-trained model database, the data set storage unit stores the data set in association with tag information composed of at least one of domain information indicating an object of data included in the data set to be stored, class information indicating a class included in the data, and data acquisition condition information regarding an acquisition condition of the data; a dataset integration unit that integrates a plurality of datasets to be stored in the dataset storage unit and outputs the integrated dataset; a dataset relevance assessment unit that assesses the similarity between the datasets stored in the dataset storage unit; A user interface for use by a user; a dataset relevance display unit of the user interface displays values ​​obtained by mapping the feature quantities of each dataset evaluated by the dataset relevance evaluation unit in two or three dimensions; A machine learning system characterized in that the dataset integration unit retrieves and integrates a dataset specified by a user via the user interface or a dataset corresponding to specified tag information from the dataset storage unit, and outputs it to the first learning unit as an integrated dataset.

2. 2. The machine learning system of claim 1, The first learning unit performs pre-learning by self-teaching learning using the integrated data set, and outputs the pre-learning model to the pre-learning model database; A machine learning system characterized in that the pre-training model database stores information regarding the dataset used in the integrated dataset used for training by the first learning unit, or tag information used in the integrated dataset, in association with the pre-training model.

3. 3. The machine learning system of claim 2, A second learning unit that performs transfer learning using the pre-trained model stored in the pre-trained model database and a given dataset and outputs a trained model, A machine learning system characterized in that the second learning unit learns only the output layer of a pre-learned model.

4. 4. The machine learning system of claim 3, The machine learning system is characterized in that the second learning unit performs learning using a quasi-Newton method or a natural gradient method and an extension thereof.

Images