JP7471135B2 - Learning model generation system and learning model generation method - Google Patents

Description

The present invention generally relates to generating models.

AI (artificial intelligence) models implemented in on-site devices (edge devices) provide feedback to users and take other actions. Such AI models need to be able to realize the functions required in the field.

In this regard, a technology has been disclosed in which an edge device requests a trained AI model managed in a cloud environment based on information such as the type of AI model and the specifications of the edge device (see Patent Document 1).

AI models need to be adapted to the on-site environment, but Patent Document 1 makes no mention of adapting the AI model to the on-site environment.

In this regard, AI model updating technologies such as Google Federated Learning are known for adapting AI models to the field. In AI model updating technologies, a server device distributes an AI model to a client, the client terminal learns the AI model, and returns the learned results to the server device, which then updates the AI model.

US Patent Application Publication No. 2019/0042955

While AI model updating technology can adapt an AI model to the on-site environment, it has the problem that it cannot generate an AI model that achieves functions different from those of the current AI model.

The present invention was made in consideration of the above points, and aims to propose a learning model generation system that can appropriately generate models.

In order to solve this problem, the present invention provides a storage unit that stores each of the multiple layers that make up an existing learning model, which is a trained model; a generation unit that acquires layers corresponding to each of the multiple layers from the storage unit based on specification information in which the multiple layers are specified, and combines the acquired layers to generate a learning model; and an output unit that outputs the learning model generated by the generation unit.

With the above configuration, layers of existing learning models are combined to generate a new learning model, thereby reducing the time it takes to generate a learning model.

According to the present invention, a model can be generated appropriately. Problems, configurations, and advantages other than those described above will become clear from the following description of the embodiment of the invention.

FIG. 1 is a diagram illustrating an example of the configuration of a learning model generation system according to a first embodiment. FIG. 2 is a diagram illustrating an example of a hardware configuration of a learning model generating device according to a first embodiment. FIG. 4 is a diagram showing an example of model information according to the first embodiment; FIG. 2 is a diagram illustrating an example of a learning model according to the first embodiment. FIG. 2 is a diagram showing an example of a screen according to the first embodiment; FIG. 2 is a diagram illustrating an example of a series of processes according to the first embodiment. FIG. 11 is a diagram illustrating an example of a generation process according to the first embodiment. A diagram showing the relationship between the accuracy of a learning model and the time required to generate the learning model according to the first embodiment. A figure showing an example of the configuration of a learning model generation system according to a second embodiment. A figure showing an example of the configuration of a learning model generation system according to a third embodiment. 13 is an example of model specification information according to the third embodiment. FIG. 13 is a diagram illustrating an example of model information according to the third embodiment. FIG. 13 illustrates an example of a generation process according to the third embodiment;

(1) First Embodiment Hereinafter, one embodiment of the present invention will be described in detail. In this embodiment, the generation of a learning model will be described.

A learning model is a model of a neural network that has been trained (trained) to obtain a specified accuracy. A learning model is, for example, a trained deep neural network model. A learning model is configured by combining multiple layer models. A layer model consists of one or more layers that perform output according to input. In other words, a layer model may be one layer or multiple layers. A layer model is configured from a network structure (hereinafter referred to as "network") and parameters set in the network. A network includes information such as node definitions and node connection relationships. Parameters include information such as the weights of edges connecting nodes and node biases.

The layer models may be provided corresponding to predetermined categories. Examples of categories include the target when solving the problem set in the learning model, i.e., categories for extracting the characteristics of the target of inference in the learning model, categories for the inference, etc. Inference in the learning model includes object detection, posture detection, angle detection, distance measurement, etc., used for supporting workers on factory lines, monitoring facilities, autonomous driving, etc.

In the learning model generation system of this embodiment, layer models of existing learning models (i.e., layers consisting of networks and parameters) are linked to predetermined categories and stored in a storage device. The learning model generation system generates a new learning model by combining the existing layer models specified from the storage device based on specification information including information indicating layer models specified for each category by a user of the learning model generation system.

In the learning model generation system, a new dataset to which a label (correct answer data) for a new learning model is added may be further linked to the new learning model and stored in a storage device in addition to an existing dataset used to train an existing learning model. The learning model generation system may use the new dataset to train (re-train) a new learning model. The learning model generation system may also re-train a new learning model using an existing dataset stored in a storage device.

In addition, the learning model may be a model learned by supervised learning, a model learned by unsupervised learning, or a model learned by reinforcement learning. In the following, an example will be described in which the model is learned by supervised learning.

In the learning model generation system, the layer models (network and parameters) of the new learning model may be linked by category and stored in a storage device. In addition, in the learning model generation system, the layer models of the new re-learned learning model may be linked by category and stored in a storage device. In the learning model generation system, the new dataset used for re-learning may be further linked to the new learning model and stored in a storage device.

According to the above configuration, the optimal layer model for a new learning model can be reused from multiple existing learning models, so the learning model generation system can generate a learning model in a short period of time. Also, for example, when multiple learning models with different functions are stored as existing learning models, the user can generate a learning model with a different function from the existing learning models.

The learning models generated by the learning model generation system can be applied to a variety of systems, such as various industrial systems and mobility systems.

Next, an embodiment of the present invention will be described with reference to the drawings. The examples are illustrative for explaining the present invention, and some parts have been omitted or simplified as appropriate for clarity of explanation. The present invention can also be implemented in various other forms. Unless otherwise specified, each component may be singular or plural.

The position, size, shape, range, etc. of each component shown in the drawings may not represent the actual position, size, shape, range, etc., in order to facilitate understanding of the invention. Therefore, the present invention is not necessarily limited to the position, size, shape, range, etc. disclosed in the drawings.

As examples of various types of information, expressions such as "table," "list," and "queue" may be used, but the various types of information may be expressed in other data structures. For example, various types of information such as "XX table," "XX list," and "XX queue" may be expressed as "XX information." When explaining identification information, expressions such as "identification information," "identifier," "name," "ID," and "number" are used, but these are interchangeable.

When there are multiple components with the same or similar functions, they may be described using the same reference numerals with different subscripts. Also, when there is no need to distinguish between these multiple components, the subscripts may be omitted.

In the embodiments, the processing performed by executing a program may be described. Here, the computer executes the program using a processor (e.g., CPU (Central Processing Unit), GPU (Graphics Processing Unit)) and performs the processing defined in the program using storage resources (e.g., memory), interface devices (e.g., communication ports), etc. Therefore, the subject of the processing performed by executing the program may be the processor. Similarly, the subject of the processing performed by executing the program may be a controller, device, system, computer, or node having a processor. The subject of the processing performed by executing the program may be a calculation unit, and may include a dedicated circuit that performs specific processing. Here, the dedicated circuit is, for example, an FPGA (Field Programmable Gate Array), an ASIC (Application Specific Integrated Circuit), a CPLD (Complex Programmable Logic Device), etc.

The program may be installed on the computer from a program source. The program source may be, for example, a program distribution server or a computer-readable storage medium. When the program source is a program distribution server, the program distribution server may include a processor and a storage resource that stores the program to be distributed, and the processor of the program distribution server may distribute the program to be distributed to other computers. In addition, in the embodiments, two or more programs may be realized as one program, and one program may be realized as two or more programs.

In FIG. 1, 100 indicates a learning model generation system according to the first embodiment as a whole (hereinafter, referred to as "system 100").

The system 100 includes a learning model generation device (hereinafter referred to as "device") 101 and a client terminal (hereinafter referred to as "terminal") 102. The device 101 and the terminal 102 are connected so as to be able to communicate with each other. The device 101 stores one or more existing learning models that are being operated in a field system 103.

The device 101 is a device that generates a learning model. The device 101 is a notebook computer, a server device, a quantum computer, etc. The hardware configuration of the device 101 will be described later with reference to FIG. 2.

The device 101 includes a memory unit 110, a generation unit 120, and an output unit 130.

The storage unit 110 stores various types of information. For example, the storage unit 110 stores model information 111, a learning model 112, and a dataset 113. The model information 111 will be described later with reference to FIG. 3.

The learning model 112 is an existing trained model. The learning model 112 is composed of a network 114 and parameters 115.

In the system 100, the network 114 and the parameters 115 of the learning model 112 are classified into one of a first category, a second category, and a third category and stored in the storage unit 110. The first category includes a first layer model composed of a first network 114-1 and a first parameter 115-1. The second category includes a second layer model composed of a second network 114-2 and a second parameter 115-2. The third category includes a third layer model composed of a third network 114-3 and a third parameter 115-3.

In this way, the storage unit 110 stores the learning model 112 as a layer model classified by category. Note that the number of categories is not limited to three. The number of categories may be two, or four or more. The layer model will be described later with reference to FIG. 4.

Although the case where the learning model 112 is stored in the memory unit 110 has been described, the number of existing learning models stored in the memory unit 110 is not limited to one. The memory unit 110 may store one or more learning models different from the learning model 112. The following describes an example where multiple existing learning models are stored in the memory unit 110.

The generation unit 120 generates a new learning model based on the model information 111 stored in the memory unit 110 and the specification information 141 input by the input unit 140 of the terminal 102.

The output unit 130 outputs the learning model generated by the generation unit 120. The output learning model is installed, for example, in the edge device 150 of the on-site system 103.

The terminal 102 is a notebook computer, a tablet terminal, a smartphone, etc. The terminal 102 is equipped with an input unit 140.

The input unit 140 inputs specification information 141 that indicates the specifications of the learning model required by the user. The specification information 141 will be described later with reference to FIG. 5.

The systems 103 include a taxi system 103-1, a bicycle rental system 103-2, a delivery system 103-3, a wholesale system 103-4, a retail system 103-5, etc. Each system 103 is equipped with an edge device 150.

The edge device 150 is a device close to the site. The edge device 150 is, for example, a small device that can be connected to the Internet. The edge device 150 is a vehicle, a smartphone, a tablet terminal, a mobile phone, a camera, industrial equipment, a sensor, etc. The edge device 150 may be connected to various sensors for acquiring data on the site.

The edge device 150 has a learning model consisting of a network 151 and parameters 152. The edge device 150 performs object detection and the like using the learning model. The edge device 150 may re-learn the learning model based on data obtained from a sensor. The edge device 150 stores the data used to re-learn the learning model as a dataset 153. The dataset 153 may be stored in the edge device 150, or may be stored in a computer that can communicate with the edge device 150.

The edge device 150 transmits the data set 153 and the like to the apparatus 101 via the Internet. The learning model consisting of the network 151 and the parameters 152 is divided into layer models corresponding to categories by the operator of the system 100 and stored in the storage unit 110. At this time, the operator of the system 100 reflects the contents stored in the storage unit 110 in the model information 111.

Figure 1 shows an example in which a new learning model for a retail system 103-5 is generated based on existing learning models for a taxi system 103-1, a bicycle rental system 103-2, a delivery system 103-3, and a wholesale system 103-4.

In the above description of the system 100, the storage unit 110, the generation unit 120, and the output unit 130 are provided in the device 101. However, the system 100 is not limited to this configuration. For example, the storage unit 110 may be provided in a computer other than the device 101.

Figure 2 shows an example of the hardware configuration of device 101.

The device 101 includes a processor 210, a main memory device 220, an auxiliary memory device 230, an input device 240, an output device 250, and a communication device 260.

The processor 210 is a device that performs arithmetic processing. The processor 210 is, for example, a CPU (Central Processing Unit), an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), an AI chip, etc.

The main memory device 220 is a device that stores programs, data, etc. The main memory device 220 is, for example, a ROM (Read Only Memory), a RAM (Random Access Memory), etc. The ROM is, for example, a SRAM (Static Random Access Memory), a NVRAM (Non Volatile RAM), a Mask ROM (Mask Read Only Memory), a PROM (Programmable ROM), etc. The RAM is, for example, a DRAM (Dynamic Random Access Memory).

The auxiliary storage device 230 is a hard disk drive, a flash memory, an SSD (Solid State Drive), an optical storage device, etc. The optical storage device is a CD (Compact Disc), a DVD (Digital Versatile Disc), etc. The programs, data, etc. stored in the auxiliary storage device 230 are loaded into the main storage device 220 as needed.

The input device 240 is a user interface that accepts information from a user. The input device 240 is, for example, a keyboard, a mouse, a card reader, a touch panel, etc.

The output device 250 is a user interface that outputs various types of information (display output, audio output, print output, etc.). The output device 250 is, for example, a display device that visualizes various types of information, an audio output device (speaker), a printer, etc. The display device is an LCD (Liquid Crystal Display), a graphics card, etc.

The communication device 260 is a communication interface that communicates with other devices via a communication medium. The communication device 260 is, for example, a network interface card (NIC), a wireless communication module, a universal serial interface (USB) module, a serial communication module, etc. The communication device 260 can also function as an input device that receives information from other devices that are communicatively connected. The communication device 260 can also function as an output device that transmits information to other devices that are communicatively connected.

The functions of the device 101 (storage unit 110, generation unit 120, output unit 130, etc.) may be realized, for example, by the processor 210 reading a program stored in the auxiliary storage device 230 into the main storage device 220 and executing it (software). The functions of the device 101 may be realized by hardware such as a dedicated circuit. The functions of the device 101 may be realized by a combination of software and hardware. In addition, some of the functions of the device 101 may be realized by another computer capable of communicating with the device 101.

Note that the hardware of the terminal 102 is the same as or similar to that of the device 101, and therefore a description thereof will be omitted. The functions of the terminal 102 (such as the input unit 140) may be realized, for example, by a processor reading a program stored in an auxiliary storage device into a main storage device and executing it (software). The functions of the terminal 102 may be realized by hardware such as a dedicated circuit. The functions of the terminal 102 may be realized by a combination of software and hardware. In addition, some of the functions of the terminal 102 may be realized by another computer capable of communicating with the terminal 102.

Figure 3 shows an example of model information 111 (model table 300).

The model table 300 stores information indicating the layer models divided by the operator of the system 100 so that the learning models correspond to categories, and information indicating the datasets of the learning models.

More specifically, in the model table 300, information on a management number 310, a first layer model name 320, a second layer model name 330, a third layer model name 340, and a data set name 350 is associated.

The management number 310 is information that can identify a learning model registered in the auxiliary storage device 230 by the operator of the system 100. The first layer model name 320 is a name indicating a layer model of the learning model classified in the first category. The first layer model name 320 is linked to the layer model of the learning model classified in the first category. The second layer model name 330 is a name indicating a layer model of the learning model classified in the second category. The second layer model name 330 is linked to the layer model of the learning model classified in the second category. The third layer model name 340 is a name indicating a layer model of the learning model classified in the third category. The third layer model name 340 is linked to the layer model of the learning model classified in the third category. The dataset name 350 is a name indicating a dataset used in the learning of the learning model. The dataset name 350 is linked to the dataset used in the learning of the learning model.

Figure 4 shows an example of a learning model 112.

The learning model 112 consists of a first layer model 410, a second layer model 420, and a third layer model 430.

The first layer model 410 is a layer model for extracting the features of the subject (object) of inference in the learning model 112. The first layer model 410 is a layer model for extracting the features of an object, such as what kind of object is to be recognized, for example, whether the object is tilted or lying down, and may be referred to as a feature extraction unit. More specifically, the first layer model 410 consists of two layers 412-1 and 412-2 in which learned parameters 411-1 and 411-2 are set.

The second layer model 420 is a layer model for taking into account the influence of the site on the inference in the learning model 112. The second layer model 420 is a layer model for adding site dependency, such as the type of illuminance and the type of weather to be recognized, and may be referred to as the environment-dependent part. More specifically, the second layer model 420 consists of three layers 422-1, 422-2, and 422-3 in which learned parameters 421-1, 421-2, and 421-3 are set.

The on-site environment will be described using the illuminance and weather at the site as examples, but is not limited to these. The on-site environment may also be other physical indicators, such as the size of the site, how crowded the site is, etc.

The third layer model 430 is a layer model for performing inference (for performing the intended function) such as detecting objects and measuring distances, and may be referred to as a functional unit. More specifically, the third layer model 430 consists of three layers 432-1, 432-2, and 432-3 in which learned parameters 431-1, 431-2, and 431-3 are set.

The third layer model 430 will be described using an example in which it includes an output layer. However, this is not the only possible configuration. The third layer model 430 does not necessarily need to include an output layer. In this case, the output layer is provided separately.

Figure 5 shows an example of a screen (generation screen 500) for generating specification information 141.

The generation screen 500 is displayed on the terminal 102. The specification information 141 is generated based on the information entered on the generation screen 500.

The generation screen 500 includes a first selection list 510, a second selection list 520, a third selection list 530, a re-learning selection list 540, an input size 550, an output destination 560, and a model generation button 570.

In the first selection list 510, information on the first layer model name 320 is displayed in a selectable manner. In the second selection list 520, information on the second layer model name 330 is displayed in a selectable manner. In the third selection list 530, information on the third layer model name 340 is displayed in a selectable manner. In the re-learning selection list 540, whether or not to re-learn a new learning model is displayed in a selectable manner.

The input size 550 is an area for inputting the size of the data to be input into the new learning model. The output destination 560 is an area for specifying the location to output the new learning model. The model generation button 570 is a button for instructing the generation of a learning model.

The user selects the desired layer model (layer) from each of the first selection list 510, the second selection list 520, and the third selection list 530. The user also inputs the data size in input size 550 and the output location in output destination 560. Then, when the user presses model generation button 570, a learning model with the content specified on the generation screen 500 is generated.

Figure 6 shows an example of a series of processes (sequence) in the system 100.

In S601, the terminal 102 sends a request (generation request) to the device 101 to start generating a learning model.

In S602, the device 101 receives the generation request, acquires the model information 111 from the memory unit 110, and generates screen information for the generation screen 500 based on the model information 111.

In S603, the device 101 transmits screen information to the terminal 102.

In S604, the terminal 102 receives the screen information, displays the generation screen 500, and accepts input from the user. When the user presses the model generation button 570, the terminal 102 generates specification information 141 based on the information entered in the generation screen 500. The specification information 141 includes the information selected in the first selection list 510, the second selection list 520, the third selection list 530, and the relearning selection list 540, as well as the information entered in the input size 550 and the output destination 560.

In S605, the terminal 102 transmits the specification information 141 to the device 101.

In S606, the device 101 receives the specification information 141 and performs a process (generation process) to generate a learning model. In the generation process, a learning model is generated. The generation process will be described later with reference to FIG. 7.

Figure 7 shows an example of the generation process performed by device 101.

In S701, the generation unit 120 refers to the specification information 141 and obtains from the storage unit 110 the layer models with the layer model names selected in the first selection list 510, the second selection list 520, and the third selection list 530.

In S702, the generation unit 120 sets the size of the input layer according to the information of the input size 550 included in the specification information 141. Note that the input layer may be included in the first layer model. In this case, the input size 550 is not provided on the generation screen 500.

In S703, the generation unit 120 determines whether the sizes of all layer models match. More specifically, the generation unit 120 determines whether the sizes of the input layer and the first layer model match, whether the sizes of the first layer model and the second layer model match, and whether the sizes of the second layer model and the third layer model match. If the generation unit 120 determines that the sizes of all layer models match, it proceeds to S706, and if it determines that there are layer models whose sizes do not match, it proceeds to S704.

In S704, the generation unit 120 converts (adjusts) the size of data between layer models of different sizes.

The generation process is not limited to the above. For example, the size of the layer models may be adjusted before storing them in the device 101 or before the generation process. In this case, the processes of S703 and S704 are not necessary.

In S705, the generation unit 120 combines the layer models to generate a new learning model.

In S706, the generation unit 120 refers to the information selected in the relearning selection list 540 included in the specification information 141 and determines whether or not relearning is necessary. If the generation unit 120 determines that relearning is necessary, it proceeds to S707, and if it determines that relearning is not necessary, it ends the generation process.

In S707, the generation unit 120 re-trains the generated new learning model based on the dataset used for re-learning.

The annotations of the dataset used for re-learning may be changed by the operator of the system 100. Here, a case will be described where "3: Palette" is selected in the first selection list 510, "2: Indoors, fluorescent light" is selected in the second selection list 520, and "1: Object detection" is selected in the third selection list 530. In this case, the dataset with management number 310 "3" that contains data to be subjected to object detection (e.g., an image) is used as the dataset to be used for re-learning. In this case, since the dataset with management number 310 "3" has been assigned a label for angle detection, the operator of the system 100 re-assigns the label for object detection.

In S708, the output unit 130 outputs the learning model generated by the generation unit 120. More specifically, the output unit 130 transmits the learning model to the location specified by the output destination 560 based on the information on the output destination 560 included in the specification information 141.

The learning model generated by the device 101 may be tuned by the operator of the system 100.

Figure 8 is a diagram (image 800) showing the relationship between the accuracy of a learning model and the time it takes to generate the learning model.

In the image diagram 800, the vertical axis indicates the accuracy of the learning model, and the horizontal axis indicates the time required to generate the learning model. Graph 810 shows a graph when a learning model is generated by device 101. Graph 810 shows a graph when a learning model is generated by an existing method.

As shown in graph 810, the time it takes for device 101 to generate the learning model is a first time _x1 , and the accuracy of the learning model is a first accuracy _y1 .

As shown in graph 820, the time required to generate a learning model by the existing method is a second time _{x2 that} is longer than the first time x1 because it requires a time to prepare a dataset and a time to construct the learning model. In addition, according to the existing method, since trained parameters are not used, the accuracy of the learning model is a second accuracy _y2 that is lower than the first accuracy _y1 .

After the learning model is generated, it is trained in the field, and the accuracy of the learning model is improved.

According to this embodiment, it is possible to generate a learning model in a shorter time than existing methods. Furthermore, according to this embodiment, it is possible to generate a learning model with higher accuracy than existing methods. Furthermore, according to this embodiment, it is possible to generate a learning model with different functions from existing learning models.

(2) Second embodiment In this embodiment, the main difference from the first embodiment is that data (on-site data) from a site where a learning model generated by the device 101 is operated is used as a data set for the learning model. In this embodiment, the same components as those in the first embodiment are denoted by the same reference numerals and descriptions thereof are omitted.

Figure 9 shows an example of a system 900 according to this embodiment.

In the retail system 903-5, the learning model 960 generated by the apparatus 901 is provided in an edge device 950-5 and is operated. Although not shown in the figure, the retail system 903-5 is provided with various sensors.

In the retail system 903-5, the parameters 952-5 may be trained using a dataset (dataset 953-5) in which the operator of the retail system 903-5 labels on-site data acquired by a sensor. If training is not performed, the operator of the system 900 or the operator of the retail system 903-5 may generate the dataset 953-5 in which the on-site data is labeled.

In the system 900, the retail system 903-5 transmits a dataset 953-5 to the device 901 at a predetermined timing. The memory unit 910 of the device 901 stores the received dataset 953-5. At this time, the device 901 may update the dataset linked to the dataset name 350 corresponding to the management number 310 of the learning model 960 with the dataset 953-5, or may add and store the dataset 953-5.

The specified timing may be, for example, a time determined by the operator of the system 900. Also, for example, the specified timing may be a timing instructed by the operator of the system 900.

According to this embodiment, the device 901 can retrain the generated learning model using field data.

In addition, in the system 900, the operator of the system 900 may store the parameters 925-5 of the learning model 960 in the device 901.

(3) Third embodiment In this embodiment, the main difference from the first embodiment is the configuration for generating a learning model according to the accuracy designated by the user. In this embodiment, the same components as those in the first embodiment are denoted by the same reference numerals and the description thereof is omitted.

Figure 10 shows an example of a system 1000 according to this embodiment.

In the system 1000, the input unit 1040 of the terminal 1002 accepts input of model specification information 1041 including accuracy information indicating the accuracy specified by the user. The terminal 1002 transmits the model specification information 1041 to the device 1001. The model specification information 1041 will be described later with reference to FIG. 11.

The generation unit 1020 of the device 1001 generates a learning model according to the accuracy specified by the user, based on the specification information 1041 and the model information 1011 stored in the storage unit 1010. The model information 1011 will be described later with reference to FIG. 12.

Figure 11 shows an example of model specification information 1041 (model specification information 1100).

The model specification information 1100 includes information that corresponds the accuracy 1110, the first layer model name 1120, the second layer model name 1130, the third layer model name 1140, and whether or not re-learning is performed 1150.

Accuracy 1110 is information indicating the accuracy desired by the user for the learning model generated by the device 1001. First layer model name 1120 is information indicating the layer model of the first category used in the learning model. Second layer model name 1130 is information indicating the layer model of the second category used in the learning model. Third layer model name 1140 is information indicating the layer model of the third category used in the learning model. Relearning 1150 is information indicating whether or not relearning of the learning model is performed.

Although not shown in the figure, the model specification information 1041 is input via a specified screen. For example, the specified screen may be a screen in which an area for inputting accuracy information is provided on the generation screen 500.

Figure 12 shows an example of model information 1011 (model table 1200).

In model table 1200, information in model table 300 is further associated with information in accuracy 1210. Accuracy 1210 is information indicating the accuracy of the learning model of management number 310.

Figure 13 shows an example of the generation process performed by device 1001.

In S1301, the generation unit 1020 refers to the model specification information 1041 and determines whether the accuracy specified by the user (specified accuracy) is higher than the accuracy of the existing learning model (model accuracy). If the generation unit 1020 determines that the specified accuracy is higher than the model accuracy, it proceeds to S1302, and if it determines that the specified accuracy is not higher than the model accuracy, it proceeds to S1303.

In the system 1000, the generation unit 1020 uses the accuracy of the learning model that included the layer model of the first category (first layer model name) as the accuracy of the learning model.

Here, the model specification information 1100 and the model table 1200 are taken as examples for explanation. The model specification information 1100 includes the accuracy 1110 "55%", the first layer model name 1120 "3: Palette", the second layer model name 1130 "2: Indoor, fluorescent light", and the third layer model name 1140 "1: Object detection". In this case, the accuracy 1210 "60%" of the learning model with the model management number 310 "3" that included the first layer model name 1120 "3: Palette" is used as the accuracy of the existing learning model. Therefore, in S1301, the generation unit 1020 determines that the accuracy "55%" specified by the user is not higher than the accuracy "60%" of the existing learning model, and proceeds to S1303.

The reason for using the accuracy of the learning model that included a layer model of the first category as the accuracy of the learning model is that the layer model of the first category is a feature extraction unit and therefore has a large impact on the accuracy. However, the above configuration is not limited to this, since the layer model that has a large impact on the accuracy differs depending on how the layer models are classified. For example, the accuracy of the learning model may be configured to use the accuracy of a learning model that included a layer model of a category different from the layer model of the first category.

In S1302, the generation unit 1020 adds a layer. For example, the generation unit 1020 adds a layer that is the same as the last layer (e.g., the deepest layer) of the first layer model after the last layer of the first layer model. However, the addition of a layer is not limited to this process. For example, the generation unit 1020 may add two or more layers of the first layer model depending on the difference between the specified accuracy and the accuracy of the learning model. Also, for example, the generation unit 1020 may add a layer to a layer model different from the first layer model.

In S1303, the generation unit 1020 deletes a layer. For example, the generation unit 1020 deletes the end of the first layer model. However, the deletion of layers is not limited to this process. For example, the generation unit 1020 may delete two or more layers of the first layer model depending on the difference between the specified accuracy and the accuracy of the learning model. Also, for example, the generation unit 1020 may delete a layer from a layer model different from the first layer model.

The learning model generated by the device 101 may be tuned by the operator of the system 1000 to match the accuracy of the specification information 141.

According to this embodiment, it becomes possible to generate a learning model according to the accuracy specified by the user.

(4) Supplementary Notes The above-described embodiment includes, for example, the following contents.

In the above embodiment, the present invention has been described as being applied to a learning model generation system, but the present invention is not limited to this and can be widely applied to various other systems, devices, methods, and programs.

In addition, in the above-described embodiment, the configuration of each table is an example, and one table may be divided into two or more tables, or all or part of two or more tables may be one table.

The screens shown in the above-described embodiment are merely examples, and any screen design may be used as long as the information accepted is the same. For example, the screen may be designed like a spreadsheet, with the name of each input value in the title row, and each row being an input cell for an individual input value. The screen may also have an interface for accepting file input, and the contents of the screen may be updated based on the information in a specified file.

In addition, in the above description, information such as programs, tables, and files that realize each function can be stored in a storage device such as a memory, a hard disk, or an SSD (Solid State Drive), or in a recording medium such as an IC card, an SD card, or a DVD.

The above-described embodiment has the following characteristic configurations:

The learning model generation system (e.g., system 100, system 900, system 1000) includes a memory unit, a generation unit, and an output unit.

The storage unit is, for example, the storage unit 110, the device 101, the processor 210, or a circuit. The storage unit stores each of the multiple layers constituting an existing learning model (e.g., the learning model 112) that is a trained model (e.g., a neural network model, a deep neural network model). The multiple layers are a layer model consisting of a first network 114-1 and a first parameter 115-1, a layer model consisting of a second network 114-2 and a second parameter 115-2, a layer model consisting of a third network 114-3 and a third parameter 115-3, etc.

The number of the existing learning models may be one or more. If there is one existing learning model, the existing learning model consists of multiple layers. If there are multiple existing learning models, the multiple learning models may be provided at one site, or at multiple sites. Furthermore, if there are multiple existing learning models, the multiple learning models may be models that perform inference on the same subject, or models that perform inference on different subjects.

Each of the plurality of items stored in the storage unit may be classified into a predetermined category or granularity, or may not be classified into a predetermined category or granularity.

The generation unit 120 is, for example, the device 101, the processor 210, and a circuit. The generation unit acquires layers corresponding to each of the multiple layers from the storage unit based on specification information (for example, specification information 141) in which multiple layers are specified, and combines the acquired layers to generate a learning model (for example, a learning model consisting of network 151-5 and parameters 152-5).

The above specification information may be information generated via a screen such as the generation screen 500, or may be a file in which information has been entered in a pre-specified format.

The output unit is, for example, the output unit 130, the device 101, the processor 210, or a circuit. The output unit outputs the learning model generated by the generation unit. The output unit may output to an output destination specified by the user, or may output to an output destination specified in advance.

In the above configuration, layers of existing learning models are combined to generate a new learning model, so the learning model generation system can reduce the time it takes to generate a learning model.

The storage unit stores each of the multiple layers constituting each of the multiple existing learning models. The multiple existing learning models include a learning model consisting of network 151-1 and parameters 152-1, a learning model consisting of network 151-2 and parameters 152-2, a learning model consisting of network 151-3 and parameters 152-3, a learning model consisting of network 151-4 and parameters 152-4, etc.

The generation unit acquires layers corresponding to the layers from the storage unit based on specification information in which layers of different learning models among the multiple models are specified (e.g., see FIG. 5). The generation unit combines the acquired layers to generate a learning model.

In the above configuration, a new learning model is generated in which layers of different learning models are combined based on specification information in which layers of different models among multiple models are specified. With the above configuration, for example, a user can generate a learning model that is different from an existing learning model.

The storage unit stores a dataset in which a label has been added to the dataset used in the existing learning. The dataset is stored in the storage unit, for example, by an operator of the learning model generation system. The generation unit re-trains the generated learning model using the dataset.

Relearning may or may not be performed. Whether or not relearning is performed may or may not be specified by the user. When specified by the user, for example, information indicating whether or not relearning is performed is input to the device 101 directly or included in the specification information 141.

In the above configuration, the generated learning model is retrained based on a dataset in which an existing dataset is labeled. With the above configuration, for example, a user can obtain a learning model whose accuracy has been improved by retraining.

The storage unit stores multiple layers constituting the learning model generated by the generation unit, and a dataset in which on-site data acquired at the site where the learning model is operated is labeled. The dataset may be generated by an operator at the site, or may be generated by an operator of the learning model generation system.

In the above configuration, multiple layers of the generated learning model and a dataset in which data from the site where the learning model is being used are labeled are stored. With the above configuration, for example, a user can select each newly stored layer to generate a re-trained learning model, thereby improving the convenience of the learning model generation system.

The specification information includes information indicating accuracy of the learning model generated by the generation unit (see, for example, FIG. 11). The storage unit stores information indicating accuracy of the existing model (see, for example, FIG. 12).
The generation unit acquires layers corresponding to each of the layers from the storage unit based on the specification information (see, for example, S702). If the accuracy of the specification information is higher than the accuracy of the existing learning model, the generation unit adds one or more layers to the acquired layers (see, for example, S1302). If the accuracy of the specification information is not higher than the accuracy of the existing learning model, the generation unit deletes one or more layers from the acquired layers (see, for example, S1303).

In the above configuration, a learning model is generated according to the specified accuracy, so the user can obtain a learning model that is generated to approach the desired accuracy.

The storage unit classifies and stores the layers constituting the existing learning model into a first category, a second category, or a third category (see, for example, Figures 12 and 13). The first category is a category suitable for classifying layers for extracting object features. The second category is a category suitable for classifying layers for extracting environmental features. The third category is a category suitable for classifying layers for inference using the extracted features.

In the above configuration, the layers of the learning model are classified and stored into categories, so that, for example, a user can select a layer from among the organized categories, making it easy to generate the desired learning model.

Furthermore, the above-mentioned configurations may be modified, rearranged, combined, or omitted as appropriate without departing from the spirit and scope of the present invention.

It should be understood that an item listed in the format "at least one of A, B, and C" can mean (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C). Similarly, an item listed in the format "at least one of A, B, or C" can mean (A), (B), (C), (A and B), (A and C), (B and C), or (A, B, and C).

100...system, 110...memory unit, 120...generation unit, 130...output unit.

Claims (6)

A storage unit that stores each of a plurality of layers constituting an existing learning model that is a trained model;
a generation unit that acquires layers corresponding to each of the layers constituting the existing learning model from the storage unit based on specification information in which multiple layers are specified by a user's input operation received via a generation screen , and combines the acquired layers to generate a learning model;
an output unit that outputs the learning model generated by the generation unit;
Equipped with
The storage unit stores each of a plurality of layers constituting each of a plurality of existing learning models,
The generation unit acquires layers corresponding to different learning models from the storage unit based on specification information in which layers of the different learning models among the plurality of models are specified, and generates a learning model by combining the acquired layers.
Learning model generation system. The storage unit stores a dataset in which a label is added to the dataset used in the existing learning,
The generation unit re-trains the generated learning model using the dataset.
The learning model generation system according to claim 1 . The storage unit stores a plurality of layers constituting the learning model generated by the generation unit and a dataset in which site data acquired at the site where the learning model is operated is labeled.
The learning model generation system according to claim 1 . The specification information includes information indicating accuracy of the learning model generated by the generation unit,
The storage unit stores information indicating accuracy of the existing model,
The generation unit acquires layers corresponding to each of the plurality of layers from the storage unit based on the specification information, and adds one or more layers to the acquired layers when the accuracy of the specification information is higher than the accuracy of the existing learning model, and deletes one or more layers from the acquired layers when the accuracy of the specification information is not higher than the accuracy of the existing learning model.
The learning model generation system according to claim 1 . the storage unit classifies and stores layers constituting the existing learning model into any one of categories suitable for classifying layers for extracting object features, categories suitable for classifying layers for extracting environmental features, and categories suitable for classifying layers for inference using the extracted features.
The learning model generation system according to claim 1 . A storage unit stores each of a plurality of layers constituting an existing learning model that is a trained model;
A generation unit acquires layers corresponding to each of the layers constituting the existing learning model from the storage unit based on specification information in which a plurality of layers are specified by a user's input operation received via a generation screen , and generates a learning model by combining the acquired layers;
An output unit outputs the learning model generated by the generation unit;
Including,
The storage unit stores each of a plurality of layers constituting each of a plurality of existing learning models,
The generation unit acquires layers corresponding to different learning models from the storage unit based on specification information in which layers of the different learning models among the plurality of models are specified, and generates a learning model by combining the acquired layers.
How to generate a learning model.

Images