Machine learning device, machine learning method, and machine learning program

Abstract

To provide a machine learning technique capable of performing transition learning according to characteristics of a domain.SOLUTION: When a first model of neural network learned using teacher data of a first domain is subjected to transition learning by teacher data of a second domain, a domain adaptive data fullness degree determining part 52 determines a domain adaptive data fullness degree on the basis of the number of teacher data of a second domain. A learning layer determining part 54 determines a layer of a learning object of a second model obtained by copying the first model, based on the domain adaptive data fullness degree. A transition learning execution part 56 subjects the layer of the learning object of the second model to transition learning using the teacher data of the second domain.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to machine learning techniques. [Background technology]

[0002] Transfer learning is a technique for applying a model trained in one domain to another domain. In transfer learning, the original domain is the source domain, and the domain to which it is transferred is the target domain. It is called a neural network. It efficiently adapts a model trained in the source domain to the target domain. It is necessary to make this happen.

[0003] In Patent Document 1, an input image is received from a source domain, and a source image is extracted from the source domain. Generates a transformed image, which is a transformation of the input image to a target domain that is different from the domain To achieve this, the network is constructed to process inputs that include input images from the source domain. A domain transformation neural network is disclosed. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Special Publication No. 2020-502665 Summary of the Invention [Problem to be solved by the invention]

[0005] Transfer learning adapts a model trained in a source domain to a target domain. When applying the algorithm, transfer learning was performed regardless of the characteristics of the domain. This resulted in the generalization of inference accuracy. There were problems with this, such as a decrease in efficiency or an increase in the amount of processing required.

[0006] The present invention has been made in view of the above circumstances, and its object is to provide a method for detecting a domain according to its characteristics. The objective of this invention is to provide a machine learning technology that can perform transfer learning based on the learning model. [Means for solving the problem]

[0007] In order to solve the above problem, a machine learning device according to an embodiment of the present invention is provided. The first model of the neural network trained using the data is used as the teacher data for the second domain. When transferring data, the domain adaptation data is used based on the number of training data in the second domain. a domain adaptation data enrichment determining unit for determining a data enrichment level; Based on the degree, the learning layer of the second model, which is a replica of the first model, is determined. and a transfer learning step for transferring a learning target layer of the second model using training data of the second domain. and a transfer learning unit.

[0008] Another aspect of the present invention is a machine learning method, which includes: The first model of the neural network trained using the training data of the second domain is When transferring data to the second domain, the domain adaptation data is enriched based on the number of training data in the second domain. determining a degree of domain adaptation data richness; and replicating the first model based on the domain adaptation data richness. determining a layer to be learned of the second model; and and transferring the training target layer of the second model using

[0009] Any combination of the above components, and the expression of the present invention may be used as a method, an apparatus, a system, a recording medium, Conversions between the body, computer program, etc. are also valid aspects of the present invention. be. [Effects of the Invention]

[0010] According to the present invention, a machine learning technique capable of transfer learning according to the characteristics of a domain is provided. It can be provided. [Brief explanation of the drawings]

[0011] [Figure 1] FIG. 1 is a configuration diagram of a machine learning device and an inference device according to an embodiment. [Figure 2] FIG. 2 is a detailed configuration diagram of a transfer learning unit of the machine learning device of FIG. 1. [Figure 3] FIG. 2 is a diagram showing the structure of a neural network model used as a source model and a target model in the machine learning device of FIG. 1. [Figure 4] FIG. 10 is a diagram illustrating a layer to be trained as a target model according to the degree of domain adaptation data richness. [Figure 5] FIG. 10 is a diagram illustrating another example of a layer to be learned according to the domain adaptation data richness. [Figure 6] FIG. 10 is a diagram illustrating yet another example of layers to be learned according to the domain adaptation data richness. [Figure 7] 2 is a flowchart illustrating a machine learning procedure performed by the machine learning device of FIG. 1. DETAILED DESCRIPTION OF THE INVENTION

[0012] FIG. 1 is a configuration diagram of a machine learning device 100 and an inference device 200 according to an embodiment. The machine learning device 100 includes a source model storage unit 30, a target domain acquisition unit 40, and a transfer The inference device 200 includes a learning unit 50 and a target model storage unit 60. It includes a domain acquisition unit 70, an inference unit 80, and an inference result output unit 90.

[0013] Transfer learning is a machine learning technique that involves learning a first task with sufficient data. The model trained on the first task is then used to train a second task that is related to the first task but for which there is insufficient data. With sufficient data, the knowledge learned can be transferred to other tasks. This allows us to obtain highly accurate results even for the second task, which has little data.

[0014] In transfer learning, the input domain of the first task and the input domain of the second task are the same type. The case where the two domains are the same and only the probability distribution is different is called "domain adaptation."

[0015] The input domain of the first task that is the source of the transfer is called the "source domain," and the input domain of the second task that is the destination of the transfer is called the "source domain." The input domain of the task is called the "target domain." We call the first model the “source model” and the trained model for the second task the “target model.”

[0016] An example of domain adaptation is to use a computer that is easy to collect data from as the source domain. · Using graphics (CG) and images on the web, and using cameras as target domains It is preferable to use real images taken by a camera, but it is also necessary to distinguish between the source and target domains. The image format is not limited to these.

[0017] A large amount of CG images is used as the source domain to train the source model, and then the target By using the image captured by the camera as the target domain, the source model can be adapted to the domain. Here, the target model is generated by transfer learning. As an example, the objects are people, cars, bicycles, dogs, and motorcycles, and the task is classification.

[0018] The machine learning device 100 performs transfer learning from a trained source model and a target domain. It is a device that generates a more targeted model.

[0019] The source domain acquisition unit 10 acquires people, cars, bicycles, dogs, and motorcycles as source domains. A large number of people, cars, and bicycles can be trained to classify the source model with high accuracy. , dog, and motorcycle CG images are obtained.

[0020] The learning unit 20 performs machine learning of a neural network model using the source domain. A source model is generated and stored in a source model storage unit 30. The source model can classify the source domain with high accuracy.

[0021] The source model stored in the source model storage unit 30 is used as a transfer source model for transfer learning. The source model is a neural network model. do.

[0022] The target domain acquisition unit 40 acquires, as the target domain, the person photographed by the camera, Get images of cars, bicycles, dogs, motorcycles. The target domain is generally the source domain. There is less data than in.

[0023] The transfer learning unit 50 uses the target domain acquired by the target domain acquisition unit 40. The source model stored in the source model storage unit 30 is transferred to the target model by using the A target model is generated and stored in the target model storage unit 60. do.

[0024] The target model stored in the target model storage unit 60 is generated by transfer learning. The target model is a neural network model. The target model is a part of the source model that was once replicated and then transferred to the target domain. The training was carried out again using the

[0025] The inference device 200 performs inference using the target model generated by the machine learning device 100. The inference device 200 is a device that performs inference and classifies images. The acquired image is inferred and the inference results are output.

[0026] The target domain acquisition unit 70 acquires the target domain of the inference target, and The inference unit 80 supplies the target model stored in the target model storage unit 60 to the The target domain is inferred based on the rule, and the inference result is supplied to the inference result output unit 90. The inference result output unit 90 outputs the classification that is the inference result.

[0027] FIG. 2 is a detailed configuration diagram of the transfer learning unit 50 of the machine learning device 100. The domain adaptation data richness determination unit 52, the learning layer determination unit 54, and the transfer learning execution unit 5 Includes 6.

[0028] The domain adaptation data enrichment determination unit 52 determines the enrichment level of the target domain based on the number of training data of the target domain. Determine domain-adaptive data richness using the

[0029] More specifically, domain adaptation data richness is the number of classes per target domain. The ratio of the number of training data for the class with the smallest number of training data to the specified number of training data TDNUM Here, the predetermined number of training data is the number of training data that is considered to be sufficient for accurate learning. The number of training data depends on the neural network model. The neural network model is VGG16, so the number is set to 3000. Using the training data from the database, we measure the accuracy of learning according to the amount of training data, and aim to achieve accuracy that exceeds the desired level. The number of training data sets to be used may be used as the predetermined number of training data sets. It is preferable to calculate it by the following formula: MIN(n,m) selects the minimum value of n or m In this case, the domain adaptation data richness is a value between 0 and 1. Domain adaptation data richness = MIN(1, (the number of training data in the target domain is the smallest (Number of teacher data in each class) / TDNUM

[0030] The learning layer determination unit 54 determines the learning layer of the neural network based on the domain adaptation data richness. The layer to be learned is determined in the target model, which is a network model.

[0031] Specifically, the learning layer determination unit 54 determines whether the target model is more suitable for the domain adaptation data richness. The number of layers to be trained is increased, and the lower the domain adaptation data richness, the lower the target model. The number of layers to be trained in the model can be reduced.

[0032] The learning layer determination unit 54 determines whether the higher the domain adaptation data richness, the higher the layer (the layer closer to the input) is. Many layers are included in the learning target, from the top layer to the bottom layer (layers close to the output), and the domain adaptation data richness is high. The lower the learning rate, the fewer layers on the lower layer side (layers closer to the output) may be used as learning targets. The layer determination unit 54 determines whether a layer closer to the input layer is the learning target layer as the domain adaptation data richness increases. include.

[0033] The learning layer determination unit 54 determines whether the target model is a target model when the domain adaptation data richness is equal to or less than a predetermined value. Alternatively, only the fully connected layer of the model may be used as the learning target layer.

[0034] In this way, when the degree of domain adaptation data richness is low, that is, when the number of training data is small, By training only the lower layers of the target model, the details of the upper layers of the source model can be By adapting the source model to the target domain while maintaining the generalization performance of feature extraction, This allows the accuracy of the target model to be maintained at a high level after transfer learning. When the degree of main adaptation data is high, that is, when the number of training data is large, the target model By training more layers of the target model, the accuracy of the target model after transfer learning can be improved. It is possible.

[0035] The transfer learning execution unit 56 uses the image of the target domain as training data to perform the learning layer The layer to be the learning target of the target model determined by the determination unit 54 is subjected to transfer learning. The layers other than the one you want to learn are not the target of learning, and the layers of the copied source model are newly learned. It is used as is without modification.

[0036] FIG. 3 shows the data used as the source model and the target model in the machine learning device 100. FIG. 1 is a diagram showing the structure of a neural network model.

[0037] In this embodiment, the source model and the target model are neural network models. VGG16 has 13 convolutional layers (CONV) and 1 fully connected layer (D The learning layer consists of three layers of convolutional layers and five layers of pooling layers. This is a fully connected layer. The pooling layer subsamples the feature map output from the convolutional layer. The source and target models are not limited to VGG16, and the number of layers is also It's not limited to form.

[0038] Figure 4 explains the layers to be trained for the target model according to the degree of domain adaptation data richness. This is an explanatory diagram.

[0039] If the domain adaptation data richness is 1.00, all layers are used for learning. If the adaptive data completeness is less than 1.00 and greater than or equal to 0.95, CONV1-1 is excluded. The layer is the learning target. CONV1-1 uses a direct copy of the source model's layer. If the domain adaptation data richness is less than 0.95 and greater than or equal to 0.90, CONV1-1 The layers excluding CONV1-1 and CONV1-2 are the target of training. The layer of the model is replicated and used as is. If it is less than .10 and greater than 0.00, all layers are not used for learning. The layers are simply copied from the source model.

[0040] In this way, the greater the domain adaptation data richness, the more convolutions close to the input layer. The layer is the learning target, and the smaller the domain adaptation data richness, the closer it is to the output layer. The convolutional layer is the learning target.

[0041] The relationship between the domain adaptation data richness and the layer to be learned is not limited to this. The higher the degree of adaptation data richness, the more layers are used from the upper layer (layers closest to the input) to the lower layer (layers closest to the output). The lower the domain adaptation data richness, the lower the layer (the layer closest to the output) is. The layer with the fewest number of layers should be the learning target. Note that Dense indicates three fully connected layers. Although three fully connected layers are controlled together, they can also be controlled individually, as with the convolutional layers. ,Domain adaptation data richness is the minimum number of training data per class in the target domain. The ratio of the number of teacher data of the class to the number of teacher data TDNUM is The main adaptive data richness is the minimum number of training data per class in the target domain. It may also be the number of training data in each class.

[0042] Figure 5 is a diagram explaining another example of a layer to be learned depending on the degree of domain adaptation data richness. be.

[0043] In the example shown in Figure 5, multiple convolution layers are bundled and controlled in units of pooling layers. For example, the two convolutional layers CONV1-1 and CONV1-2 are bundled together to form the domain If the adaptive data richness is less than 1.00 and more than 0.95, it is a target for learning. The two convolutional layers, CONV2-1 and CONV2-2, are combined into one for domain adaptation data filling. If the actual degree is less than 1.00 and more than 0.85, it is a learning target. Adjacent convolutional layers are combined into one and used as the learning target. The feature extraction of the source model can be maintained at each resolution of the map.

[0044] FIG. 6 illustrates another example of a layer to be learned depending on the domain adaptation data richness. This is a diagram.

[0045] In the example shown in Figure 6, layers close to the input layer are not always the target of learning. Four convolutional layers: CONV1-1, CONV1-2, CONV2-1, and CONV2-2 For other convolutional layers, the domain adaptation data richness is large. The larger the value, the fewer convolutional layers that are closer to the output layer will be trained.

[0046] This allows us to directly utilize the detailed edge-level feature extraction of the source domain. Accuracy is improved by adapting to the target domain from abstracted features. This allows the amount of calculation to be reduced.

[0047] FIG. 7 is a flowchart illustrating the machine learning procedure performed by the machine learning device 100 of FIG. do.

[0048] The domain adaptation data richness determination unit 52 of the transfer learning unit 50 of the machine learning device 100 Determine the degree of domain adaptation data richness based on the number of training data in the get domain (S1 0).

[0049] The learning layer determination unit 54 replicates the source model based on the domain adaptation data richness. The layer to be trained for the target model is determined (S20).

[0050] The transfer learning execution unit 56 uses the image of the target domain as training data to The layer to be trained of the model is subjected to transfer learning (S30).

[0051] The various processes of the machine learning device 100 and the inference device 200 described above are performed by a CPU and memory. It can be realized as a device using hardware such as a ROM ( Firmware stored in read-only memory (ROM) or flash memory, etc., It can also be realized by software such as a computer. The program, software program, etc. are recorded on a computer-readable recording medium. It can be provided to a server or transmitted over a wired or wireless network. It is also possible to transmit and receive the data as a terrestrial or satellite digital broadcast.

[0052] As described above, according to the machine learning device 100 of this embodiment, The target model is subjected to transfer learning according to the degree of domain adaptation data richness based on the number of training data. Modifying the layers of our neural network allows us to process data according to the characteristics of the domain. It is possible to generate a target model that is highly efficient, has high inference accuracy, and generalization performance.

[0053] The present invention has been described above based on the embodiments. The embodiments are merely examples, and each structure thereof is not intended to be limiting. The fact that various variations are possible in the combination of components and each treatment process, and that such variations It will be understood by those skilled in the art that such modifications are also within the scope of the present invention. [Explanation of symbols]

[0054] 10 source domain acquisition unit, 20 learning unit, 30 source model storage unit, 40 Target domain acquisition part, 50 Transfer learning part, 52 Domain adaptation data richness Decision unit, 54 learning layer decision unit, 56 transfer learning execution unit, 60 target model storage memory unit, 70 target domain acquisition unit, 80 inference unit, 90 inference result output unit, 100 Machine learning devices, 200 Inference devices.

Claims

1. The first model of the neural network trained using the training data of the first domain is When transferring learning using the training data of the second domain, the number of training data in the second domain is determined based on the number of training data. a domain adaptation data richness determining unit for determining a domain adaptation data richness based on the domain adaptation data richness; training a second model that replicates the first model based on the domain adaptation data richness; a learning layer determination unit that determines a target layer; Transfer learning a learning target layer of the second model using training data of the second domain. a transfer learning unit; The learning layer determination unit determines whether the target domain adaptation data is sufficient or not when the domain adaptation data is sufficient or not. A machine learning device characterized in that not all layers of a network model are used as learning targets.

2. The learning layer determination unit is configured to determine the number of layers to be learned as the domain adaptation data richness increases. and the smaller the domain adaptation data richness, the smaller the number of layers to be learned. The machine learning device according to claim 1 .

3. The learning layer determination unit is configured to determine a learning layer closer to an input layer as the domain adaptation data richness increases. The machine learning device according to claim 1 or 2, characterized in that the layer to be learned is included.

4. The learning layer determination unit determines whether the domain adaptation data richness is equal to or less than a predetermined value.

2. The machine learning device according to claim 1, wherein only the first layer is used as a learning target layer.

5. The first model of the neural network trained using the training data of the first domain is When transferring learning using the training data of the second domain, the number of training data in the second domain is determined based on the number of training data. a domain adaptation data richness determination step of determining a domain adaptation data richness based on the domain adaptation data richness; training a second model that replicates the first model based on the domain adaptation data richness; a learning layer determination step for determining a target layer; Transfer learning a learning target layer of the second model using training data of the second domain. a transfer learning step; When the domain adaptation data richness is equal to or less than a predetermined value, the learning layer determination step A machine learning method characterized in that not all layers of the target model are used as learning targets.

6. The first model of the neural network trained using the training data of the first domain is When transferring learning using the training data of the second domain, the number of training data in the second domain is determined based on the number of training data. a domain adaptation data richness determination step of determining a domain adaptation data richness based on the domain adaptation data richness; training a second model that replicates the first model based on the domain adaptation data richness; a learning layer determination step for determining a target layer; Transfer learning a learning target layer of the second model using training data of the second domain. and a transfer learning step, When the domain adaptation data richness is equal to or less than a predetermined value, the learning layer determination step A machine learning program characterized by not targeting all layers of the target model as learning targets.

Images