Machine learning devices, inference devices, machine learning methods, and machine learning programs

Abstract

To provide a machine learning technique capable of optimizing the number of tasks that can be learned additionally and the accuracy of the added tasks for a target task.SOLUTION: According to the depth of a neural network model layer, an initialization rate determination unit 22 determines a first initialization rate for initializing the weight of a neural network model of a first task. A machine learning execution unit 24 machine-learns the first task to generate a learned neural network model of the first task. An initialization unit 26 initializes the weight of the learned neural network model of the first task based on the first initialization rate to generate an initialized learned neural network model of the first task for use in a second task.SELECTED DRAWING: Figure 2

Description

Technical Field

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

Background Art

[0002] Humans can learn new knowledge through long-term experience and can maintain the old knowledge without forgetting it. On the other hand, the knowledge of a convolutional neural network (CNN) depends on the dataset used for learning, and in order to adapt to changes in the data distribution, it is necessary to relearn the parameters of the CNN for the entire dataset. In CNN, as learning progresses for a new task, the estimation accuracy for the old task decreases. Thus, in CNN, when continuous learning is performed, catastrophic forgetting, in which the learning results of the old tasks are forgotten during the learning of the new task, cannot be avoided.

[0003] As a method for avoiding catastrophic forgetting, incremental learning or continual learning has been proposed. Incremental learning or continual learning is a learning method in which, when a new task or new data occurs, instead of learning the model from scratch, the currently learned model is improved and learned. One method of incremental learning is PackNet (Non-Patent Document 1).

[0004] In the incremental learning by PackNet, the weights used are changed in the order of the added tasks,

Prior Art Documents

Non-Patent Documents

[0004] [Non-Patent Document 1] Mallya, Arun, and Svetlana Lazebnik. “Packnet: Adding multiple tasks to a single network by iterative pruning.” Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2018. [Overview of the project] [Problems that the invention aims to solve]

[0005] In PackNet, the number of additional tasks that can be trained and the accuracy of those additional tasks are the target tasks. There was a problem in that it did not improve in that regard.

[0006] This invention was made in view of these circumstances, and its purpose is to enable additional learning A machine learning technique that can optimize the number of tasks and the accuracy of added tasks for the target task. The purpose is to provide a service. [Means for solving the problem]

[0007] To solve the above problems, a machine learning device according to one aspect of the present invention is a neural network The weights of the neural network model for the first task are initially determined according to the depth of the layers in the neural network model. An initialization rate determination unit that determines a first initialization rate, and a machine learning unit that performs the first task A machine learning execution unit that generates a pre-trained neural network model of the first type, and the first type The weights of the trained neural network model of the Suku are initialized based on the first initialization rate. The initialized and trained neural network from the first task is then converted for use in the second task. It includes an initialization unit that generates a work model.

[0008] Another aspect of the present invention is an inference device. This device selects one task from a group of tasks. A task input unit for selecting tasks, and a neural network model that has already learned the aforementioned multiple tasks. Dell's weights are set to a new weight where all weights except those used in the selected task are set to 0. An inference model generation unit that generates a neural network model, and the selected task before It includes an inference unit that performs inference based on a new neural network model.

[0009] Yet another aspect of the present invention is a machine learning method. This method uses a neural network Initialize the weights of the neural network model for the first task according to the depth of the model layers. An initialization rate determination step to determine a first initialization rate, and the first task to be performed by machine learning A machine learning execution step that generates a pre-trained neural network model for one task, The weights of the trained neural network model for the first task are based on the first initialization rate. Then, initialize the initialized and trained neural for use in the second task. This includes an initialization step to generate a network model.

[0010] Furthermore, any combination of the above components, or any expression of the present invention, may be used to describe a method, apparatus, system, or recording medium. The conversion between the human body, computer programs, etc., is also valid as an embodiment of the present invention. be. [Effects of the Invention]

[0011] According to the present invention, the number of additional tasks that can be learned and the accuracy of the additional tasks are compared to the target task. We can provide machine learning techniques that can optimize performance. [Brief explanation of the drawing]

[0012] [Figure 1] It is a configuration diagram of a machine learning device and an inference device according to an embodiment. [Figure 2] It is a detailed configuration diagram of the continuous learning unit of the machine learning device in FIG. 1. [Figure 3] It is a flowchart for explaining the operation of the continuous learning unit in FIG. 2 for Task 1. [Figure 4] It is a diagram showing the structure of a neural network model used in the machine learning execution unit in FIG. 2. [Figure 5] FIGS. 5(a) and 5(b) are diagrams for explaining a predetermined value of the initialization rate of the neural network model for Task 1. [Figure 6] It is a diagram for explaining the number of input / output channels of each layer of the neural network model, the number of weights between the input / output channels, the total number of weights of each layer, and the number of parameters. [Figure 7] It is a flowchart for explaining the operation of the continuous learning unit in FIG. 2 for Task 2. [Figure 8] It is a diagram for explaining a predetermined value of the initialization rate of the neural network model for Task 3. [Figure 9] It is a flowchart for explaining the operation of the inference device in FIG. 1 for Task N.

Mode for Carrying Out the Invention

[0013] 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 task input unit 10, a continuous learning unit 20, and a storage unit 30. The inference device 200 includes a task input unit 40, a task determination unit 50, an inference model generation unit 60, an inference unit 70, and an inference result output unit 80.

[0014] In continuous learning, it is required to learn a new task without catastrophic forgetting. In this embodiment The machine learning device 100, in particular, adds new tasks to the trained model during continuous learning. The purpose is to enable learning.

[0015] The machine learning device 100 learns the target model and effective parameters from multiple tasks through continuous learning. This is a device that generates meter information. For simplicity, the following tasks will be described here. I will explain assuming there are three, but the number and types of tasks are arbitrary.

[0016] Task 1 involves image recognition using the ImageNet dataset, which is the first dataset. This is a recognition task. Task 2 is a second dataset, the Places365 dataset. This is an image recognition task using the third dataset, CUBS B. This is an image recognition task using the IRDs dataset. The task input to the inference device 200 is... KN is one of Task 1 to Task 3, which are tasks for which the target model has been trained. This is a task. Here, a different dataset was assigned to each task, but each task This is not limited to cases where the tasks are different. It may also be divided into classes. For example, 10 different classes in the ImageNet dataset You may assign them to Task 1, Task 2, and Task 3 respectively. Also, the images for each task are shown in the diagram. The image may also be an image input to the task input unit 10 from an image acquisition unit such as a camera, which is not shown. For example, Task 1 uses an existing image dataset, and Tasks 2 and beyond use a camera, etc., which are not shown in the diagram. These may also be used as a dataset of images input to the task input unit 10.

[0017] The task input unit 10 can process multiple tasks (in this case, Task 1, Task 2, and Task 3) in succession. These will be supplied to the learning unit 20 in stages.

[0018] The continuous learning unit 20 sequentially uses multiple tasks (in this case, Task 1, Task 2, and Task 3). The neural network model is continuously trained to obtain information on the target model and effective parameters. Generate information.

[0019] The target model is a pre-trained neural network model generated by the continuous learning unit 20. It is Dell. The target model will eventually be able to perform multiple tasks through continuous learning (here, the This will be a pre-trained neural network for Task 1, Task 2, and Task 3. Effective parameters The data is applied to the trained neural network model generated by the continuous learning unit 20. , parameters such as the weights of the trained neural network model to be enabled for each task This is identifying information. Details of the valid parameter information will be described later.

[0020] The memory unit 30 stores the target model and effective parameter information.

[0021] The inference device 200 uses the target model and effective parameters generated by the machine learning device 100. This device uses data to generate inference results for multiple tasks.

[0022] The task input unit 40 supplies task N to the inference unit 70. The task determination unit 50 determines the inference unit The task N supplied to 70 is determined to be one of the pre-trained tasks (here, task 1, task 2). The system determines which task (either Task 2 or Task 3) is being performed and supplies the determination result to the inference model generation unit 60. In this embodiment, the user specifies which of Task 1 to Task 3 they are performing. However, it could be automated using some method.

[0023] The inference model generation unit 60 generates target models acquired from the memory unit 30 of the machine learning device 100. Dell stores effective parameter information and, based on the target model and effective parameter information... An inference model is generated and supplied to the inference unit 70.

[0024] The inference unit 70 performs tasks based on the inference model generated by the inference model generation unit 60. N is inferred and the inference result is supplied to the inference result output unit 80. The inference result output unit 90 outputs the inference result Output the result.

[0025] Figure 2 is a detailed configuration diagram of the continuous learning unit 20 of the machine learning device 100. Continuous Learning Unit 20 This includes a task similarity derivation unit 21, an initialization rate determination unit 22, a machine learning execution unit 24, and an initialization unit 26. , and also includes a fine-tuning section 28.

[0026] Figure 3 is a flowchart illustrating the operation of the continuous learning unit 20 for Task 1. Refer to Figure 2 and Figure 3 to describe the configuration and operation of the continuous learning unit 20 for Task 1.

[0027] The task similarity derivation unit 21 calculates that for task 1, since it is the first task, the task similarity is... It will not be calculated.

[0028] The initialization rate determination unit 22 determines the neural network according to the depth of the neural network layers. The initialization rate of the work model is determined to a predetermined value (S10). Task 1 uses a neural network. All weights in the work model are subject to initialization. The predetermined values ​​will be described later.

[0029] The machine learning execution unit 24 performs machine learning on a neural network model for Task 1. Then, a pre-trained neural network model is generated (S20).

[0030] Figure 4 shows the structure of the neural network model used in the machine learning execution unit 24. This is a diagram.

[0031] In this embodiment, the neural network model is a deep neural network. Let's call it VGG16. VGG16 has 13 convolutional layers (CONV) and fully connected layers (D It consists of 3 convolutional layers and 5 pooling layers. The layers to be trained are the convolutional layers and It is a fully connected layer. The pooling layer subsamples the feature map, which is the output of the convolutional layer. These are layers. Layers closer to the input are called shallow layers, and layers closer to the output are called deep layers. The model is not limited to VGG16, nor is the number of layers limited to this embodiment.

[0032] Figures 5(a) and 5(b) show the initial results of the neural network model for Task 1. This diagram illustrates the predetermined value of the pre-processing rate.

[0033] The initialization rate is set to a predetermined value for each layer of the neural network. In Figure 5(a), C ONV1-1, CONV1-2, CONV2-1, CONV2-2, CONV3-1, C For ONV3-2 and CONV3-3, the initialization rate is set to 0%, and for CONV4-1, CONV4-2, CONV4-3, CONV5-1, CONV5-2, CONV5-3, For Dense6, Dense7, and Dense8, the initialization rate is set to 50%.

[0034] In Figure 5(b), the initialization rate for CONV1-1 and CONV1-2 is set to 10%. Furthermore, the initialization rate for CONV2-1 and CONV2-2 is set to 20%, CON For V3-1, CONV3-2, and CONV3-3, the initialization rate is set to 30%, C For ONV4-1, CONV4-2, and CONV4-3, the initialization rate is set to 40%. , CONV5-1, CONV5-2, CONV5-3, Dense6, Dense7, D For ense8, the initialization rate is set to 50%.

[0035] Regarding the hierarchy of neural network models, the initialization rate is higher in deeper layers than in shallower layers. It is preferable to set it so that it becomes less available. The higher the initialization rate, the more available it will be for Task 2 and beyond. The weights increase. Below, we will discuss the initialization rate of the neural network model for Task 1. The predetermined value will be explained using the example shown in Figure 5(a).

[0036] Refer again to Figures 2 and 3. The initialization unit 26 uses the trained neural network model Initialize Dell's weights based on the initialization rate of each layer (S30). Here, initialization means This involves setting the weights of the neural network to 0 (zero). For each layer of the twerk model, the weights in each layer are allocated in order from the weights closest to 0 to the initialization rate. Initialize the weights of the combination to 0.

[0037] Weights that were not initialized will be used in Task 1, and weights that were initialized will be used in Task 1. The weight will be used from Task 2 onwards.

[0038] The effective parameter information for Task 1 is the weights used in Task 1, i.e., the learning of Task 1. This is information that identifies weights that have not been initialized after learning. Initialization unit 26 is effective for task 1. The parameter information is stored in the storage unit 30.

[0039] The effective parameter information is assigned to each weight of the neural network model, with 1 bit assigned to each. This is binary information assigned to each unit. The initialization unit 26 is a neural network model. For all the weights of the element, if the weight is 0, the sign is "0", and if the weight is not 0, the sign is " The value "1" may be assigned and stored in the storage unit 30 as a code sequence.

[0040] Figure 6 shows the number of input / output channels and the intervals between input / output channels in each layer of the neural network model. This diagram explains the number of weights, the total number of weights in each layer, and the number of parameters.

[0041] If the initialization rate is 50%, for example, CONV4-1, then there are 1,179,648 units. Initialize 589,824 weights, which is 50% of the total weights.

[0042] Refer again to Figures 2 and 3. The fine-tuning unit 28 changes the initialized weights. To avoid making changes, the trained neural network model for Task 1 is refined. Tune and generate the target model (S40). Target for fine tuning. The weights used are the uninitialized weights used in Task 1.

[0043] Next, we will explain the operation of the continuous learning unit 20 for Task 2.

[0044] Figure 7 is a flowchart illustrating the operation of the continuous learning unit 20 for Task 2. Refer to Figure 2 and Figure 7 to describe the configuration and operation of the continuous learning unit 20 for Task 2.

[0045] The task similarity derivation unit 21 uses Task 1, which is a trained task, and the target task. The distance of the probability density functions of the data distributions for Task 2 is derived as the task similarity (S50). Here, the Jensen-Shannon diverge is used as the distance between two probability density functions. Use JS divergence. JS divergence takes values ​​from 0 to 1. The smaller the JS divergence, the closer the distance between the two probability density functions is, and the JS divergence The larger the gens, the greater the distance between the two probability density functions. Therefore, JS divergence The settings are configured so that the smaller the divergence, the greater the task similarity, and the JS divergence is large. Set it so that the higher the value, the lower the task similarity.

[0046] Here, JS divergence was used to derive task similarity, but The K. Leibler divergence (KLD) and other methods evaluate the distance between two probability density functions. Any scale may be used as long as it is a suitable scale.

[0047] The initialization rate determination unit 22 determines the number of layers in the neural network according to the depth of the neural network layers and the task similarity. The initialization rate of the get model is determined to a predetermined value (S60). The predetermined value will be described later.

[0048] The weights to which the initialization rate applies are those that have not been assigned to any task. Weights assigned to any task are not included in the initialization process.

[0049] Figure 8 illustrates a predetermined value for the initialization rate of the neural network model for Task 2. This is a diagram.

[0050] The initialization rate is determined based on the layer depth and task similarity of the neural network as follows: Set to a predetermined value.

[0051] CONV1-1 to CONV3-3 are not initialized in Task 1, which is a pre-trained task. Since there are no weights to initialize, the initialization rate is 0.

[0052] When task similarity is high, i.e., when JS divergence (JSD) is small, the hierarchy Set the initialization rate higher for shallower levels and lower for deeper levels.

[0053] When task similarity is high, i.e., JSD is small, task similarity is low, i.e., J The initialization rate is set higher compared to when the SD card is larger.

[0054] When the task similarity is large, i.e., the JSD is small, CONV4-X(X=1,2,3 The weights of ) will not be updated.

[0055] More specifically, as an example, as shown in Figure 8, when JSD < 0.1, CONV For 4-1, CONV4-2, and CONV4-3, the initialization rate is set to 100%, C For ONV5-1, CONV5-2, and CONV5-3, the initialization rate is set to 95%. For Dense6, Dense7, and Dense8, the initialization rate is set to 80%. .

[0056] If 0.1 ≤ JSD < 0.5, then CONV4-1, CONV4-2, CONV4-3, For CONV5-1, CONV5-2, and CONV5-3, the initialization rate is set to 90%. Furthermore, for Dense6, Dense7, and Dense8, the initialization rate is set to 75%. ru.

[0057] If 0.5 ≤ JSD < 0.9, then CONV4-1, CONV4-2, CONV4-3, CONV5-1, CONV5-2, CONV5-3, Dense6, Dense7, De For nse8, the initialization rate is set to 75%.

[0058] If 0.9 ≤ JSD, then CONV4-1, CONV4-2, CONV4-3, CONV 5-1, CONV5-2, CONV5-3, Dense6, Dense7, Dense8 The initialization rate for this is set to 50%.

[0059] Therefore, in the case of tasks with high similarity, higher-level features are similar to those of the trained task. Therefore, layers that learn higher-level features have a higher initialization rate, and subsequent additions are processed. The weights initialized for the purpose can be retained.

[0060] If Task 1 and Task 2 are highly similar, the weights assigned to Task 1 will be changed to match those of Task 2. Because the probability of being able to share for inference increases, new to assign to Task 2 Also, the number of weights to initialize can be reduced. Conversely, if the similarity between Task 1 and Task 2 is low If not, the weights assigned to Task 1 can be shared for inference in Task 2. Because the probability decreases, it is necessary to increase the number of weights to be newly initialized in order to assign them to Task 2. There is a need.

[0061] Refer again to Figures 2 and 7. The machine learning execution unit 24 is the target task. Using method 2, transfer learning is performed on the target model without changing the weights of the pre-trained task. Then a trained neural network model is generated (S70). Here the trained task The weights of 'K' are the weights used in Task 1. They are assigned to the trained tasks before and after transfer learning. The weights used remain unchanged. Note that we will not modify the weights of the pre-trained task here. Training a target model involves transferring the weights of a trained task to another task. Although this was called transfer learning, it could also simply be called learning.

[0062] The initialization unit 26 initializes the weights of the trained neural network model based on the initialization rate. Initialize and generate the first candidate for the target model (S80).

[0063] Weights that are not initialized, including the weights of the pre-trained task, are allocated as weights used in Task 2. It can be identified.

[0064] The valid parameter information for Task 2 is the weights used in Task 2, and is not initialized. This is information that identifies the weight. The initialization unit 26 stores the effective parameter information of task 2 in the storage unit. Remember it as 30.

[0065] The fine-tuning unit 28 does not change the weights of the trained task or the initialized weights. In this way, the first candidate for the target model for Task 2 is fine-tuned. Generate a second candidate for the get model (S90).

[0066] The fine-tuning unit 28 selects the first candidate target model and the second target model. Of the two candidates, the one with higher accuracy will be selected as the final target model (S1 00). Basically, the second candidate for the target model is selected as the final target model. However, in order to improve the generalization performance of the target model, the weights of the target model are Using evaluation data different from the training data used for learning, the target model at the end of learning The inference accuracy of the first and second candidate for the rule is evaluated, and the candidate with higher accuracy is selected as the final candidate. It is preferable to decide on this as the get model.

[0067] Thus, based on the depth of the neural network model layers and the similarity between tasks Set the initialization rate of the pre-trained neural network model and train it for a new task. This enables continuous learning, where new tasks are taught to match the characteristics of the task. This reduces the wasted use of weights and increases the number of additional tasks that can be trained. This can be done. Also, the initialization of useful weights is reduced, and additional tasks This allows for maintaining a high level of inference accuracy.

[0068] Task 3 will be processed in the same way as Task 2, but the method for deriving task similarity will be different. They are different.

[0069] The task similarity deriving unit 21 calculates the task similarity 31 between task 3 and task 1, and the task similarity between task 3 and task 1. Derive the task similarity of Task 2, 32. Of the task similarity 31 and task similarity 32, The task with the higher similarity to the other task will be designated as the pre-trained task.

[0070] Generally, the task similarity derivation unit 21 selects a target task from among multiple trained tasks. Select the single pre-trained task with the greatest similarity to the given task as the pre-trained task.

[0071] However, if the number of tasks increases, deriving task similarity for all tasks is not possible. This is inefficient. Therefore, the targets for deriving task similarity should be selected as follows: It's also possible. (1) Select new tasks as priority for derivation. For example, newly entered tasks A predetermined number of tasks are left as targets for derivation, in the order of "K". (2) Tasks with a low initialization rate (unrelated tasks) are selected as priority for derivation. For example, a predetermined number of tasks are selected for derivation, in order of their initialization rate. (3) Select tasks with an initialization rate smaller than a predetermined value (unrelated tasks) as the target for derivation. Determine the value. For example, a predetermined number of tasks with an initialization rate less than a predetermined value are kept as the target for derivation. (4) Combinations of (1) and (2) above (5) Combinations of (1) and (3) above Thus, among the multiple pre-trained tasks, the one with the greatest similarity to the target task is selected. Alternatively, by using a relatively large task as a pre-trained task, the amount of time required for the target task can be reduced. The weight can be reduced.

[0072] Next, the inference device 200 and its operation will be described. Figure 9 shows the operation of the inference device 200 for task N. This is a flowchart explaining the operation.

[0073] The task determination unit 50 determines whether the task N input to the inference unit 70 is one of tasks 1 to 3. It is determined which task it is (S200). In this embodiment, the user is determined which task it is Specify whether to use it or not.

[0074] The inference model generation unit 60 generates an inference model based on the trained target model and effective parameter information. Generate a neural network model for inference (hereinafter referred to as the "inference model") (S2 10) The target model is a neural network trained on tasks 1 through 3. This is a work model. Task N is determined to be task i (where i is one of 1 to 3). In this case, the inference model generation unit 60 generates a target model based on the effective parameter information of task i. Dell generates an inference model where all weights except those used in task i are set to 0. Specifically, the inference model generation unit 60 reads the code sequence of the valid parameter information and the code If the sign is "1", the weight corresponding to that sign remains unchanged, on the other hand, if the sign is "0", If so, the weight corresponding to that sign may be changed to 0.

[0075] The inference unit 70 uses the inference model generated for task i to determine the inference result for task N that was input to it. Generate (S220).

[0076] In this embodiment, the initialization unit 26 initializes the weights of the trained neural network model. Initialization was performed in units of weights based on the initialization rate, but the initialization unit 26 is a pre-trained neural network The weights of the filter model may be initialized on a per-filter basis based on the initialization rate.

[0077] The various processes of the machine learning device 100 and inference device 200 described above are performed by the CPU and memory It can of course be realized as a device using hardware such as ROM ( Firmware stored in read-only memory or flash memory, etc. This can also be achieved through software such as a computer. Gram, recording software programs onto a recording medium that can be read by a computer, etc. It can be provided, or sent and received to the server via a wired or wireless network. It is also possible to transmit and receive this data as data broadcasting on terrestrial or satellite digital television.

[0078] As described above, according to the machine learning device 100 of this embodiment, the trained task and the target Utilization rate of target model weights that are continuously learned according to the similarity or correlation of the target task. By changing this setting, you can increase the number of additional tasks that can be trained and the accuracy of the added tasks relative to the target task. It can be optimized for that purpose.

[0079] The present invention has been described above based on embodiments. The embodiments are illustrative and their respective components The fact that various variations are possible in the combination of constituent elements and each processing process, and such variations Those skilled in the art will understand that this also falls within the scope of the present invention. [Explanation of Symbols]

[0080] 10 Task input unit, 20 Continuous learning unit, 21 Task similarity derivation unit, 22 Initial Periodization rate determination unit, 24 Machine learning execution unit, 26 Initialization unit, 28 Fine tuning 30 Input unit, 30 Memory unit, 40 Task input unit, 50 Task determination unit, 60 Inference unit Model generation unit, 70 Inference unit, 80 Inference result output unit, 100 Machine learning device, 200 Reasoning Apparatus.

Claims

1. A task similarity derivation unit that derives the task similarity between a trained task and a target task, The target is determined according to the depth of the layers in the neural network model and the task similarity. The initial initialization rate determines the initial weights of the neural network model for the task. The part that determines the concentration ratio, The initialized trained neural network model of the aforementioned trained task is the target Transfer learning is performed on the task to train the neural network model for the target task. A machine learning execution unit that generates Dell, The weights of the trained neural network model for the target task are initialized as follows: Initialize based on the above, and initialize the target task for use in the next task. The neural network model and the weights that have not been initialized after training for the target task. It is characterized by including an initialization unit that generates valid parameter information, which is information that identifies the user. A machine learning device.

2. A task similarity derivation step for deriving the task similarity between a trained task and a target task. P and, The target is determined according to the depth of the layers in the neural network model and the task similarity. The initial initialization rate determines the initial weights of the neural network model for the task. Steps to determine the conversion rate, The initialized trained neural network model of the aforementioned trained task is the target Transfer learning is performed on the task to train the neural network model for the target task. Machine learning execution steps to generate Dell, The weights of the trained neural network model for the target task are initialized as follows: Initialize based on the above, and initialize the target task for use in the next task. The neural network model and the weights that have not been initialized after training for the target task. It includes an initialization step that generates valid parameter information, which is information that identifies the specific parameter. A machine learning method performed by a computer.

3. A task similarity derivation step for deriving the task similarity between a trained task and a target task. P and, The target is determined according to the depth of the layers in the neural network model and the task similarity. The initial initialization rate determines the initial weights of the neural network model for the task. Steps to determine the conversion rate, The initialized trained neural network model of the aforementioned trained task is the target Transfer learning is performed on the task to train the neural network model for the target task. Machine learning execution steps to generate Dell, The weights of the trained neural network model for the target task are initialized as follows: Initialize based on the above, and initialize the target task for use in the next task. The neural network model and the weights that have not been initialized after training for the target task. Initialization step to generate valid parameter information, which is information that identifies the computer A machine learning program characterized by its execution.

Images