JP2021163424A - Ai system development simulator and hyper parameter optimization method - Google Patents

Abstract

To provide an AI system development simulator configured to select a machine learning algorithm and optimize a hyper parameter thereof in a short time.SOLUTION: An AI system development simulator includes: a hyper parameter storage unit which stores optimized hyper parameters obtained in the past by repeating trial of machine learning; a machine learning algorithm selection unit which selects one of a plurality of machine learning algorithms on the basis of recognition object information and recognition content information; a hyper parameter extraction unit which extracts an optimized hyper parameter obtained in the past from the hyper parameter storage unit; and a simulation unit which sets the extracted optimized hyper parameter obtained in the past to an initial value of a hyper parameter of the selected machine learning algorithm, and simulates execution of trial of machine learning using the machine learning algorithm to optimize the hyper parameter of the machine learning algorithm.SELECTED DRAWING: Figure 1

Description

The present invention relates to an AI simulator and a hyperparameter optimization method for streamlining the development of an AI system.

In order to develop an AI system that recognizes images and sounds, an appropriate machine learning algorithm (program for generating a model) is required according to the recognition target (attribute of information to be input) and recognition content (class to be identified). Must be selected. In addition, once the machine learning algorithm to be adopted is decided, when data is input to the machine learning algorithm and normal parameter machine learning (for example, learning of weights and biases in deep learning) is executed, hyperparameters (for example, deep layer) are executed. It is necessary to appropriately set the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, the number of epochs, etc. in learning.

Many methods for automatically adjusting hyperparameters have already been proposed (Non-Patent Document 1).

"Various automatic adjustment of hyperparameters", URL: https://qiita.com/cvusk/items/1f3b178f34c39beb29ff

However, in the conventional automatic adjustment method listed in Non-Patent Document 1 and the like, since it is necessary for an engineer to manually specify the range of hyperparameters to be optimized, a range larger than necessary tends to be specified. Become. In addition, in setting hyperparameters including selection of machine learning algorithms, it is necessary to repeat machine learning trials to optimize hyperparameters. Efficient optimization of hyperparameters depends on the skill level of the technician who does it. For this reason, in the development of AI systems that recognize images and sounds, a long time is often spent for selecting machine learning algorithms and optimizing their hyperparameters, which is a major obstacle to shortening the development period. It has become.

The present invention provides an AI system development simulator and a hyperparameter optimization method capable of selecting a machine learning algorithm and optimizing its hyperparameters in a short time.

In order to solve the above problem, the AI system development simulator according to claim 1 is an AI system development simulator that selects a machine learning algorithm and optimizes its hyper parameters, and recognizes for inputting information about a recognition target. A target information input unit, a recognition content information input unit for inputting information related to recognition content, a machine learning algorithm storage unit that stores a plurality of machine learning algorithms, and a machine learning algorithm storage unit obtained in the past by repeating machine learning trials. The hyper-parameter storage unit that stores the optimized hyper-parameters in association with the recognition target, recognition content, and machine learning algorithm used for the machine learning, the information input by the recognition target information input unit, and the recognition. A machine learning algorithm selection unit that selects one machine learning algorithm from a plurality of machine learning algorithms stored in the machine learning algorithm information storage based on the information input by the content information input unit, and the above-mentioned machine learning algorithm selection unit. From the hyper parameter storage unit based on the information input by the recognition target information input unit, the information input by the recognition content information input unit, and the machine learning algorithm selected by the machine learning algorithm selection unit. The machine learning algorithm selection unit uses the hyper parameter extraction unit for extracting the optimized hyper parameters obtained in the past and the optimized hyper parameters obtained in the past extracted by the hyper parameter extraction unit. A simulation unit that optimizes the hyperparameters of the machine learning algorithm by setting the initial values of the hyperparameters of the machine learning algorithm selected by It is characterized by having.

In the AI system development simulator of claim 1 configured as described above, information on the recognition target is input to the recognition target information input unit, and information on the recognition content is input to the recognition content information input unit. Multiple machine learning algorithms are stored in the machine learning algorithm storage unit, and the hyperparameters stored in the hyperparameter storage unit use the optimized hyperparameters obtained in the past by repeating machine learning trials for the machine learning. It is stored in association with the recognized recognition target, recognition content, and machine learning algorithm. The machine learning algorithm selection unit is a plurality of machine learning algorithms stored in the machine learning algorithm storage unit based on the information input by the recognition target information input unit and the information input by the recognition content information input unit. Select one machine learning algorithm from the list. The hyperparameter extraction unit is based on the information input by the recognition target information input unit, the information input by the recognition content information input unit, and the machine learning algorithm selected by the machine learning algorithm selection unit. Extract the optimized hyperparameters obtained in the past from the storage unit. The simulation unit sets the optimized hyperparameters obtained in the past extracted by the hyperparameter extraction unit to the initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection unit, and sets the machine. The hyperparameters of the machine learning algorithm are optimized by simulating machine learning trials using the learning algorithm.

According to the AI system development simulator of claim 1, it is possible to select a machine learning algorithm and optimize its hyperparameters in a short time.

In the AI system development simulator of claim 2, the hyperparameter includes any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs.

According to the AI system development simulator of claim 2, the selection of the machine learning algorithm and the optimization of any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs included in the hyperparameters thereof are optimized. It can be done in a short time.

In the AI system development simulator of claim 3, the recognition target is the data of the content screen constituting the content video included in the broadcast content or the distribution content, and the recognition content is an image element in the content screen. ..

According to the AI system development simulator of claim 3, selection of a machine learning algorithm that recognizes image elements in the content screen by targeting the data of the content screen constituting the content video included in the broadcast content or the distributed content, and the selection thereof. Hyperparameters can be optimized in a short time.

In the AI system development simulator of claim 4, the image element is any one of characters, a person's face, a program set, and a background.

According to the AI system development simulator of claim 4, the data of the content screen constituting the content video included in the broadcast content or the distributed content is the recognition target, and among the characters, the face of the person, the program set, and the background in the content screen. It is possible to select a machine learning algorithm that recognizes any of the above and optimize its hyperparameters in a short time.

In the AI system development simulator of claim 5, the recognition target is the content sound included in the broadcast content or the distribution content, and the recognition content includes any of a person's voice, a musical sound, a sound effect, and an environmental sound. Is included.

According to the AI system development simulator of claim 5, a machine learning algorithm that recognizes the content sound included in the broadcast content or the distributed content and recognizes any of a person's voice, a musical sound, a sound effect sound, and an environmental sound. Selection and optimization of its hyperparameters can be performed in a short time.

The hyper-parameter optimization method of claim 6 is a method of selecting a machine learning algorithm and optimizing the hyper-parameters, and inputs a recognition target input step for inputting information on a recognition target and information on the recognition content. The machine learning of the recognition content information input step for input, the machine learning algorithm storage step for storing a plurality of machine learning algorithms, and the optimized hyperparameters obtained in the past by repeating the machine learning trial. The hyperparameter storage step that is stored in association with the recognition target, recognition content, and machine learning algorithm used in the above, the information input by the recognition target information input step, and the information input by the recognition content information input step. Based on the above, a machine learning algorithm selection step for selecting one machine learning algorithm from a plurality of machine learning algorithms stored by the machine learning algorithm storage step, information input by the recognition target information input step, and the above. Based on the information input by the recognition content information input step and the machine learning algorithm selected by the machine learning algorithm selection step, the optimized hyper obtained in the past stored by the hyper parameter storage step. The hyperparameter extraction step for extracting parameters and the optimized hyperparameters obtained in the past extracted by the hyperparameter extraction step are used as the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection step. It is characterized by having a simulation step of optimizing the hyperparameters of the machine learning algorithm by setting an initial value and simulating a machine learning trial by the machine learning algorithm.

In the hyperparameter optimization method of claim 6, information on the recognition target is input by the recognition target input step, and information on the recognition content is input by the recognition content information input step. The machine learning algorithm storage step stores multiple machine learning algorithms, and the hyperparameter storage step uses the optimized hyperparameters obtained in the past by repeating machine learning trials for the machine learning. It is stored in association with the recognition target, recognition content, and machine learning algorithm. In the machine learning algorithm selection step, among a plurality of machine learning algorithms stored in the machine learning algorithm storage step based on the information input in the recognition target information input step and the information input in the recognition content information input step. One machine learning algorithm is selected from. In the hyperparameter extraction step, hyperparameters are based on the information input by the recognition target information input step, the information input by the recognition content information input step, and the machine learning algorithm selected by the machine learning algorithm selection step. The optimized hyperparameters obtained in the past stored by the storage step are extracted. Then, in the simulation step, the optimized hyperparameters obtained in the past extracted by the hyperparameter extraction step are set as the initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection step. , The process of optimizing the hyperparameters of the machine learning algorithm is executed by simulating the trial of machine learning by the machine learning algorithm.

According to the hyperparameter optimization method of claim 6, the machine learning algorithm can be selected and the hyperparameters can be optimized in a short time.

In the hyperparameter optimization method of claim 7, the hyperparameter includes any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs.

According to the hyperparameter optimization method of claim 7, selection of a machine learning algorithm and optimization of any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs included in the hyperparameters. Can be done in a short time.

In the hyperparameter optimization method of claim 8, the recognition target is the data of the content screen constituting the broadcast content or the content video included in the distribution content, and the recognition content is an image element in the content screen. be.

According to the hyperparameter optimization method of claim 8, selection of a machine learning algorithm that recognizes image elements in the content screen by targeting data on the content screen constituting the content video included in the broadcast content or the distributed content and selection of a machine learning algorithm. The hyperparameters can be optimized in a short time.

In the hyperparameter optimization method of claim 9, the image element is any one of characters, a person's face, a program set, and a background.

According to the hyperparameter optimization method of claim 9, the data of the content screen constituting the content video included in the broadcast content or the distributed content is recognized, and the characters, the face of a person, the program set, and the background in the content screen are recognized. It is possible to select a machine learning algorithm that recognizes one of them and optimize its hyperparameters in a short time.

In the hyperparameter optimization method of claim 10, the recognition target is the content voice included in the broadcast content, and the recognition content includes any one of a person's voice, a musical sound, a sound effect, and an environmental sound. included.

According to the hyperparameter optimization method of claim 10, a machine learning algorithm that recognizes a person's voice, a musical sound, a sound effect, or an environmental sound by targeting the content sound included in the broadcast content or the distributed content. And its hyperparameters can be optimized in a short time.

According to the AI system development simulator of claim 1 and the hyperparameter optimization method of claim 6, it is possible to select a machine learning algorithm and optimize the hyperparameters in a short time.
As a result, more optimal results can be obtained in a short time, and the speed of AI system development can be accelerated. In addition, by sharing the obtained data and the like among developers, the speed of hyperparameter optimization can be further increased. Therefore, it is possible to reduce the cost associated with AI development.

According to the AI system development simulator of claim 2 and the hyperparameter optimization method of claim 7, the selection of the machine learning algorithm and the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the epoch included in the hyperparameters. Optimization of any of the numbers can be done in a short time.

According to the AI system development simulator of claim 3 and the hyperparameter optimization method of claim 8, the data of the content screen constituting the broadcast content or the content video included in the distributed content is recognized, and the image element in the content screen is recognized. It is possible to select a machine learning algorithm that recognizes the above and optimize its hyperparameters in a short time.

According to the AI system development simulator of claim 4 and the hyperparameter optimization method of claim 9, the data of the content screen constituting the broadcast content or the content video included in the distributed content is recognized, and the characters in the content screen are recognized. It is possible to select a machine learning algorithm that recognizes a person's face, a program set, or a background and optimize its hyperparameters in a short time.

According to the AI system development simulator of claim 5 and the hyperparameter optimization method of claim 10, the content sound included in the broadcast content or the distributed content is the recognition target, and among the human voice, music sound, sound effect sound and environmental sound. It is possible to select a machine learning algorithm that recognizes any of the above and optimize its hyperparameters in a short time.

It is a functional block diagram which shows one Embodiment of the AI system development simulator of this invention. It is a flowchart which illustrates the operation content of the AI system development simulator shown in FIG.

Hereinafter, the AI system development simulator of one embodiment will be described with reference to the attached drawings. This AI system development simulator is an AI system development simulator that selects a machine learning algorithm and optimizes its hyperparameters. An example of an AI system to be developed is a system that automatically generates metadata of broadcast content or distribution content composed of video and audio. A television program can be mentioned as a typical example of broadcast content. The distributed content includes so-called video content distributed via an information communication network represented by the Internet. The metadata includes keywords for searching a program in which a specific person appears, a program having a specific content, and the like from many broadcasted programs.

[composition]
As shown in FIG. 1, the AI system development simulator 100 of one embodiment includes a recognition target information input unit 110, a recognition content information input unit 120, a machine learning algorithm storage unit 130, a hyperparameter storage unit 140, and machine learning. It has an algorithm selection unit 150, a hyperparameter extraction unit 160, and a simulation unit 170. The AI system development simulator 100 is realized by installing and executing a program for realizing the AI system development simulator according to the present invention on a computer. The program includes a program for executing the hyperparameter optimization method according to the present invention using a computer.

The recognition target information input unit 110 is a functional block for inputting information related to the recognition target. The information regarding the recognition target includes information indicating data attributes such as whether the data to be recognized (input data) is image data or audio data, as well as the size of the input data, the number of channels, and the like. Examples of the recognition target include data on the content screen constituting the content video included in the broadcast content or the distributed content, data of the content audio included in the broadcast content or the distributed content, and the like.

The recognition content information input unit 120 is a functional block for inputting information related to the recognition content. The information about the recognition content includes the number of classes. As an example of the recognized content, an image element in the content screen constituting the content video included in the broadcast content or the distributed content can be mentioned. Examples of the image element include characters, a person's face, a program set, a background, and the like. For example, when the recognition content is a character, the number of classes is the number of character types to be recognized (identified). When the recognition target is the data of the content voice included in the broadcast content or the distribution content, examples of the recognition content include a person's voice, a musical sound, a sound effect, an environmental sound, and the like. For example, when the recognition content is the voice of a plurality of specific persons, the number of classes is the number of persons to be recognized (identified).

The machine learning algorithm storage unit 130 is a functional block that stores a plurality of machine learning algorithms. A machine learning algorithm is a program for generating a model.

The hyperparameter storage unit 140 has a function of storing optimized hyperparameters obtained in the past by repeating machine learning trials in association with the recognition target, recognition content, and machine learning algorithm used for the machine learning. It is a block. The hyperparameter stored in the hyperparameter storage unit 140 includes any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs. Further, the hyperparameter storage unit 140 may store setting information regarding the learned model used in the past for image recognition and voice recognition, and information regarding the hardware used at that time. The configuration information about the trained model includes information about the machine learning algorithm used to generate the trained model. The hardware information includes hardware spec information. The hardware spec information includes CPU (Central Processing Unit) spec information, GPU (Graphics processing unit) spec information, and the like.

The machine learning algorithm selection unit 150 has a plurality of information stored in the machine learning algorithm storage 130 based on the information input by the recognition target information input unit 110 and the information input by the recognition content information input unit 120. It is a functional block that selects one machine learning algorithm from the machine learning algorithms.

The hyperparameter extraction unit 160 is based on the information input by the recognition target information input unit 110, the information input by the recognition content information input unit 120, and the machine learning algorithm selected by the machine learning algorithm selection unit 150. It is a functional block that extracts the optimized hyperparameters obtained in the past from the hyperparameter storage unit 140.

The simulation unit 170 sets the optimized hyperparameters obtained in the past extracted by the hyperparameter extraction unit 160 as the initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection unit 150. , It is a functional block that optimizes the hyperparameters of the machine learning algorithm by simulating a machine learning trial by the machine learning algorithm. The simulation unit 170 repeats the simulated trial of machine learning by the machine learning algorithm while automatically changing the hyperparameters of the machine learning algorithm that tries the machine learning in a simulated manner.

Spec information of the hardware used for machine learning can be input to the simulation unit 170 as a condition. The simulation unit 170 saves the set conditions (hyperparameters, hardware spec information) and results (recognition accuracy, required time, consumed energy, etc.) each time one simulation process is completed, and a plurality of them. The conditions are optimized based on the conditions and results of the simulations executed once. The conditions may include the number of classes, the size of training data, the number of steps, the batch size, the number of iterations, and the like. In addition, the result may include information indicating that learning was not possible due to an error.

[motion]
As shown in FIG. 2, the AI system development simulator 100 includes a recognition target information input step S1, a recognition content information input step S2, a machine learning algorithm storage step S3, a hyperparameter storage step S4, and a machine learning algorithm selection step S5. And the hyperparameter extraction step S6 and the simulation step S7 are executed.

The recognition target information input step S1 is a step for inputting information regarding the recognition target.

The recognition content information input step S2 is a step for inputting information regarding the recognition content.

The machine learning algorithm storage step S3 is a step of storing a plurality of machine learning algorithms.

In the hyperparameter storage step S4, the optimized hyperparameters (number of layers, number of nodes, batch size, learning rate, number of iterations, number of epochs, etc.) obtained in the past by repeating machine learning trials are stored. It is a step to memorize the recognition target, recognition content and machine learning algorithm used for machine learning in association with each other.

The machine learning algorithm selection step S5 is a plurality of machines stored in the machine learning algorithm storage step S3 based on the information input in the recognition target information input step S1 and the information input in the recognition content information input step S2. This is the step of selecting one machine learning algorithm from the learning algorithms.

The hyperparameter extraction step S6 is based on the information input in the recognition target information input step S1, the information input in the recognition content information input step S2, and the machine learning algorithm selected in the machine learning algorithm selection step S5. This is a step of extracting the optimized hyperparameters obtained in the past stored by the hyperparameter storage step.

In the simulation step S7, the optimized hyperparameters obtained in the past extracted by the hyperparameter extraction step S6 are set as the initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection step. This is a step of optimizing the hyperparameters of the machine learning algorithm by simulating a machine learning trial by the machine learning algorithm.

[Action effect]
As described above, in the AI system development simulator 100 of one embodiment, the information regarding the recognition target is input to the recognition target information input unit 110, and the information regarding the recognition content is input to the recognition content information input unit 120. A plurality of machine learning algorithms are stored in the machine learning algorithm storage unit 130, and the hyperparameter storage unit 140 stores optimized hyperparameters (number of layers, nodes) obtained in the past by repeating machine learning trials. Number, batch size, learning rate, number of iterations, number of epochs, etc.) are stored in association with the recognition target, recognition content, and machine learning algorithm used for the machine learning. The machine learning algorithm selection unit 150 is stored in the machine learning algorithm storage unit 130 based on the information input by the recognition target information input unit 110 and the information input by the recognition content information input unit 120. Select one machine learning algorithm from the machine learning algorithms of. The hyperparameter extraction unit 160 is based on the information input by the recognition target information input unit 110, the information input by the recognition content information input unit 120, and the machine learning algorithm selected by the machine learning algorithm selection unit 150. Then, the optimized hyperparameters obtained in the past are extracted from the hyperparameter storage unit. Then, the simulation unit 170 sets the optimized hyperparameters obtained in the past extracted by the hyperparameter extraction unit 160 as the initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection unit 150. Then, the hyperparameters of the machine learning algorithm are optimized by simulating the trial of machine learning by the machine learning algorithm.

In this way, according to the AI system development simulator of one embodiment, the selection of machine learning algorithms and the optimization of their hyperparameters, which conventionally depended on the skill level of engineers, are fully automated and in a short time. It can be carried out.

Further, according to the AI system development simulator 100 of one embodiment, the data of the content screen constituting the broadcast content or the content video included in the distributed content is targeted for recognition, and the image elements (characters, person's face, etc.) in the content screen are recognized. It is possible to select a machine learning algorithm that recognizes a program set, background, etc. and optimize its hyperparameters in a short time.

Further, according to the AI system development simulator 100 of one embodiment, a machine that recognizes a person's voice, a musical sound, a sound effect, or an environmental sound by targeting the content sound included in the broadcast content or the distributed content. It is possible to select a learning algorithm and optimize its hyperparameters in a short time.

In this way, by being able to select machine learning algorithms and optimize their hyperparameters in a short time, the development period of a system that automatically generates metadata of broadcast content or distribution content consisting of video and audio has been extended. Can be significantly shortened.

Further, according to the AI system development simulator 100 of one embodiment, by setting the spec information of the hardware (CPU, GPU, etc.) as the simulation conditions, the recognition accuracy, the required time, and the consumption are consumed as the result of the simulation. Since the energy, etc. can be confirmed, it is possible to evaluate the hardware used for image recognition and voice recognition. For example, if a GPU of a certain grade is used, the memory capacity is insufficient, so it is possible to evaluate that it is necessary to use a GPU of another grade.
As a result, it is possible to construct an AI system with high cost performance including the hardware performance of the CPU, GPU, memory, etc., and it is possible to reduce the development cost required for AI system development.

100 AI system development simulator 110 Recognition target information input unit 120 Recognition content information input unit 130 Machine learning algorithm storage unit 140 Hyperparameter storage unit 150 Machine learning algorithm selection unit 160 Hyperparameter extraction unit 170 Simulation unit S1 Recognition target information input step S2 Recognition Content information input step S3 Machine learning algorithm storage step S4 Hyperparameter storage step S5 Machine learning algorithm selection step S6 Hyperparameter extraction step S7 Simulation step

Claims (10)

An AI system development simulator that selects machine learning algorithms and optimizes their hyperparameters.
A recognition target information input unit for inputting information about the recognition target,
A recognition content information input unit for inputting information related to recognition content,
A machine learning algorithm storage unit that stores multiple machine learning algorithms,
A hyperparameter storage unit that stores optimized hyperparameters obtained in the past by repeating machine learning trials in association with the recognition target, recognition content, and machine learning algorithm used for the machine learning.
Based on the information input by the recognition target information input unit and the information input by the recognition content information input unit, one of a plurality of machine learning algorithms stored in the machine learning algorithm information storage. Machine learning algorithm selection section that selects the machine learning algorithm of
The hyperparameter storage unit is based on the information input by the recognition target information input unit, the information input by the recognition content information input unit, and the machine learning algorithm selected by the machine learning algorithm selection unit. The hyperparameter extraction unit that extracts the optimized hyperparameters obtained in the past from
The optimized hyperparameters obtained in the past extracted by the hyperparameter extraction unit are set as initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection unit, and the machine learning is performed. An AI system development simulator characterized by having a simulation unit that optimizes hyperparameters of the machine learning algorithm by simulating a machine learning trial by the algorithm. The AI system development simulator according to claim 1, wherein the hyperparameter includes any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs. The recognition target is data on the content screen that constitutes the content video included in the broadcast content or the distributed content.
The AI system development simulator according to claim 1 or 2, wherein the recognition content is an image element in the content screen. The AI system development simulator according to claim 3, wherein the image element is any one of characters, a person's face, a program set, and a background. The recognition target is content audio included in broadcast content or distribution content, and is
The AI system development simulator according to claim 3 or 4, wherein the recognition content includes any one of a person's voice, a musical sound, a sound effect, and an environmental sound. It is a method of selecting a machine learning algorithm and optimizing its hyperparameters.
The recognition target information input step for inputting information about the recognition target, and
Recognition content information input step for inputting information about recognition content,
Machine learning algorithm storage steps that store multiple machine learning algorithms,
A hyperparameter storage step that stores optimized hyperparameters obtained in the past by repeating machine learning trials in association with the recognition target, recognition content, and machine learning algorithm used for the machine learning.
One machine from a plurality of machine learning algorithms stored by the machine learning algorithm storage step based on the information input by the recognition target information input step and the information input by the recognition content information input step. Machine learning algorithm selection step to select learning algorithm and
The hyperparameter storage step is based on the information input by the recognition target information input step, the information input by the recognition content information input step, and the machine learning algorithm selected by the machine learning algorithm selection step. The hyperparameter extraction step for extracting the optimized hyperparameters obtained in the past, which was stored in
The optimized hyperparameters obtained in the past extracted by the hyperparameter extraction step are set as initial values of the hyperparameters of the machine learning algorithm selected by the machine learning algorithm selection step, and the machine learning is performed. A hyperparameter optimization method comprising: a simulation step for optimizing the hyperparameters of the machine learning algorithm by simulating a machine learning trial by the algorithm. The hyperparameter optimization method according to claim 6, wherein the hyperparameter includes any one of the number of layers, the number of nodes, the batch size, the learning rate, the number of iterations, and the number of epochs. The recognition target is data on the content screen that constitutes the content video included in the broadcast content or the distributed content.
The hyperparameter optimization method according to claim 6 or 7, wherein the recognition content is an image element in the content screen. The hyperparameter optimization method according to claim 8, wherein the image element is any one of characters, a person's face, a program set, and a background. The recognition target is content audio included in broadcast content or distribution content, and is
The hyperparameter optimization method according to claim 8 or 9, wherein the recognition content includes any one of a person's voice, a musical sound, a sound effect, and an environmental sound.

Images