WO2022176481A1 - 機械学習用データ生成方法、メタ学習方法、機械学習用データ生成装置及びプログラム - Google Patents
機械学習用データ生成方法、メタ学習方法、機械学習用データ生成装置及びプログラム Download PDFInfo
- Publication number
- WO2022176481A1 WO2022176481A1 PCT/JP2022/001626 JP2022001626W WO2022176481A1 WO 2022176481 A1 WO2022176481 A1 WO 2022176481A1 JP 2022001626 W JP2022001626 W JP 2022001626W WO 2022176481 A1 WO2022176481 A1 WO 2022176481A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- augmentation
- machine learning
- learning
- domain
- Prior art date
Links
- 238000010801 machine learning Methods 0.000 title claims abstract description 130
- 238000000034 method Methods 0.000 title claims abstract description 95
- 238000013434 data augmentation Methods 0.000 claims abstract description 101
- 238000012549 training Methods 0.000 claims abstract description 69
- 230000003416 augmentation Effects 0.000 claims abstract description 48
- 230000008859 change Effects 0.000 claims description 25
- 230000002787 reinforcement Effects 0.000 claims description 14
- 230000007613 environmental effect Effects 0.000 claims description 13
- 238000013075 data extraction Methods 0.000 claims description 11
- 239000000284 extract Substances 0.000 claims description 9
- 230000008569 process Effects 0.000 claims description 9
- 230000002194 synthesizing effect Effects 0.000 claims description 7
- 238000001914 filtration Methods 0.000 claims description 6
- 238000003384 imaging method Methods 0.000 claims description 6
- 238000000605 extraction Methods 0.000 description 19
- 238000012545 processing Methods 0.000 description 18
- 238000013500 data storage Methods 0.000 description 15
- 238000010586 diagram Methods 0.000 description 15
- 238000003860 storage Methods 0.000 description 11
- 238000012552 review Methods 0.000 description 5
- 230000003190 augmentative effect Effects 0.000 description 3
- 238000004891 communication Methods 0.000 description 3
- 238000009826 distribution Methods 0.000 description 3
- 238000007689 inspection Methods 0.000 description 3
- 239000000463 material Substances 0.000 description 3
- 238000012986 modification Methods 0.000 description 3
- 230000004048 modification Effects 0.000 description 3
- 238000012360 testing method Methods 0.000 description 3
- 230000002950 deficient Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 230000006698 induction Effects 0.000 description 2
- 230000009467 reduction Effects 0.000 description 2
- 238000009966 trimming Methods 0.000 description 2
- 238000012795 verification Methods 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000004458 analytical method Methods 0.000 description 1
- 230000015572 biosynthetic process Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000004590 computer program Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 230000006870 function Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 238000011946 reduction process Methods 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000003786 synthesis reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/004—Artificial life, i.e. computing arrangements simulating life
- G06N3/006—Artificial life, i.e. computing arrangements simulating life based on simulated virtual individual or collective life forms, e.g. social simulations or particle swarm optimisation [PSO]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T11/00—2D [Two Dimensional] image generation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/77—Processing image or video features in feature spaces; using data integration or data reduction, e.g. principal component analysis [PCA] or independent component analysis [ICA] or self-organising maps [SOM]; Blind source separation
- G06V10/774—Generating sets of training patterns; Bootstrap methods, e.g. bagging or boosting
Definitions
- the present invention relates to a machine learning data generation method, a meta-learning method, a machine learning data generation device, and a program.
- Patent Literature 1 describes a training method for a meta-learning network.
- Generalization learning requires a lot of domain data to improve generalization performance, but acquiring new data is costly.
- the present invention has been made in view of the circumstances described above, and aims to provide a method of generating training data capable of improving the generalization ability of a learning model.
- a machine learning data generation method is a data generation method for generating data for domain generalization in machine learning, wherein a computer uses learning data for training a machine learning model. performing the augmentation as data; and extracting, by a computer, a dataset containing both the original data and the data generated by the data augmentation as the dataset for the domain generalization. , is included.
- a domain is a dataset acquired in an environment.
- Machine learning includes supervised learning, unsupervised learning, and reinforcement learning.
- supervised learning a dataset is a collection of pairs of data and teacher labels.
- unsupervised learning a dataset is a collection of data.
- reinforcement learning a dataset is the state of the space (environment) in which an agent exists.
- Domain generalization is a method of constructing a robust model against domain shifts between unknown domains and distributions by using training sets extracted from multiple distributions.
- the domain shift represents the shift in the distribution followed by the data between the training set and the test set.
- a training set is a collection of data used for model training
- a test set is a collection of data used for model verification (testing).
- the original data and the data generated by the data augmentation are each held in units of domains, and in the extracting step, at least one domain of the original data and the data generated by the data augmentation At least one domain of the obtained data may be extracted as a data set for said domain generalization.
- the domain for learning can be extracted so as to include both the domain of the original data and the domain of the generated data, thereby preventing the induction of over-learning due to bias toward the original data or the generated data. can be done.
- the original data may include a target portion and a non-target portion, and in the step of performing the augmentation, data augmentation may be performed to change the non-target portion in the original data.
- a target portion is a portion that directly affects a subject (task) to be learned by machine learning, and is a portion targeted by the task.
- the non-target portion is a portion that does not affect the task and mainly corresponds to the environment (for example, background, brightness, etc.) in which data is acquired.
- the target part is the part that affects the relationship between the data and the teacher label, for example, the information corresponding to the target object to be recognized.
- An object is an object to be judged, which is the purpose of using machine learning.
- a recognition target in image recognition for example, a vehicle in vehicle recognition
- voice data excluding noise and environmental sound in voice recognition text in meaning extraction, and the like.
- the non-target portion in supervised learning includes the environment in which the target exists (for example, background, brightness, etc.).
- the target part is the part that affects the relationship between the data and the features to be obtained from the data, for example, the information of the part corresponding to the object to be clustered.
- An object is an object to be judged, which is the purpose of using machine learning, as in supervised learning.
- the untargeted part in unsupervised learning includes the environment in which the target exists (for example, background, brightness, etc.).
- the target part is the information that affects the acquisition of the "reward” (task achievement) in the "environment” in which the "agent” is placed.
- the information is information about the object and factors affecting gripping of the object (for example, the inclination of the place where the object is placed, friction, etc.).
- the non-target portion in reinforcement learning is the information of the portion that does not affect the reward, such as the color and brightness of the place where the target object is placed.
- the original data is image data
- the target portion is an image of an object
- the environment around the object and/or the Data augmentation may be performed to change the imaging conditions of the object.
- the step of performing the augmentation by performing at least one of changing the brightness of the image of the original data, changing the background, and changing the color tone, the environment around the object is changed.
- Alternate data augmentation may be implemented. As a result, it is possible to generate a variety of training data assuming changes in shooting environments such as time and weather.
- the step of performing the augmentation by performing at least one of rotation, inversion, enlargement, reduction, movement, trimming, and filtering of the image of the original data, the image of the object is captured.
- Data augmentation with varying conditions may be implemented. As a result, it is possible to generate a variety of training data assuming differences in imaging conditions.
- the original data is voice data
- the target portion is a specific voice
- the step of performing the augmentation includes data augmentation for changing environmental sounds and noises in the voice data of the original data.
- the data augmentation may be performed by synthesizing the audio data of the original data with the audio of the environmental sound. As a result, it is possible to generate a variety of training data assuming differences in locations where voice data was obtained.
- the original data is signal data
- the target part is a specific signal pattern
- data augmentation is performed to change noise in the signal data of the original data.
- the data augmentation may be performed by synthesizing noise with the signal data of the original data. As a result, it is possible to generate a variety of training data assuming differences in the locations where the signal data were acquired.
- the original data is text data
- the target portion is a specific text pattern
- data augmentation is performed to change the writing style of the text data of the original data. You may make it As a result, in training of a learning model such as semantic extraction, various training data effective for generalization learning can be generated.
- the data augmentation may be performed by changing the beginning and/or ending of the text data of the original data. As a result, it is possible to generate various training data assuming different styles of writing.
- the original data is data related to the state of the environment in which the agent is placed in reinforcement learning
- the target portion is information on the portion that affects acquisition of reward
- the step of performing the augmentation includes the Data augmentation may be performed to change the conditions of the portion of the environmental state in the original data that does not affect acquisition of the reward.
- the domain of the original data and the domain of the data generated by the data augmentation may be extracted so as to include a predetermined ratio.
- the domain for learning can be extracted so as to include both the domain of the original data and the domain of the generated data, thereby preventing the induction of over-learning due to bias toward the original data or the generated data. can be done.
- the predetermined ratio may be specified by the user, or may be held as a parameter in advance.
- a meta-learning method includes the step of performing domain generalization by meta-learning using the data set for domain generalization generated by the above data generation method for machine learning.
- a set of a plurality of data sets including at least one domain of the original data and at least one domain of the data generated by the data augmentation may be used to perform domain generalization by meta-learning. This makes it possible to perform meta-learning using a data set that always includes both original data and data generated by data augmentation, prevent over-learning, and improve the generalization ability of domain generalization.
- a domain generalization learning method includes the step of performing domain generalization learning using the data set for domain generalization generated by the above data generation method for machine learning.
- a machine learning data generation device is a data generation device that generates data for domain generalization in machine learning, wherein learning data used for training a machine learning model is used as original data.
- a data generation unit that performs augmentation; and a training data extraction unit that extracts a dataset containing both the original data and the data generated by the data augmentation as a dataset for the domain generalization.
- a program according to one aspect of the present invention is a computer that generates data for domain generalization in machine learning. and a training data extraction unit that extracts a data set containing both the original data and the data generated by the data augmentation as a data set for the domain generalization.
- FIG. 4 is a flowchart showing an example of the operation of the machine learning data generation device according to the embodiment of the present invention; 4 is a flowchart showing an example of the operation of the machine learning data generation device according to the embodiment of the present invention; The figure explaining an example of the data generation processing for machine-learnings which concerns on embodiment of this invention.
- FIG. 4 is a diagram showing an example of a training data set extraction method by the machine learning data generation device according to the embodiment of the present invention
- FIG. 4 is a diagram showing an example of a training data set extraction method by the machine learning data generation device according to the embodiment of the present invention
- FIG. 5 is a diagram illustrating data augmentation according to data type by machine learning data generation processing according to the embodiment of the present invention
- FIG. 4 is a flowchart showing an example of the operation of the machine learning data generation device according to the embodiment of the present invention; The figure explaining an example of the data generation processing for machine-learnings which concerns on embodiment of this invention. The figure explaining an example of the data generation processing for machine-learnings which concerns on embodiment of this invention.
- FIG. 5 is a diagram illustrating data augmentation according to data type by machine learning data generation processing according to the embodiment of the present invention; The figure explaining an example of the data generation processing for machine-learnings which concerns on embodiment of this invention. The figure explaining an example of the data generation processing for machine-learnings which concerns on embodiment of this invention.
- this embodiment an embodiment according to one aspect of the present invention (hereinafter also referred to as "this embodiment") will be described based on the drawings.
- the embodiments described below are merely examples of the present invention in all respects. It goes without saying that various modifications and variations can be made without departing from the scope of the invention. That is, in implementing the present invention, a specific configuration according to the embodiment may be appropriately adopted.
- the data appearing in this embodiment are explained in terms of natural language, more specifically, they are specified in computer-recognizable pseudo-language, commands, parameters, machine language, and the like.
- FIG. 1 is a diagram for explaining an outline of a machine learning data generation process by a machine learning data generation device 10 according to the present invention.
- data augmentation is performed using image data captured by a camera installed at a point A as original data (original data) to generate training data.
- the original data is divided into four domain data sets and stored: image data taken in rainy weather, image data taken in snowfall, image data taken in daytime, and image data taken at night.
- image data taken in rainy weather image data taken in snowfall
- image data taken in daytime image data taken at night.
- each data after augmentation is divided into data for rainy weather, data for snowfall, daytime data, and nighttime data.
- the data are divided into datasets and saved. Note that data augmentation is sometimes called data augmentation.
- a predetermined number of data are extracted so as to include both the original data and generated data domains, and are used as a training data set for meta-learning. For example, from the original data, we extract the data in each domain of rain data and nighttime data, and from the generated data, we extract the data in each domain of snowfall data and daytime data, and execute meta-learning as a training dataset. . Note that the original data and the generated data are data having a predetermined structure for meta-learning.
- the machine learning data generation device 10 includes, as its hardware resources, a processor 11, a main memory 12, a camera interface 13, an input/output interface 14, a display interface 15, a communication interface 16, and a storage device 17. computer system.
- the storage device 17 is a computer-readable recording medium such as a disk medium (eg, magnetic recording medium or magneto-optical recording medium) or semiconductor memory (eg, volatile memory or non-volatile memory). Such a recording medium can also be called a non-transitory recording medium, for example.
- a generalization learning program 20 is stored in the storage device 17 .
- the generalization learning program 20 is a computer program for causing the processor 11 to execute the meta-learning method according to this embodiment.
- the generalization learning program 20 is loaded from the storage device 17 into the main memory 12 and interpreted and executed by the processor 11, thereby executing the meta-learning method according to the present embodiment.
- a camera 51 is connected to the camera interface 13 .
- Camera 51 may include, for example, an image sensor that captures color images.
- the camera 51 may be built in the machine learning data generation device 10 or may be externally attached to the machine learning data generation device 10 .
- An image captured by the camera 51 is stored in the original data storage section 31 of the storage device 17 as original data.
- An input device 52 and an output device 53 are connected to the input/output interface 14 .
- the input device 52 is, for example, a keyboard, mouse, touch pad, or the like.
- the output device 53 is a device that outputs various processing results and the like.
- the output device 53 is, for example, a printer.
- a display device 54 is connected to the display interface 15 .
- the display device 54 displays a user interface for accepting instructions from the user, and original data and generated data for data augmentation.
- FIG. 3 is a block diagram showing the functional modules executed by processor 11. As shown in FIG. As shown in FIG. 3, the functional module includes a data generation unit 21, a training data extraction unit 22, and a learning execution unit .
- An original data storage unit 31 and a generated data storage unit 32 are mounted in the storage device 17 .
- the original data storage unit 31 stores original data for data augmentation captured by the camera 51 or the like.
- the generated data storage unit 32 stores data generated by data augmentation.
- FIG. 1 a machine learning data generation method by the machine learning data generation device 10 according to the present embodiment will be described with reference to FIGS. 4 to 8.
- FIG. an example will be described in which data augmentation is performed on image data serving as original data to generate new image data.
- FIG. 4 is a flow chart showing an example of a machine learning data generation method according to an embodiment of the present invention.
- supervised learning is assumed as the machine learning method, but it can be applied to unsupervised learning and reinforcement learning as well.
- step S101 the original data is stored for each domain in the original data storage unit 31 of the machine learning data generation device 10.
- the original data is a set of teacher data created by annotating image data taken at a point A in advance.
- the original data is divided into four domains, data for rainy weather, data for snowfall, data for daytime, and data for nighttime, and stored according to the weather and time of day when the image was taken.
- the machine-learning data generation device 10 may acquire an image captured by the camera at the point A via a communication line, or the machine-learning data generation device 10 may acquire image data stored in an external storage device. may be copied to the storage device 17 of the
- step S102 the machine learning data generation device 10 accepts the data augmentation method.
- the data augmentation method can be input by the user via the input device 52, for example.
- a data augmentation method for example, as shown in FIG. 1, specific processing such as image inversion can be specified.
- step S103 the data generation unit 21 of the machine learning data generation device 10 performs data augmentation on the original data based on the designated method. For example, a process of horizontally reversing each image data stored in the original data storage unit 31 is executed.
- step S104 the data generation unit 21 of the machine learning data generation device 10 stores the image data generated by the data augmentation in the generated data storage unit 32.
- the generated data may also be stored for each domain corresponding to the domain of the original data.
- the image data obtained by augmenting the rain data is stored in the domain of the rain data in the generated data storage unit 32 .
- step S ⁇ b>105 the machine learning data generation device 10 executes meta-learning-based generalization learning by the learning execution unit 23 .
- step S105 the processing of steps S1051 to S1055 is repeated the number of times of learning.
- MLDG Metal-Learning for Domain Generalization
- any generalized learning method based on meta-learning may be used.
- the training data extraction unit 22 extracts data from the storage device 17 so as to include both the domain of the original data stored in the original data storage unit 31 and the domain of the generated data stored in the generated data storage unit 32. to extract multiple domains. Based on the domain extraction parameter, the training data extraction unit 22 extracts the domain of the original data and the domain of the generated data so as to include a predetermined ratio.
- the domain extraction parameter may be a value specified each time via the input device 52, or may be stored in advance by storing a domain extraction parameter definition table in the storage device 17. FIG.
- step S1052 the training data extraction unit 22 divides the multiple domains extracted in step S1051 into dataset 1 (training domain) and dataset 2 (verification domain).
- step S1053 the learning executing unit 23 calculates a loss (loss1), which is the difference between the correct label and the predicted label, for data set 1, and temporarily updates network parameters (machine learning parameters).
- step S1054 the learning execution unit 23 calculates a loss (loss2) for data set 2 using the network parameters updated in step S1053 as initial values.
- step S1055 the learning execution unit 23 updates the network parameters so as to minimize the weighted sum of loss1 and loss2.
- the network parameters are optimized by repeating the above steps S1051 to S1055 for the number of times of learning.
- step S201 the original data is stored in the original data storage unit 31 for each domain, as in step S101 of FIG.
- step S202 similar to step S102 in FIG. 4, data augmentation methods are accepted, but in the example of FIG. 5, a plurality of methods (n ways) can be accepted at the same time.
- the data generation unit 21 sequentially performs data augmentation using n designated methods to create generated data. Specifically, first, the data augmented by the first method (for example, inversion) is temporarily stored as generated data, and then the data that has undergone the inversion is subjected to the second method ( For example, enlargement) is performed and stored as generated data. Similarly, after performing augmentation by the n-th method, the generated data is stored as final generated data for each domain in the generated data storage unit 32 .
- the first method for example, inversion
- the second method For example, enlargement
- step S ⁇ b>205 the machine learning data generation device 10 executes meta-learning-based generalization learning by the learning execution unit 23 .
- step S205 processing similar to steps S1051 to S1055 in FIG. 4 is repeated for the number of times of learning.
- the original data includes a part that should remain unchanged before and after data augmentation (target part) and a part that should be expanded or changed by data augmentation (non-target part).
- target part a part that should remain unchanged before and after data augmentation
- non-target part a part that should be expanded or changed by data augmentation
- the target (non-target part) to be expanded or changed by data augmentation is the surrounding environment (e.g. landscape, road shape, etc.) and shooting conditions (e.g. magnification, shooting direction, weather, time of day, etc.). is.
- Specific augmentation methods for image data include rotating, reversing, enlarging, reducing, moving, trimming, and filtering the image of the original data, as well as changing the brightness of the image of the original data (weather, time, etc.). Examples include variations in time zone), background changes, and color tone changes (weather and time zone variations).
- FIG. 6 is a diagram for explaining an example of performing filter processing, for example. In the example of FIG. 6, the original data of daytime data and nighttime data are subjected to filtering processing for extracting images in specific weather (snowfall or cloudy weather).
- data augmentation may be performed to generate different domains by performing different types of augmentation on the same original data.
- a generated data domain A' is obtained by performing a rotation process on the original data domain A
- a generated data domain A is obtained by performing a reduction process on the domain A.
- the same type of expansion may be performed at different strengths (levels).
- levels levels
- for domain B of the original data 45 It is also possible to store the generated data domain B′ that has been rotated by degrees (strength), and the generated data domain B′′ that has been rotated by 90 degrees (strength).
- We also perform different types of data augmentation for different domains of the original data (rotation and shrinkage for domain A, 45 degree rotation and 90 degree rotation for domain B), as shown in the example of FIG. You may do so.
- the extracted ⁇ domains are divided into two data sets (Dataset1, Dataset2), and meta-learning is repeatedly performed so as to minimize the sum of losses of both.
- the extraction ratio ⁇ of the domain of the original data is changed each time learning is performed, and the number of extractions ⁇ is constant.
- the extracted ⁇ domains are divided into two data sets (Dataset1, Dataset2), and meta-learning is repeatedly performed so as to minimize the sum of losses of both.
- parameters such as the extraction ratio and the number of extractions may be repeatedly learned with constant values without changing for each learning.
- the extraction parameter is not limited to the above example, and for example, instead of the ratio of the original data, the number of domains of the original data to be extracted may be used as a parameter.
- FIG. 10 shows various training data to which the machine learning data generation method can be applied when the machine learning method is supervised learning or unsupervised learning, and an example of data augmentation when using each data. It is a figure which shows.
- data type is the type of training data
- task example is an example of a model that uses the training data
- target is the part that should remain unchanged before and after data augmentation (target part )
- change target indicates a portion to be expanded/changed by data augmentation (non-target portion)
- augmentation example indicates the content of change to be performed on the change target.
- the target object is a specific object (for example, a vehicle for vehicle recognition, a face for face recognition, etc.).
- subject to change and augmentation include changing the shooting time by changing the light and dark, representing the difference between indoors and outdoors by replacing the background, representing the difference in season and scenery by changing the color, fogging of the lens and noise.
- the difference in focus is expressed by filtering
- the inclination of the camera is changed by rotation
- the difference in shooting position is expressed by movement or enlargement/reduction.
- the target object may be a specific speech (for example, human voice, etc.).
- objects to be changed and augmentation include addition of environmental sounds (for example, running sounds of cars, operating sounds of machines, etc.) by synthesis.
- the target object may be a specific waveform pattern (for example, abnormal sound, etc.).
- objects to be changed and augmented include addition of environmental sounds (such as operating sounds of machines, etc.), vibrations, and noise from microphones and sensors using synthetic signals.
- the target can be a specific text (for example, review articles, etc.).
- modification objects and augmentations include changing the tone of a sentence by replacing word endings (eg, interjections, symbols, etc.).
- FIG. 11 and 12 are diagrams showing an example of a machine learning data generation method when the machine learning method is supervised learning and the data type is voice.
- the original data is voice data recorded indoors (quiet place), and noise is added to the data assuming the inside of a train station.
- data augmentation is performed by synthesizing noise (for example, train sounds, talking voices, announcement sounds, etc.) at various stations (for example, subways, bullet trains, etc.) with the original data.
- noise for example, train sounds, talking voices, announcement sounds, etc.
- stations for example, subways, bullet trains, etc.
- data obtained by diversifying the data acquisition environment is generated using audio data recorded indoors (quiet place) as the original data. Specifically, by adding sounds such as car and train sounds, people's voices, rain sounds, etc., data that is assumed to have been acquired near railway tracks, in offices, outdoors, etc. is generated. By using data generated in this way for meta-learning, it is possible to generalize machine learning so as to be able to cope with various data acquisition environments.
- FIG. 13 is a diagram showing an example of a machine learning data generation method when the machine learning method is supervised learning and the data type is a signal.
- FIG. 13 shows an example of generating training data for a model that analyzes machine vibration data and detects anomalies.
- the original data is vibration data during manufacturing of the products A, B, and C acquired on the floor a.
- vibration data assumed to have been acquired on the floor b is generated by data augmentation. Specifically, noise is added assuming that there are people on floor b, and vibrations generated from machines that are on floor b but not on floor a are added as noise.
- FIG. 14 is a diagram showing an example of a machine learning data generation method when the machine learning method is supervised learning and the data type is text.
- FIG. 14 shows an example of generating training data for a learning model that classifies texts regarding product evaluations posted on various sites on the Internet according to satisfaction levels.
- the original data are reviews posted on shopping sites and review sites, and review articles on news sites. Data augmentation is performed on these original data to convert the writing style into a colloquial style. Specifically, it performs processing such as dividing sentences into shorter sentences, changing the endings of words to look like spoken words, and adding exclamation points to the beginnings of words. As a result, data is generated on the assumption that it will be written on a bulletin board or posted on an SNS. By generating training data in this way, generalized machine learning can be achieved even for reviews that contain a lot of colloquialism.
- FIG. 16 is a diagram showing an example of a machine learning data generation method when the machine learning method is supervised learning and the data type is images.
- Visual Slam is used in automated guided vehicles (AGV) and autonomous mobile robots (AMR), and is a technology that simultaneously estimates its own position from images captured by a camera and creates an environmental map.
- the original data is a set of an image captured by the camera and the coordinates and point cloud estimated by the robot, and the information on the coordinates and point cloud is the teacher label.
- data augmentation is performed to increase environmental variations such as weather and time of day for a domain (sunny/daytime) that includes a plurality of image data shot in sunny daytime and teacher labels.
- Data conversion is performed only on the input data (image data) and not on the combined teacher labels.
- data sets of multiple domains for example, sunny/evening, snowy/daytime, etc.
- FIG. 17 is a diagram showing an example of a machine learning data generation method when the machine learning method is unsupervised learning and the data type is images.
- the original data is an inspection image obtained by an inspection machine. It is assumed that defective products are rare in products and that most of the inspection images used for learning are good products.
- Data augmentation is performed on these images to increase the variation of shooting conditions such as cameras and lights. As a result, data sets of multiple domains (for example, rotation, enlargement, color change, etc.) extended to various shooting conditions are generated.
- FIG. 18 is a diagram showing various types of training data to which the machine learning data generation method can be applied and an example of data augmentation when using each data when the machine learning method is reinforcement learning.
- training data original data
- environment acquisition method For example, if the task is to grasp an object by a robot arm (agent), an image of the grasping environment (for example, an image including the object itself and the floor on which the object is placed) is the original data. Also, if the task is automatic driving, for example, the image of the road that can be seen from the driver's seat becomes the original data.
- the target part of the original data is "the part that affects the acquisition of rewards"
- the non-target part is the "data acquisition environment” that does not affect the acquisition of rewards.
- the target part includes not only objects such as objects that are the direct target of the task, but also elements that affect the reward (for example, in the case of grasping by a robot arm, the material and angle of the floor on which the object is placed) ) are also included.
- the non-target portion includes the brightness of the room, the color of the floor, and the like.
- FIG. 19 is a diagram showing an example of a machine learning data generation method when the machine learning method is reinforcement learning and the task is gripping a robot arm.
- the original data is an image of a space including a grasped object (for example, a product) and a table on which the grasped object is placed when the robot arm performs a grasping operation on the simulator.
- conditions other than elements that affect the success or failure of gripping e.g. product shape, material, weight, stand angle, surface material, etc.
- shooting conditions such as camera, light, etc. , base color, etc.
- a data set extended to various shooting conditions for example, brightness, etc.), color, etc. is generated.
- FIG. 20 is a diagram showing an example of a machine learning data generation method when the machine learning method is reinforcement learning and the task is Embodied Question Answering (EQA) (reference document 3).
- EQA gives a question to an agent
- the agent moves to the necessary place to obtain the information necessary to answer the question and outputs the answer.
- the agent searches for a position where the car can be seen.
- the original data is an image of the field of view seen from the robot.
- FIG. 20 also assumes that learning is performed using a physical simulator.
- the object itself here, a car
- the factors that affect the movement of the agent for example, the angle of the floor surface, Friction, etc.
- the brightness and color of the space are non-objective. Therefore, data augmentation is performed to increase variations in brightness (eg, time of day, etc.) and space color (eg, wall color, etc.) to generate a data set in which these conditions are expanded.
- data augmentation is performed using training data for meta-learning as source data, and a data set including both the source data and data generated by data augmentation was extracted as a dataset for meta-learning.
- learning data capable of preventing over-learning and improving the generalization ability of meta-learning.
- the target part of the original data is not changed, and data augmentation is performed by changing the non-target part other than the target part.
- various training data effective for generalization learning can be generated without changing parts that should remain unchanged before and after data augmentation.
- a data generation method for generating data for domain generalization in machine learning comprising: a computer (10) performing augmentation using learning data used for training a machine learning model as source data; a computer (10) extracting a data set containing both the original data and the data generated by the data augmentation as a data set for the domain generalization; data generation for machine learning.
- Appendix 2 The original data and the data generated by the data augmentation are each held in units of domains, In the extracting step, The data generation for machine learning according to appendix 1, wherein at least one domain of the original data and at least one domain of the data generated by the data augmentation are extracted as a data set for the domain generalization. Method.
- the original data includes a target portion and a non-target portion, In the step of performing the augmentation, 3.
- Appendix 4 the original data is image data, the target portion is an image of an object, In the step of performing the augmentation, 3.
- the original data is signal data
- the target part is a specific signal pattern
- the data generation method for machine learning according to appendix 3 wherein data augmentation for changing noise is performed in the signal data of the original data.
- Appendix 12 In the step of performing the augmentation, 12.
- the original data is data on the state of the environment in which the agent is placed in reinforcement learning
- the target part is information on the part that affects the acquisition of rewards
- (Appendix 15) Meta-learning including the step of performing domain generalization by meta-learning using the data set for domain generalization generated by the machine learning data generation method according to any one of Appendices 1 to 14 Method.
- Appendix 17 A domain generalization learning method including the step of performing domain generalization learning using the data set for domain generalization generated by the machine learning data generation method according to any one of Appendices 1 to 14. .
- a data generator (10) for generating data for domain generalization in machine learning a data generation unit (21) that performs data augmentation using learning data used for training a machine learning model as source data; a training data extraction unit (22) for extracting a data set containing both the original data and the data generated by the data augmentation as a data set for the domain generalization; A device (10).
- (Appendix 19) a computer (10) that generates data for domain generalization in machine learning; a data generation unit (21) that performs data augmentation using learning data used for training a machine learning model as source data; A program that functions as a training data extraction unit (22) that extracts a data set containing both the original data and the data generated by the data augmentation as a data set for the domain generalization.
- Machine learning data generator 11 For Processor 12
- Main memory 13 For Camera interface 14
- Input/output interface 15 For Display interface 16
- Data generator 22 For Training data Extraction unit 23
- Learning execution unit 31 For Original data storage unit 32
- Generated data storage unit 51 For Camera 52
- Input device 53 For Output device 54... Display device
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Computing Systems (AREA)
- Evolutionary Computation (AREA)
- Mathematical Physics (AREA)
- Medical Informatics (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Computational Linguistics (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
- Feedback Control In General (AREA)
- Machine Translation (AREA)
Abstract
Description
データオーギュメンテーションは、機械学習における学習プロセスの学習であるメタ学習の分野においても有効な技術である。例えば特許文献1には、メタ学習ネットワークの訓練方法について記載されている。
本発明の一側面に係る機械学習用データ生成方法は、機械学習におけるドメイン汎化のためのデータを生成するデータ生成方法であって、コンピュータが、機械学習モデルの訓練に用いる学習用データを元データとしてオーギュメンテーションを実施する工程と、コンピュータが、前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する工程と、を含むものである。
上記構成により、学習工程(学習ループ内)で元データとオーギュメンテーションにより生成されたデータの両方を必ず含むデータセットを抽出し、これを用いてメタ学習を実施することができ、過学習を防いで、メタ学習によるドメイン汎化の汎化能力を向上させることができる。
上記構成により、ドメイン汎化学習の前段で元データとオーギュメンテーションにより生成されたデータの両方を必ず含むデータセットを抽出し、これを用いてドメイン汎化学習を実施することができ、過学習を防いで、ドメイン汎化学習の汎化能力を向上させることができる。
上記構成により、コストをかけずに汎化学習のための新たなデータを生成し、汎化学習における汎化性能を向上させることができる。
上記構成により、コストをかけずに汎化学習のための新たなデータを生成し、汎化学習における汎化性能を向上させることができる。
図1を用いて、本発明が適用される場面の一例について説明する。図1は、本発明による機械学習用データ生成装置10による機械学習用データ生成処理の概要を説明する図である。図1の例では、ある地点Aに設置されたカメラで撮影した画像データをオリジナルデータ(元データ)としてデータオーギュメンテーションを実施し、訓練データを生成している。オリジナルデータは、雨天時に撮影した画像データ、降雪時に撮影した画像データ、昼間に撮影した画像データ、夜間に撮影した画像データの4つのドメインのデータセットに分けて保存されている。図1の例では、これらの各ドメインのオリジナルデータに対して画像を反転させる処理を実施し、オーギュメンテーション後の各データを、雨天時のデータ、降雪時のデータ、昼間のデータ、および夜間のデータのデータセットに分けて保存している。なお、データオーギュメンテーションは、データ拡張と呼ばれることもある。
(1.ハードウェア構成)
次に、図2を参照しながら、本実施形態に関わる機械学習用データ生成装置10のハードウェア構成の一例について説明する。
機械学習用データ生成装置10は、そのハードウェア資源として、プロセッサ11と、メインメモリ12と、カメラインタフェース13と、入出力インタフェース14と、ディスプレイインタフェース15と、通信インタフェース16と、記憶装置17とを備えるコンピュータシステムである。
次に、図3を用いて、本開示の実施形態に係る機械学習用データ生成装置10の機能構成の一例を説明する。図3は、プロセッサ11によって実行される機能モジュールを示すブロック図である。図3に示すように、機能モジュールには、データ生成部21、訓練データ抽出部22、学習実行部23が含まれる。
次に、本実施形態に係る機械学習用データ生成装置10による機械学習用データ生成方法について、図4~8を用いて説明する。なお、ここでは、元データとなる画像データにデータオーギュメンテーションを実施し、新たな画像データを生成する例について説明する。
[参考文献1]“Learning to Generalize: Meta-Learning for Domain Generalization”, Da Li et al., URL:https://arxiv.org/abs/1710.03463
ステップS1054において、学習実行部23は、ステップS1053で更新したネットワークパラメータを初期値として、データセット2に対して損失(loss2)を算出する。
[参考文献2]“Multi-Task Adversarial Network for Disentangled Feature Learning”, Yang Liu et al.,
URL:https://openaccess.thecvf.com/content_cvpr_2018/html/Liu_Multi-Task_Adversarial_Network_CVPR_2018_paper.html
ステップS202においては、図4のステップS102と同様にデータオーギュメンテーションの方法を受け付けるが、図5の例では、複数の方法(n通り)を同時に受け付けることができる。
[参考文献3]“Embodied Question Answering”, Abhishek Das et al.,
URL:https://arxiv.org/abs/1711.11543
機械学習におけるドメイン汎化のためのデータを生成するデータ生成方法であって、
コンピュータ(10)が、機械学習モデルの訓練に用いる学習用データを元データとしてオーギュメンテーションを実施する工程と、
コンピュータ(10)が、前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する工程と、を含む機械学習用データ生成方法。
前記元データと前記データオーギュメンテーションにより生成されたデータは、それぞれドメイン単位で保持されており、
前記抽出する工程では、
前記元データの少なくとも1つのドメインと、前記データオーギュメンテーションにより生成されたデータの少なくとも1つのドメインを、前記ドメイン汎化のためのデータセットとして抽出する、付記1に記載の機械学習用データ生成方法。
前記元データは、対象部分と非対象部分を含み、
前記オーギュメンテーションを実施する工程では、
前記元データにおける前記非対象部分を変化させるデータオーギュメンテーションを実施する、付記1または2に記載の機械学習用データ生成方法。
前記元データは画像データであり、前記対象部分は対象物の画像であり、
前記オーギュメンテーションを実施する工程では、
前記元データの画像における、前記対象物の周囲の環境及び/又は前記対象物の撮像条件を変化させるデータオーギュメンテーションを実施する、付記3に記載の機械学習用データ生成方法。
前記オーギュメンテーションを実施する工程では、
前記元データの画像の明度の変更、背景の変更、及び色調の変更のうちの少なくとも1つを実施することにより、前記対象物の周囲の環境を変化させるデータオーギュメンテーションを実施する、付記4に記載の機械学習用データ生成方法。
前記オーギュメンテーションを実施する工程では、
前記元データの画像の回転、反転、拡大、縮小、移動、トリミング、及びフィルタ処理のうちの少なくとも1つを実施することにより、前記対象物の撮像条件を変化させるデータオーギュメンテーションを実施する、付記4に記載の機械学習用データ生成方法。
前記元データは音声データであり、前記対象部分は特定の音声であり、
前記オーギュメンテーションを実施する工程では、
前記元データの音声データにおける、環境音やノイズを変化させるデータオーギュメンテーションを実施する、付記3に記載の機械学習用データ生成方法。
前記オーギュメンテーションを実施する工程では、
前記元データの音声データに環境音の音声を合成することによりデータオーギュメンテーションを実施する、付記7に記載の機械学習用データ生成方法。
前記元データは信号データであり、前記対象部分は特定の信号パターンであり、
前記オーギュメンテーションを実施する工程では、
前記元データの信号データにおける、ノイズを変化させるデータオーギュメンテーションを実施する、付記3に記載の機械学習用データ生成方法。
前記オーギュメンテーションを実施する工程では、
前記元データの信号データにノイズを合成することによりデータオーギュメンテーションを実施する、付記9に記載の機械学習用データ生成方法。
前記元データはテキストデータであり、前記対象部分は特定のテキストパターンであり、
前記オーギュメンテーションを実施する工程では、
前記元データのテキストデータにおける、文体を変化させるデータオーギュメンテーションを実施する、付記3に記載の機械学習用データ生成方法。
前記オーギュメンテーションを実施する工程では、
前記元データのテキストデータの語頭及び/又は語尾を変えることによりデータオーギュメンテーションを実施する、付記11に記載の機械学習用データ生成方法。
前記抽出する工程では、
前記元データのドメインと前記データオーギュメンテーションにより生成されたデータのドメインが、所定の割合で含まれるように抽出する、付記2に記載の機械学習用データ生成方法。
前記元データは強化学習におけるエージェントが置かれた環境の状態に関するデータであり、前記対象部分は報酬の獲得に影響を与える部分の情報であり、
前記オーギュメンテーションを実施する工程では、
前記元データにおける、前記環境の状態のうち、前記報酬の獲得に影響を与えない部分の条件を変化させるデータオーギュメンテーションを実施する、付記3に記載の機械学習用データ生成方法。
付記1から14のいずれか1項に記載の機械学習用データ生成方法によって生成された前記ドメイン汎化のためのデータセットを用いて、メタ学習によるドメイン汎化を実施する工程を含む、メタ学習方法。
前記メタ学習によるドメイン汎化を実施する工程では、
前記元データの少なくとも1つのドメインと、前記データオーギュメンテーションにより生成されたデータの少なくとも1つのドメインとを含む複数のデータセットの集合を用いて、メタ学習によるドメイン汎化を実施する、付記15に記載のメタ学習方法。
付記1から14のいずれか1項に記載の機械学習用データ生成方法によって生成された前記ドメイン汎化のためのデータセットを用いて、ドメイン汎化学習を実施する工程を含むドメイン汎化学習方法。
機械学習におけるドメイン汎化のためのデータを生成するデータ生成装置(10)であって、
機械学習モデルの訓練に用いる学習用データを元データとしてデータオーギュメンテーションを実施するデータ生成部(21)と、
前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する訓練データ抽出部(22)とを備えた機械学習用データ生成装置(10)。
機械学習におけるドメイン汎化のためのデータを生成するコンピュータ(10)を、
機械学習モデルの訓練に用いる学習用データを元データとしてデータオーギュメンテーションを実施するデータ生成部(21)と、
前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する訓練データ抽出部(22)、として機能させるプログラム。
Claims (19)
- 機械学習におけるドメイン汎化のためのデータを生成するデータ生成方法であって、
コンピュータが、機械学習モデルの訓練に用いる学習用データを元データとしてオーギュメンテーションを実施する工程と、
コンピュータが、前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する工程と、を含む機械学習用データ生成方法。 - 前記元データと前記データオーギュメンテーションにより生成されたデータは、それぞれドメイン単位で保持されており、
前記抽出する工程では、
前記元データの少なくとも1つのドメインと、前記データオーギュメンテーションにより生成されたデータの少なくとも1つのドメインを、前記ドメイン汎化のためのデータセットとして抽出する、請求項1に記載の機械学習用データ生成方法。 - 前記元データは、対象部分と非対象部分を含み、
前記オーギュメンテーションを実施する工程では、
前記元データにおける前記非対象部分を変化させるデータオーギュメンテーションを実施する、請求項1または2に記載の機械学習用データ生成方法。 - 前記元データは画像データであり、前記対象部分は対象物の画像であり、
前記オーギュメンテーションを実施する工程では、
前記元データの画像における、前記対象物の周囲の環境及び/又は前記対象物の撮像条件を変化させるデータオーギュメンテーションを実施する、請求項3に記載の機械学習用データ生成方法。 - 前記オーギュメンテーションを実施する工程では、
前記元データの画像の明度の変更、背景の変更、及び色調の変更のうちの少なくとも1つを実施することにより、前記対象物の周囲の環境を変化させるデータオーギュメンテーションを実施する、請求項4に記載の機械学習用データ生成方法。 - 前記オーギュメンテーションを実施する工程では、
前記元データの画像の回転、反転、拡大、縮小、移動、トリミング、及びフィルタ処理のうちの少なくとも1つを実施することにより、前記対象物の撮像条件を変化させるデータオーギュメンテーションを実施する、請求項4に記載の機械学習用データ生成方法。 - 前記元データは音声データであり、前記対象部分は特定の音声であり、
前記オーギュメンテーションを実施する工程では、
前記元データの音声データにおける、環境音やノイズを変化させるデータオーギュメンテーションを実施する、請求項3に記載の機械学習用データ生成方法。 - 前記オーギュメンテーションを実施する工程では、
前記元データの音声データに環境音の音声を合成することによりデータオーギュメンテーションを実施する、請求項7に記載の機械学習用データ生成方法。 - 前記元データは信号データであり、前記対象部分は特定の信号パターンであり、
前記オーギュメンテーションを実施する工程では、
前記元データの信号データにおける、ノイズを変化させるデータオーギュメンテーションを実施する、請求項3に記載の機械学習用データ生成方法。 - 前記オーギュメンテーションを実施する工程では、
前記元データの信号データにノイズを合成することによりデータオーギュメンテーションを実施する、請求項9に記載の機械学習用データ生成方法。 - 前記元データはテキストデータであり、前記対象部分は特定のテキストパターンであり、
前記オーギュメンテーションを実施する工程では、
前記元データのテキストデータにおける、文体を変化させるデータオーギュメンテーションを実施する、請求項3に記載の機械学習用データ生成方法。 - 前記オーギュメンテーションを実施する工程では、
前記元データのテキストデータの語頭及び/又は語尾を変えることによりデータオーギュメンテーションを実施する、請求項11に記載の機械学習用データ生成方法。 - 前記抽出する工程では、
前記元データのドメインと前記データオーギュメンテーションにより生成されたデータのドメインが、所定の割合で含まれるように抽出する、請求項2に記載の機械学習用データ生成方法。 - 前記元データは強化学習におけるエージェントが置かれた環境の状態に関するデータであり、前記対象部分は報酬の獲得に影響を与える部分の情報であり、
前記オーギュメンテーションを実施する工程では、
前記元データにおける、前記環境の状態のうち、前記報酬の獲得に影響を与えない部分の条件を変化させるデータオーギュメンテーションを実施する、請求項3に記載の機械学習用データ生成方法。 - 請求項1から14のいずれか1項に記載の機械学習用データ生成方法によって生成された前記ドメイン汎化のためのデータセットを用いて、メタ学習によるドメイン汎化を実施する工程を含む、メタ学習方法。
- 前記メタ学習によるドメイン汎化を実施する工程では、
前記元データの少なくとも1つのドメインと、前記データオーギュメンテーションにより生成されたデータの少なくとも1つのドメインとを含む複数のデータセットの集合を用いて、メタ学習によるドメイン汎化を実施する、請求項15に記載のメタ学習方法。 - 請求項1から14のいずれか1項に記載の機械学習用データ生成方法によって生成された前記ドメイン汎化のためのデータセットを用いて、ドメイン汎化学習を実施する工程を含む、ドメイン汎化学習方法。
- 機械学習におけるドメイン汎化のためのデータを生成するデータ生成装置であって、
機械学習モデルの訓練に用いる学習用データを元データとしてデータオーギュメンテーションを実施するデータ生成部と、
前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する訓練データ抽出部とを備えた機械学習用データ生成装置。 - 機械学習におけるドメイン汎化のためのデータを生成するコンピュータを、
機械学習モデルの訓練に用いる学習用データを元データとしてデータオーギュメンテーションを実施するデータ生成部と、
前記元データと、前記データオーギュメンテーションにより生成されたデータの両方を含むデータセットを前記ドメイン汎化のためのデータセットとして抽出する訓練データ抽出部、として機能させるプログラム。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
US18/264,038 US20240232704A9 (en) | 2021-02-16 | 2022-01-18 | Machine learning data generation method, meta-learning method, machine learning data generation apparatus, and program |
EP22755806.1A EP4296911A1 (en) | 2021-02-16 | 2022-01-18 | Machine learning data generation method, meta learning method, machine learning data generation device, and program |
CN202280011672.5A CN116964600A (zh) | 2021-02-16 | 2022-01-18 | 机器学习用数据生成方法、元学习方法、机器学习用数据生成装置及程序 |
Applications Claiming Priority (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2021022715 | 2021-02-16 | ||
JP2021-022715 | 2021-02-16 | ||
JP2021122787A JP2022124989A (ja) | 2021-02-16 | 2021-07-27 | 機械学習用データ生成方法、メタ学習方法、機械学習用データ生成装置及びプログラム |
JP2021-122787 | 2021-07-27 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2022176481A1 true WO2022176481A1 (ja) | 2022-08-25 |
Family
ID=82930850
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2022/001626 WO2022176481A1 (ja) | 2021-02-16 | 2022-01-18 | 機械学習用データ生成方法、メタ学習方法、機械学習用データ生成装置及びプログラム |
Country Status (3)
Country | Link |
---|---|
EP (1) | EP4296911A1 (ja) |
TW (1) | TWI800234B (ja) |
WO (1) | WO2022176481A1 (ja) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019116643A1 (ja) * | 2017-12-12 | 2019-06-20 | ソニー株式会社 | 情報処理装置および情報処理方法 |
JP2020144849A (ja) | 2019-03-04 | 2020-09-10 | 富士通株式会社 | メタ学習ネットワークを訓練する装置及び方法 |
JP2021022715A (ja) | 2019-07-30 | 2021-02-18 | パナソニックIpマネジメント株式会社 | 回路基板、電源装置及び照明器具 |
JP2021122787A (ja) | 2020-02-05 | 2021-08-30 | 株式会社フジタ | 吸着システム、および吸着方法 |
Family Cites Families (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019016643A1 (en) * | 2017-07-20 | 2019-01-24 | Adari Swarna Kumari | ELECTRONIC NEGOTIABLE INSTRUMENT GENERATING SYSTEM AND METHOD |
US20200335086A1 (en) * | 2019-04-19 | 2020-10-22 | Behavioral Signal Technologies, Inc. | Speech data augmentation |
-
2022
- 2022-01-18 WO PCT/JP2022/001626 patent/WO2022176481A1/ja active Application Filing
- 2022-01-18 EP EP22755806.1A patent/EP4296911A1/en active Pending
- 2022-01-24 TW TW111102966A patent/TWI800234B/zh active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
WO2019116643A1 (ja) * | 2017-12-12 | 2019-06-20 | ソニー株式会社 | 情報処理装置および情報処理方法 |
JP2020144849A (ja) | 2019-03-04 | 2020-09-10 | 富士通株式会社 | メタ学習ネットワークを訓練する装置及び方法 |
JP2021022715A (ja) | 2019-07-30 | 2021-02-18 | パナソニックIpマネジメント株式会社 | 回路基板、電源装置及び照明器具 |
JP2021122787A (ja) | 2020-02-05 | 2021-08-30 | 株式会社フジタ | 吸着システム、および吸着方法 |
Non-Patent Citations (10)
Title |
---|
CAGLI ELEONORA; DUMAS CéCILE; PROUFF EMMANUEL: "Convolutional Neural Networks with Data Augmentation Against Jitter-Based Countermeasures", SAT 2015 18TH INTERNATIONAL CONFERENCE, AUSTIN, TX, USA, SEPTEMBER 24-27, 2015, vol. 10529 Chap.3, no. 558, 25 August 2017 (2017-08-25), Berlin, Heidelberg , pages 45 - 68, XP047446632, ISBN: 3540745491, DOI: 10.1007/978-3-319-66787-4_3 * |
DA LI ET AL., LEARNING TO GENERALIZE: META-LEARNING FOR DOMAIN GENERALIZATION, Retrieved from the Internet <URL:https://arxiv.org/abs/1710.03463> |
DA LI; YONGXIN YANG; YI-ZHE SONG; TIMOTHY M. HOSPEDALES: "Learning to Generalize: Meta-Learning for Domain Generalization", ARXIV.ORG, 10 October 2017 (2017-10-10), 201 Olin Library Cornell University Ithaca, NY 14853 , XP080827512 * |
DWIBEDI DEBIDATTA; MISRA ISHAN; HEBERT MARTIAL: "Cut, Paste and Learn: Surprisingly Easy Synthesis for Instance Detection", 2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER VISION (ICCV), 22 October 2017 (2017-10-22), pages 1310 - 1319, XP033282989, DOI: 10.1109/ICCV.2017.146 * |
ICHIKAWA OSAMU; FUKUDA TAKASHI; SUZUKI MASAYUKI; KURATA GAKUTO; RAMABHADRAN BHUVANA: "Harmonic feature fusion for robust neural network-based acoustic modeling", 2017 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP), 5 March 2017 (2017-03-05), pages 5195 - 5199, XP033259401, DOI: 10.1109/ICASSP.2017.7953147 * |
ICHIKAWA, OSAMU ET AL.: "Neural network-based acoustic model integrating harmonic information for robust speech recognition", PROCEEDINGS OF 2017 SPRING MEETING OF ACOUSTICAL SOCIETY OF JAPAN; 2017.03.15-17, 1 January 2017 (2017-01-01) - 17 March 2017 (2017-03-17), JP , pages 9 - 12, XP009539315, ISSN: 1880-7658 * |
KEN MIYAMOTO, OSAMU TSUKAHARA: "Data Augmentation based on RGB-D SLAM 3D Data", IEICE TECHNICAL REPORT, PRMU, vol. 118, no. 460 (PRMU2018-124), 1 February 2019 (2019-02-01), JP , pages 53 - 57, XP009539311, ISSN: 2432-6380 * |
OUCHI TOMOHITO: "Research on data extensions in pointer-generator models", PROCEEDINGS OF THE TWENTY-SIXTH ANNUAL MEETING OF THE ASSOCIATION FOR NATURAL LANGUAGE PROCESSING., 9 March 2020 (2020-03-09), pages 941 - 944, XP055961724 * |
TOMOKI YONEZAWA, SHINTARO ONO, YOSHIHIRO SUDA: "A Study of Vehicle Detection at Night by Data Augmentation using Image Style Transfer", IEICE TECHNICAL REPORT, ITS, vol. 119, no. 102 ( ITS2019-1), 28 June 2019 (2019-06-28), JP , pages 1 - 6, XP009539314, ISSN: 0913-5685 * |
YANG LIU ET AL., MULTI-TASK ADVERSARIAL NETWORK FOR DISENTANGLED FEATURE LEARNING, Retrieved from the Internet <URL:https://openaccess.thecvf.com/contentcvpr2018/html/Liu_Multi-Task_Adversarial_Network-CVPR_2018_paper.htmI> |
Also Published As
Publication number | Publication date |
---|---|
TWI800234B (zh) | 2023-04-21 |
US20240135246A1 (en) | 2024-04-25 |
TW202234296A (zh) | 2022-09-01 |
EP4296911A1 (en) | 2023-12-27 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111259779A (zh) | 一种基于中心点轨迹预测的视频动作检测方法 | |
CN110853656B (zh) | 基于改进神经网络的音频篡改识别方法 | |
CN114333070A (zh) | 一种基于深度学习的考生异常行为检测方法 | |
CN111581470B (zh) | 用于对话系统情景匹配的多模态融合学习分析方法和系统 | |
Kong et al. | Acoustic scene generation with conditional SampleRNN | |
KR102148378B1 (ko) | 기계 학습 모델을 이용한 관심 이벤트의 알림 제공 장치 및 방법 | |
CN111460876A (zh) | 用于识别视频的方法和装置 | |
CN113283336A (zh) | 一种文本识别方法与系统 | |
CN114550110A (zh) | 一种基于无监督域适应的车辆重识别方法及系统 | |
CN115512191A (zh) | 一种联合问答的图像自然语言描述方法 | |
Ivanko et al. | Speaker-dependent visual command recognition in vehicle cabin: methodology and evaluation | |
Tan et al. | Attention augmented convolutional neural network for acoustics based machine state estimation | |
WO2022176481A1 (ja) | 機械学習用データ生成方法、メタ学習方法、機械学習用データ生成装置及びプログラム | |
JP2022124989A (ja) | 機械学習用データ生成方法、メタ学習方法、機械学習用データ生成装置及びプログラム | |
KR102096598B1 (ko) | 애니메이션 생성 방법 | |
CN117115310A (zh) | 一种基于音频和图像的数字人脸生成方法及系统 | |
CN116128866A (zh) | 基于USRNet与改进MobileNet-SSD算法的输电线路绝缘子故障检测方法 | |
US20240232704A9 (en) | Machine learning data generation method, meta-learning method, machine learning data generation apparatus, and program | |
Salam et al. | You Only Look Once (YOLOv3): Object Detection and Recognition for Indoor Environment | |
CN115273224A (zh) | 一种基于高低分辨率双模态蒸馏的视频人体行为识别方法 | |
CN116964600A (zh) | 机器学习用数据生成方法、元学习方法、机器学习用数据生成装置及程序 | |
JP7516900B2 (ja) | 音声イベントの認識装置及び方法 | |
CN110472528B (zh) | 一种地铁环境目标训练集生成方法及系统 | |
KR102667003B1 (ko) | 콘텐츠 정보를 이용한 이미지 생성 모델 학습 방법 및 시스템 | |
Kousalya et al. | Group Emotion Detection using Convolutional Neural Network |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22755806 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202280011672.5 Country of ref document: CN |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18264038 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022755806 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
ENP | Entry into the national phase |
Ref document number: 2022755806 Country of ref document: EP Effective date: 20230918 |