WO2022244059A1 - 情報処理システム、情報処理方法、及び記録媒体 - Google Patents

情報処理システム、情報処理方法、及び記録媒体 Download PDF

Info

Publication number
WO2022244059A1
WO2022244059A1 PCT/JP2021/018619 JP2021018619W WO2022244059A1 WO 2022244059 A1 WO2022244059 A1 WO 2022244059A1 JP 2021018619 W JP2021018619 W JP 2021018619W WO 2022244059 A1 WO2022244059 A1 WO 2022244059A1
Authority
WO
WIPO (PCT)
Prior art keywords
pseudo
label
evaluation
data
model
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/JP2021/018619
Other languages
English (en)
French (fr)
Japanese (ja)
Inventor
亮 高本
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
NEC Corp
Original Assignee
NEC Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by NEC Corp filed Critical NEC Corp
Priority to JP2023522008A priority Critical patent/JP7552890B2/ja
Priority to PCT/JP2021/018619 priority patent/WO2022244059A1/ja
Priority to US18/290,346 priority patent/US20240289633A1/en
Publication of WO2022244059A1 publication Critical patent/WO2022244059A1/ja
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/096—Transfer learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks

Definitions

  • This disclosure relates to the technical fields of information processing systems, information processing methods, and recording media.
  • Patent Literature 1 discloses a technique of assigning a positive label or a negative label to a feature vector of unlabeled pixels.
  • Patent Literature 2 discloses a technique of generating training data with pseudo-labels by calculating a score using a ratio of labels with a short distance, etc., and setting the score to unlabeled data.
  • Patent Document 3 discloses a technology for learning a conversion rule from the source domain to the target domain using labeled data of the source domain and unlabeled data of the target domain.
  • the purpose of this disclosure is to improve the technology disclosed in prior art documents.
  • One aspect of the information processing system disclosed herein uses a data input means for inputting labeled data and unlabeled data, and a teacher model trained using the labeled data to apply a pseudo-label to the unlabeled data. and an evaluation model trained using at least one of the labeled data and the unlabeled data to evaluate the pseudo-label given to the unlabeled data, and a predetermined
  • a student model is learned using pseudo-label evaluation means for outputting the pseudo-label that meets the evaluation criteria as an evaluation pseudo-label, and pseudo-label data obtained by adding the evaluation pseudo-label to the labeled data and the unlabeled data. and model output means for outputting the learned student model.
  • One aspect of the information processing method of this disclosure is to input labeled data and unlabeled data, use a teacher model trained using the labeled data, give a pseudo label to the unlabeled data, The pseudo-label given to the unlabeled data is evaluated using an evaluation model trained using at least one of the labeled data and the unlabeled data, and the pseudo-label that meets a predetermined evaluation criterion. is output as an evaluation pseudo-label, a student model is learned using the pseudo-label data obtained by assigning the evaluation pseudo-label to the labeled data and the unlabeled data, and the learned student model is output.
  • labeled data and unlabeled data are input to a computer, and a pseudo-label is given to the unlabeled data using a teacher model trained using the labeled data. and using an evaluation model learned using at least one of the labeled data and the unlabeled data, the pseudo-label given to the unlabeled data is evaluated, and the method that satisfies a predetermined evaluation criterion is evaluated.
  • a computer program is recorded that causes the processing method to be performed.
  • FIG. 2 is a block diagram showing the hardware configuration of the information processing system according to the first embodiment
  • FIG. 1 is a block diagram showing a functional configuration of an information processing system according to a first embodiment
  • FIG. 4 is a flowchart showing the flow of operations of the information processing system according to the first embodiment
  • It is a block diagram which shows the functional structure of the information processing system which concerns on 2nd Embodiment.
  • 9 is a flowchart showing the flow of operations of an information processing system according to the second embodiment
  • FIG. 11 is a block diagram showing a method of learning an evaluation model in an information processing system according to a third embodiment
  • FIG. 14 is a block diagram showing a method of learning an evaluation model in an information processing system according to a fourth embodiment
  • FIG. 14 is a block diagram showing a method of learning an evaluation model in an information processing system according to a fifth embodiment;
  • FIG. 14 is a block diagram showing a method of learning an evaluation model in an information processing system according to a sixth embodiment;
  • FIG. 21 is a block diagram showing the configuration of a model evaluation unit in an information processing system according to a seventh embodiment;
  • FIG. FIG. 22 is a block diagram showing a functional configuration of an information processing system according to an eighth embodiment;
  • FIG. FIG. 21 is a flow chart showing the flow of operations of an information processing system according to the eighth embodiment;
  • FIG. FIG. 22 is a block diagram showing a functional configuration of an information processing system according to a ninth embodiment;
  • FIG. FIG. 21 is a flow chart showing the flow of operations of an information processing system according to a ninth embodiment;
  • FIG. 1 An information processing system according to the first embodiment will be described with reference to FIGS. 1 to 3.
  • FIG. 1 An information processing system according to the first embodiment will be described with reference to FIGS. 1 to 3.
  • FIG. 1 An information processing system according to the first embodiment will be described with reference to FIGS. 1 to 3.
  • FIG. 1 An information processing system according to the first embodiment will be described with reference to FIGS. 1 to 3.
  • FIG. 1 An information processing system according to the first embodiment will be described with reference to FIGS. 1 to 3.
  • FIG. 1 is a block diagram showing the hardware configuration of an information processing system according to the first embodiment.
  • an information processing system 10 includes a processor 11, a RAM (Random Access Memory) 12, a ROM (Read Only Memory) 13, and a storage device .
  • Information processing system 10 may further include an input device 15 and an output device 16 .
  • Processor 11 , RAM 12 , ROM 13 , storage device 14 , input device 15 and output device 16 are connected via data bus 17 .
  • the processor 11 reads a computer program.
  • processor 11 is configured to read a computer program stored in at least one of RAM 12, ROM 13 and storage device .
  • the processor 11 may read a computer program stored in a computer-readable recording medium using a recording medium reader (not shown).
  • the processor 11 may acquire (that is, read) a computer program from a device (not shown) arranged outside the information processing system 10 via a network interface.
  • the processor 11 controls the RAM 12, the storage device 14, the input device 15 and the output device 16 by executing the read computer program.
  • the processor 11 implements functional blocks for executing processing related to machine learning.
  • the processor 11 may be configured as, for example, a CPU (Central Processing Unit), GPU (Graphics Processing Unit), FPGA (Field-Programmable Gate Array), DSP (Demand-Side Platform), and ASIC (Application Specific Integrate).
  • the processor 11 may be configured with one of these, or may be configured to use a plurality of them in parallel.
  • the RAM 12 temporarily stores computer programs executed by the processor 11.
  • the RAM 12 temporarily stores data temporarily used by the processor 11 while the processor 11 is executing the computer program.
  • the RAM 12 may be, for example, a D-RAM (Dynamic RAM).
  • the ROM 13 stores computer programs executed by the processor 11 .
  • the ROM 13 may also store other fixed data.
  • the ROM 13 may be, for example, a P-ROM (Programmable ROM).
  • the storage device 14 stores data that the information processing system 10 saves for a long period of time.
  • Storage device 14 may act as a temporary storage device for processor 11 .
  • the storage device 14 may include, for example, at least one of a hard disk device, a magneto-optical disk device, an SSD (Solid State Drive), and a disk array device.
  • the input device 15 is a device that receives input instructions from the user of the information processing system 10 .
  • Input device 15 may include, for example, at least one of a keyboard, mouse, and touch panel.
  • the output device 16 is a device that outputs information about the information processing system 10 to the outside.
  • the output device 16 may be a display device (eg, display) capable of displaying information regarding the information processing system 10 .
  • FIG. 2 is a block diagram showing the functional configuration of the information processing system according to the first embodiment.
  • the information processing system 10 includes a labeled data input unit 110, an unlabeled data input unit 120, and a teacher model learning unit as processing blocks for realizing its functions. 130 , a pseudo-label generation unit 140 , a pseudo-label evaluation unit 150 , a student model learning unit 160 and a model output unit 170 .
  • Each of the labeled data input unit 110, the unlabeled data input unit 120, the teacher model learning unit 130, the pseudo label generating unit 140, the pseudo label evaluating unit 150, the student model learning unit 160, and the model output unit 170 for example, It may be implemented by processor 11 (see FIG. 1).
  • the labeled data input unit 110 is configured so that labeled data can be input.
  • the unlabeled data input section 120 is configured to be able to input unlabeled data.
  • the “label” is information indicating the correct answer assigned to the data (so-called correct answer label). is not assigned.
  • An example of labeled data and unlabeled data is image data.
  • the image data may include the eye area and face area of the living body.
  • the image data may be data including a plurality of consecutive images in time series (that is, moving image data separated by a predetermined period of time). Note that although the configuration in which labeled data and unlabeled data are input from separate input units is described here, labeled data and unlabeled data can be input from one common input unit. good.
  • labeled data input section 110 and unlabeled data input section 120 one data input section capable of inputting both labeled data and unlabeled data may be provided.
  • the teacher model learning unit 130 is configured to be able to learn a teacher model using the labeled data input to the labeled data input unit 110.
  • the training model here is a model for generating pseudo-labels to be assigned to unlabeled data.
  • a “pseudo-label” is a pseudo-correct label generated by a model trained on labeled data. Since existing techniques can be appropriately adopted for a specific learning method of the teacher model, detailed description thereof will be omitted here.
  • Consistency Regularization may be used to improve accuracy.
  • the information processing system 10 according to the first embodiment may not include the teacher model learning unit 130 . In this case, learning of a teacher model using labeled data may be performed outside the system, and the learned teacher model may be input to the information processing system 10 .
  • the pseudo-label generating unit 140 is configured to be able to generate pseudo-labels to be assigned to unlabeled data using the teacher model learned by the teacher model learning unit 130. In addition, the pseudo-label generator 140 is configured to be able to assign the generated pseudo-label to unlabeled data.
  • the pseudo-label evaluation unit 150 is configured to be able to evaluate the pseudo-labels generated by the pseudo-label generation unit 140. Specifically, the pseudo-label evaluation unit 150 is configured to be able to evaluate whether or not the pseudo-label satisfies a predetermined evaluation criterion. Note that the “predetermined evaluation criteria” here are criteria for determining whether the quality of the pseudo-label is sufficiently high, and are set in advance. The pseudo-label evaluation unit 150 is configured to output pseudo-labels (evaluation pseudo-labels) that meet a predetermined evaluation criterion, but not to output pseudo-labels that do not meet the predetermined evaluation criterion.
  • the pseudo-label evaluation unit 150 may evaluate the pseudo-label as being of low quality and not output it. In addition, the pseudo-label evaluating unit 150 may evaluate the pseudo-label as high quality and output it if the error is not more than twice the average error in the test set of labeled data.
  • the pseudo-label evaluation unit 150 is configured to evaluate pseudo-labels using an evaluation model learned using at least one of labeled data and unlabeled data. A method of learning the evaluation model will be described in detail in another embodiment described later.
  • the student model learning unit 160 uses the labeled data and the unlabeled data (hereinafter referred to as “pseudo-label data” as appropriate) to which the evaluation pseudo-labels output from the pseudo-label evaluation unit 150 are used to create a student model. Configured to be learnable. Note that the student model here is a model for generating pseudo-labels to be assigned to unlabeled data, like the teacher model. As for the specific learning method of the student model, existing techniques can be appropriately adopted, so detailed description thereof will be omitted here. When learning the student model, for example, existing distillation techniques may be used in combination.
  • the model output unit 170 is configured to be able to output a learned student model. Also, the model output unit 170 may be configured to be capable of outputting a teacher model and an evaluation model in addition to the learned student model.
  • FIG. 3 is a flow chart showing the operation flow of the information processing system according to the first embodiment.
  • the teacher model learning unit 130 uses the labeled data input from the labeled data input unit 110 to generate teacher A model is learned (step S101).
  • the pseudo-label generation unit 140 generates pseudo-labels using the trained teacher model and assigns them to the unlabeled data input from the unlabeled data input unit 120 (step S102).
  • the pseudo-label evaluation unit 150 learns the evaluation model (step S103). Then, the pseudo-label evaluating unit 150 removes low-quality pseudo-labels generated by the pseudo-label generating unit 140 (step S104). In other words, the pseudo-label evaluation unit 150 outputs only high-quality evaluation pseudo-labels.
  • the student model learning unit 160 learns a student model using the labeled data and the pseudo-labeled data to which the evaluation pseudo-label is assigned (step S105). After that, the model output unit 170 outputs the trained model (step S106).
  • pseudo-labels As described with reference to FIGS. 1 to 3, in the information processing system 10 according to the first embodiment, by evaluating generated pseudo-labels, high-quality pseudo-labels (that is, pseudo-labels output as evaluation results) ) is used for learning. In this way, it becomes possible to assign appropriate pseudo-labels to unlabeled data. More specifically, when dealing with regression problems, pseudo-labels can be given to unlabeled data with high accuracy. Therefore, for example, it is possible to reduce the cost of labeling unlabeled data.
  • FIG. 4 An information processing system 10 according to the second embodiment will be described with reference to FIGS. 4 and 5.
  • FIG. The second embodiment may differ from the above-described first embodiment only in a part of configuration and operation, and the other parts may be the same as those of the first embodiment. Therefore, in the following, portions different from the already described first embodiment will be described in detail, and descriptions of other overlapping portions will be omitted as appropriate.
  • FIG. 4 is a block diagram showing the functional configuration of an information processing system according to the second embodiment.
  • symbol is attached
  • the information processing system 10 includes a labeled data input unit 110, an unlabeled data input unit 120, and a teacher model learning unit as processing blocks for realizing its functions. 130 , a pseudo-label generation unit 140 , a pseudo-label evaluation unit 150 , a student model learning unit 160 , a model output unit 170 and a domain conversion unit 180 . That is, the information processing system 10 according to the second embodiment further includes a domain conversion unit 180 in addition to the configuration of the first embodiment (see FIG. 2).
  • the domain conversion unit 180 is configured to convert labeled data before input to the labeled data input unit 110 and unlabeled data before input to the unlabeled data input unit 120 into a common domain. It is That is, the domain conversion unit 180 is configured to be able to execute processing for aligning the domain of labeled data and the domain of unlabeled data. Note that the converted domain may be a completely different domain from the original domain. That is, labeled data and unlabeled data may be transformed into a third domain different from their original domain.
  • the processing performed by the domain conversion unit 180 may be image conversion processing when the data is image data.
  • the domain transformation unit 180 may perform a process of applying a Laplacian filter to the image to transform the original data into edge-detected data of the image.
  • the domain conversion unit 180 may convert the domain by Style Transfer (eg, AdaIN).
  • the domain converter 180 may convert the domain by changing illumination or resolution.
  • the domain conversion unit 180 may extract the feature amount of the data and calculate the distance between domains from the feature amount using Kullback-Leibler divergence.
  • FIG. 5 is a flow chart showing the operation flow of the information processing system according to the second embodiment.
  • the same reference numerals are assigned to the same processes as those shown in FIG.
  • the domain conversion unit 180 first converts labeled data and unlabeled data into a common domain (step S201).
  • the labeled data and unlabeled data after domain conversion are input to the labeled data input section 110 and the unlabeled data input section 120, respectively.
  • the teacher model learning unit 130 learns a teacher model using the labeled data input from the labeled data input unit 110 (step S101). Then, the pseudo-label generation unit 140 generates pseudo-labels using the trained teacher model and assigns them to the unlabeled data input from the unlabeled data input unit 120 (step S102).
  • the pseudo-label evaluation unit 150 learns the evaluation model (step S103). Then, the pseudo-label evaluating unit 150 removes low-quality pseudo-labels generated by the pseudo-label generating unit 140 (step S104).
  • the student model learning unit 160 learns a student model using the labeled data before domain conversion and the pseudo-labeled data obtained by adding evaluation pseudo-labels to the unlabeled data before domain conversion (step S202). .
  • the model output unit 170 outputs the trained model (step S106).
  • labeled data and unlabeled data are converted into domains common to each other.
  • the labeled data and the unlabeled data have different domains, it is possible to perform appropriate learning.
  • it is no longer necessary to use only data having a common domain and for example, it is possible to easily increase the number of data used for learning.
  • FIG. 10 An information processing system 10 according to the third embodiment will be described with reference to FIG. Note that the third embodiment shows an example of a method of learning an evaluation model, and the configuration, operation, etc. of the system may be the same as those of the first and second embodiments. Therefore, in the following, portions different from the already described embodiments will be described in detail, and descriptions of other overlapping portions will be omitted as appropriate.
  • FIG. 6 is a block diagram showing an evaluation model learning method in the information processing system according to the third embodiment.
  • the pseudo-labeled unlabeled data (specifically, data generated by the pseudo-label generation unit 140) is supplied to the pseudo-label evaluation unit 150.
  • unlabeled data with pseudo-labels is input.
  • the pseudo-label evaluation unit 150 learns the evaluation model 151 using the unlabeled data to which the pseudo-label data is assigned. Note that the pseudo-label evaluation unit 150 may perform learning using the pseudo-labels themselves in addition to the unlabeled data to which the pseudo-labels are assigned.
  • the number of epochs is preferably less than 10 to prevent overfitting.
  • the number of epochs may be set, for example, based on the size of the dataset or batch size used for learning. According to the research conducted by the inventor of the present application, it has been found that appropriate learning can be performed in many cases by setting the number of epochs to one.
  • the evaluation model 151 is learned using unlabeled data to which pseudo-labels are assigned. In this way, the evaluation model 151 can be learned appropriately, so that pseudo-label evaluation can be performed appropriately.
  • FIG. 7 is a block diagram showing an evaluation model learning method in the information processing system according to the fourth embodiment.
  • the pseudo-label evaluation unit 150 receives labeled data and pseudo-labeled unlabeled data (specifically, pseudo-label generation The pseudo-labeled unlabeled data generated in section 140) is input. Then, the pseudo-labeled evaluation unit 150 learns the evaluation model 151 using the labeled data and the unlabeled data to which the pseudo-labeled data is added. More specifically, the pseudo-label evaluation unit 150 first learns an evaluation model using labeled data. Note that the number of labeled data used for this learning may be relatively small. After that, the pseudo-label evaluation unit 150 learns the evaluation model 151 using the unlabeled data to which the pseudo-labels are assigned. A relatively large number of unlabeled data may be used for this learning.
  • the evaluation model 151 is first learned with labeled data, and then learned with unlabeled data to which pseudo-labels are added. In this way, if the labeled data is first used for learning, the evaluation model can be more appropriately compared to the case where the labeled data is not used (i.e., the case where only unlabeled data is used for learning). 151 can be learned. Therefore, pseudo-labels can be evaluated appropriately.
  • the fifth embodiment shows an example of a method of learning an evaluation model in the same manner as in the above-described third and fourth embodiments, and the configuration and operation of the system are the same as those in the first and second embodiments. may be identical. Therefore, in the following, portions different from the already described embodiments will be described in detail, and descriptions of other overlapping portions will be omitted as appropriate.
  • FIG. 8 is a block diagram showing an evaluation model learning method in the information processing system according to the fifth embodiment.
  • labeled data is input to the pseudo-label evaluation unit 150 .
  • the pseudo-label evaluation unit 150 learns the evaluation model 151 using the labeled data. It is preferable that the number of labeled data used for this learning is relatively large.
  • the evaluation model 151 is learned in the same way as the teacher model in the teacher model learning unit 130 (see FIG. 2). Therefore, in this case, there are two teacher models. Specifically, the configuration is such that a pseudo-label generated by one teacher model is evaluated by another teacher model.
  • the evaluation model 151 is learned using labeled data. In this way, the evaluation model 151 can be learned appropriately, so that pseudo-label evaluation can be performed appropriately.
  • FIG. 10 An information processing system 10 according to the sixth embodiment will be described with reference to FIG. Note that the sixth embodiment shows an example of a method of learning an evaluation model in the same manner as in the third to fifth embodiments described above, and the configuration and operation of the system are the same as those in the first and second embodiments. may be identical. Therefore, in the following, portions different from the already described embodiments will be described in detail, and descriptions of other overlapping portions will be omitted as appropriate.
  • FIG. 9 is a block diagram showing an evaluation model learning method in the information processing system according to the sixth embodiment.
  • the output of the teacher model (in other words, the label estimated by the teacher model from the unlabeled data) and the label Yes data labels and are entered.
  • the pseudo-label evaluation unit 150 first receives only the labeled data, calculates the difference between the output of the teacher model and the label of the labeled data, and learns the evaluation model 151 .
  • the pseudo-label evaluation unit 150 evaluates the pseudo-label using the difference value described above as quality. Specifically, the pseudo-label evaluation unit 150 inputs pseudo-label data to the trained evaluation model 151, estimates the difference between the estimated (pseudo) label and the true label, and calculates a large difference value. are evaluated as poor estimation accuracy of the teacher model, and those with small calculated difference values are evaluated as good estimation accuracy of the teacher model.
  • the evaluation model 151 is learned using the difference between the output of the teacher model and the label of the labeled data. In this way, the evaluation model 151 can be learned appropriately, so that pseudo-label evaluation can be performed appropriately.
  • the seventh embodiment shows a configuration example of the pseudo-label evaluation unit 150, and the system configuration and various operations may be the same as those of the first to sixth embodiments described above. Therefore, in the following, portions different from the already described embodiments will be described in detail, and descriptions of other overlapping portions will be omitted as appropriate.
  • FIG. 10 is a block diagram showing the configuration of the model evaluation unit in the information processing system according to the seventh embodiment.
  • the pseudo-label evaluation unit 150 is configured with a plurality of evaluation models 151 .
  • the pseudo-label evaluation unit 150 includes three evaluation models 151a, 151b, and 151c is given here, the number is not particularly limited.
  • the pseudo-label evaluation unit 150 may have two evaluation models 151 or four or more evaluation models 151 .
  • the multiple evaluation models 151 are models learned separately. However, the plurality of evaluation models 151 may be learned using a common data set, or may be learned using different data sets. Note that the plurality of evaluation models 151 may perform learning using data to which perturbation is applied. The method of applying perturbation to data is not particularly limited.
  • the plurality of evaluation models 151 may be learned by the learning methods described in the third to sixth embodiments.
  • each of the plurality of evaluation models 151 may be learned by different learning methods.
  • the evaluation model 151a is learned by the learning method described in the third embodiment (that is, the method of learning using unlabeled data: see FIG. 6)
  • the evaluation model 151b is learned by the learning method described in the fourth embodiment ( That is, the method of learning using labeled data and unlabeled data: see FIG. 7)
  • the evaluation model 151c is learned by the learning method described in the fifth embodiment (i.e., the method of learning using labeled data: (see FIG. 8).
  • the pseudo-label evaluation unit 150 evaluates pseudo-labels using the plurality of evaluation models 151 described above. Specifically, the pseudo-label evaluation unit 150 first outputs evaluation results from each of the plurality of evaluation models 151, and outputs one final evaluation result according to the plurality of evaluation results. More specifically, the evaluation result of each of the plurality of evaluation models 151 is determined by majority or the average value is calculated, and the overall evaluation result is output.
  • the information processing system 10 uses a plurality of evaluation models 151 to evaluate pseudo-labels. In this way, pseudo-labels can be evaluated more appropriately than when evaluation is performed using only one evaluation model.
  • FIG. 11 and 12 An information processing system 10 according to the eighth embodiment will be described with reference to FIGS. 11 and 12.
  • FIG. 10 It should be noted that the eighth embodiment may differ from the above-described first to seventh embodiments only in a part of the configuration and operation, and the other parts may be the same as those of the first to seventh embodiments. . Therefore, in the following, portions different from the already described embodiments will be described in detail, and descriptions of other overlapping portions will be omitted as appropriate.
  • FIG. 11 is a block diagram showing the functional configuration of an information processing system according to the eighth embodiment.
  • symbol is attached
  • the information processing system 10 includes a labeled data input unit 110, an unlabeled data input unit 120, and a teacher model learning unit as processing blocks for realizing its functions. 130 , a pseudo-label generation unit 140 , a pseudo-label evaluation unit 150 , a student model learning unit 160 and a model output unit 170 .
  • the labeled data input unit 110, the unlabeled data input unit 120, the teacher model learning unit 130, the pseudo label generation unit 140, and the pseudo label evaluation unit 150 are It is configured as a label learning unit 200 .
  • Pseudo label learning unit 200 repeats learning of a teacher model by teacher model learning unit 130, generation of pseudo labels by pseudo label generation unit 140, and pseudo label evaluation by pseudo label evaluation unit 150, thereby generating pseudo labels. It is configured so that learning about labels can be performed more appropriately.
  • the pseudo-label learning unit 200 is configured to reflect the evaluation result of the pseudo-label evaluation unit 150 and execute the learning of the teacher model. For example, the error calculated by the pseudo-label evaluation unit 150 is propagated back to the teacher model learning unit 130 so that the teacher model can be re-learned.
  • the pseudo-label learning unit 200 is set to repeat a series of processes until a predetermined number of times is reached. The predetermined number of times may be a value obtained by prior simulation or the like.
  • FIG. 12 is a flow chart showing the operation flow of the information processing system according to the eighth embodiment.
  • the same reference numerals are given to the same processes as those shown in FIG.
  • the teacher model learning unit 130 uses the labeled data input from the labeled data input unit 110 to generate teacher A model is learned (step S101).
  • the pseudo-label generation unit 140 generates pseudo-labels using the trained teacher model and assigns them to the unlabeled data input from the unlabeled data input unit 120 (step S102).
  • the pseudo-label evaluation unit 150 learns the evaluation model (step S103). Then, the pseudo-label evaluating unit 150 removes low-quality pseudo-labels generated by the pseudo-label generating unit 140 (step S104).
  • the information processing system 10 determines whether or not the series of processes up to this point has been repeated a predetermined number of times (step S701). Then, when it is determined that the series of processes has not been repeated a predetermined number of times (step S701: NO), the evaluation result of the pseudo-label evaluation unit 150 is reflected (step S702), and the process is executed again from step S101. .
  • step S701 YES
  • the final evaluation pseudo-label is output, and the student model learning unit 160 applies the labeled data and the evaluation pseudo-label.
  • a student model is learned using the obtained pseudo-label data (step S105).
  • the model output unit 170 outputs the trained model (step S106).
  • the pseudo-label learning unit 200 learns about pseudo-labels (specifically, learns a teacher model, generates pseudo-labels, and pseudo-label evaluation) are performed repeatedly. In this way, each model is learned to be in a more appropriate state, so that a more appropriate pseudo-label (that is, an evaluation pseudo-label) can be output.
  • FIG. 10 An information processing system 10 according to the ninth embodiment will be described with reference to FIGS. 13 and 14.
  • FIG. 13 is a block diagram showing the functional configuration of an information processing system according to the ninth embodiment.
  • symbol is attached
  • the information processing system 10 includes a labeled data input unit 110, an unlabeled data input unit 120, and a teacher model learning unit as processing blocks for realizing its functions. 130 , a pseudo-label generation unit 140 , a pseudo-label evaluation unit 150 , a student model learning unit 160 , a model output unit 170 and a model adjustment unit 190 . That is, the information processing system 10 according to the ninth embodiment further includes a model adjustment unit 190 in addition to the configuration of the first embodiment (see FIG. 2).
  • the model adjustment unit 190 is configured so that some layers of a trained model can be adjusted using labeled data. Specifically, the model adjustment unit 190 is configured to be able to perform Fine Tuning on the learned model.
  • FIG. 14 is a flow chart showing the operation flow of the information processing system according to the ninth embodiment.
  • the same reference numerals are assigned to the same processes as those shown in FIG.
  • the teacher model learning unit 130 uses the labeled data input from the labeled data input unit 110 to generate teacher A model is learned (step S101).
  • the pseudo-label generation unit 140 generates pseudo-labels using the trained teacher model and assigns them to the unlabeled data input from the unlabeled data input unit 120 (step S102).
  • the pseudo-label evaluation unit 150 learns the evaluation model (step S103). Then, the pseudo-label evaluating unit 150 removes low-quality pseudo-labels generated by the pseudo-label generating unit 140 (step S104).
  • the student model learning unit 160 learns a student model using the labeled data and the pseudo-labeled data to which the evaluation pseudo-label is assigned (step S105).
  • the model adjustment unit 190 adjusts the learned model using the labeled data (step S801).
  • the model output unit 170 outputs the adjusted learned model (step S106).
  • the information processing system 10 adjusts the model before outputting the trained model. In this way, it is possible to make the output learned model more suitable.
  • a processing method of recording a program for operating the configuration of each embodiment so as to realize the functions of each embodiment described above on a recording medium, reading the program recorded on the recording medium as a code, and executing it on a computer is also implemented. Included in the category of form. That is, a computer-readable recording medium is also included in the scope of each embodiment. In addition to the recording medium on which the above program is recorded, the program itself is also included in each embodiment.
  • a floppy (registered trademark) disk, hard disk, optical disk, magneto-optical disk, CD-ROM, magnetic tape, non-volatile memory card, and ROM can be used as recording media.
  • the program recorded on the recording medium alone executes the process, but also the one that operates on the OS in cooperation with other software and the function of the expansion board to execute the process. included in the category of
  • appendix 1 uses a data input means for inputting labeled data and unlabeled data, and a teacher model trained using the labeled data to assign pseudo-labels to the unlabeled data. and an evaluation model trained using at least one of the labeled data and the unlabeled data to evaluate the pseudo-label given to the unlabeled data, and a predetermined evaluation criterion
  • the information processing system includes model learning means and model output means for outputting the learned student model.
  • Appendix 2 The information processing system according to appendix 2, further comprising domain conversion means for converting the labeled data and the unlabeled data input to the data input means into a domain common to each other.
  • appendix 3 The information processing system according to appendix 3 is the information processing system according to appendix 1 or 2, wherein the evaluation model is trained using only the unlabeled data.
  • the information processing system according to appendix 4 is the information processing system according to appendix 1 or 2, wherein the evaluation model is learned using a part of the labeled data and then trained using the unlabeled data. is.
  • the information processing system according to appendix 5 is the information processing system according to appendix 1 or 2, wherein the evaluation model is trained using only the labeled data.
  • Appendix 7 The information processing system according to any one of appendices 1 to 6, wherein the pseudo-label evaluation means evaluates the pseudo-labels using a plurality of separately learned evaluation models. System.
  • the information processing system according to appendix 8 further comprises teacher model learning means for learning the teacher model using the labeled data, wherein the teacher model learning means uses the evaluation result of the pseudo-label evaluation means to: 8.
  • Appendix 10 In the information processing method according to appendix 10, labeled data and unlabeled data are input, a teacher model trained using the labeled data is used to assign a pseudo-label to the unlabeled data, and the label Evaluating the pseudo-labels assigned to the unlabeled data using an evaluation model trained using at least one of the labeled data and the unlabeled data, and evaluating the pseudo-labels that meet predetermined evaluation criteria.
  • the recording medium inputs labeled data and unlabeled data into a computer, uses a teacher model trained using the labeled data to give a pseudo-label to the unlabeled data,
  • the pseudo-label given to the unlabeled data is evaluated using an evaluation model trained using at least one of the labeled data and the unlabeled data, and the pseudo-label that meets a predetermined evaluation criterion.
  • an evaluation pseudo-label learning a student model using the pseudo-label data obtained by assigning the evaluation pseudo-label to the labeled data and the unlabeled data, and outputting the learned student model;
  • the computer program according to Supplementary Note 12 inputs labeled data and unlabeled data to a computer, uses a teacher model trained using the labeled data to give a pseudo-label to the unlabeled data,
  • the pseudo-label given to the unlabeled data is evaluated using an evaluation model trained using at least one of the labeled data and the unlabeled data, and the pseudo-label that meets a predetermined evaluation criterion.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Artificial Intelligence (AREA)
  • Mathematical Physics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Computational Linguistics (AREA)
  • Biophysics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
PCT/JP2021/018619 2021-05-17 2021-05-17 情報処理システム、情報処理方法、及び記録媒体 Ceased WO2022244059A1 (ja)

Priority Applications (3)

Application Number Priority Date Filing Date Title
JP2023522008A JP7552890B2 (ja) 2021-05-17 2021-05-17 情報処理システム、情報処理方法、及び記録媒体
PCT/JP2021/018619 WO2022244059A1 (ja) 2021-05-17 2021-05-17 情報処理システム、情報処理方法、及び記録媒体
US18/290,346 US20240289633A1 (en) 2021-05-17 2021-05-17 Information processing system, information processing method, and recording medium

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2021/018619 WO2022244059A1 (ja) 2021-05-17 2021-05-17 情報処理システム、情報処理方法、及び記録媒体

Publications (1)

Publication Number Publication Date
WO2022244059A1 true WO2022244059A1 (ja) 2022-11-24

Family

ID=84141382

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2021/018619 Ceased WO2022244059A1 (ja) 2021-05-17 2021-05-17 情報処理システム、情報処理方法、及び記録媒体

Country Status (3)

Country Link
US (1) US20240289633A1 (https=)
JP (1) JP7552890B2 (https=)
WO (1) WO2022244059A1 (https=)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116342979A (zh) * 2023-03-30 2023-06-27 北京百度网讯科技有限公司 伪标签处理方法、装置、电子设备和存储介质

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20230075369A1 (en) * 2021-09-08 2023-03-09 Sap Se Pseudo-label generation using an ensemble model

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160253597A1 (en) * 2015-02-27 2016-09-01 Xerox Corporation Content-aware domain adaptation for cross-domain classification
WO2019155873A1 (ja) * 2018-02-06 2019-08-15 オムロン株式会社 評価装置、動作制御装置、評価方法、及び評価プログラム

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20160253597A1 (en) * 2015-02-27 2016-09-01 Xerox Corporation Content-aware domain adaptation for cross-domain classification
WO2019155873A1 (ja) * 2018-02-06 2019-08-15 オムロン株式会社 評価装置、動作制御装置、評価方法、及び評価プログラム

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
I. ZEKI YALNIZ; HERV\'E J\'EGOU; KAN CHEN; MANOHAR PALURI; DHRUV MAHAJAN: "Billion-scale semi-supervised learning for image classification", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 2 May 2019 (2019-05-02), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081271363 *
QIZHE XIE; MINH-THANG LUONG; EDUARD HOVY; QUOC V. LE: "Self-training with Noisy Student improves ImageNet classification", ARXIV.ORG, CORNELL UNIVERSITY LIBRARY, 201 OLIN LIBRARY CORNELL UNIVERSITY ITHACA, NY 14853, 19 June 2020 (2020-06-19), 201 Olin Library Cornell University Ithaca, NY 14853 , XP081680806 *

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116342979A (zh) * 2023-03-30 2023-06-27 北京百度网讯科技有限公司 伪标签处理方法、装置、电子设备和存储介质

Also Published As

Publication number Publication date
JP7552890B2 (ja) 2024-09-18
JPWO2022244059A1 (https=) 2022-11-24
US20240289633A1 (en) 2024-08-29

Similar Documents

Publication Publication Date Title
KR102546817B1 (ko) 지식 증류에서 총 cam 정보를 이용한 교사 지원 어텐션 전달의 학습 방법 및 장치
CN111699497B (zh) 使用离散潜变量的序列模型的快速解码
RU2007127915A (ru) Способ вычисления цифровых видеоголограмм и устройство для его осуществления
JP2018169672A (ja) 教師画像を生成する方法、コンピュータおよびプログラム
JP2019046269A (ja) 機械学習用訓練データの生成
CN112640037A (zh) 学习装置、推理装置、学习模型的生成方法及推理方法
TWI800234B (zh) 機器學習用數據生成方法、元學習方法、機器學習用數據生成裝置及程式
US20190286946A1 (en) Learning program, learning method, and learning apparatus
JP7552890B2 (ja) 情報処理システム、情報処理方法、及び記録媒体
JP7586172B2 (ja) 情報処理装置およびプログラム
WO2020240808A1 (ja) 学習装置、分類装置、学習方法、分類方法、学習プログラム、及び分類プログラム
JP6622369B1 (ja) 訓練データを生成する方法、コンピュータおよびプログラム
CN120782945A (zh) 用于图像生成的方法、系统和存储介质
TW202439239A (zh) 使用超取樣技術的方法及電子裝置
US11568303B2 (en) Electronic apparatus and control method thereof
CN116091905A (zh) 一种域自适应撞击坑检测方法、系统、装置及存储介质
JP7192873B2 (ja) 情報処理プログラム、情報処理方法および学習装置
WO2020137641A1 (ja) 復元装置、復元方法、およびプログラム
WO2021220343A1 (ja) データ生成装置、データ生成方法、学習装置及び記録媒体
WO2024254348A1 (en) On-device image generation with generative adversarial networks
US20250165828A1 (en) Classical simulation of a quantum system
CN115053235B (zh) 使用插补的序列建模
JP7188589B2 (ja) 復元装置、復元方法、およびプログラム
Cao et al. Generation of inverse assist features using generative adversarial networks
WO2020144853A1 (ja) 学習装置、学習方法および学習プログラム

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: 21940679

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2023522008

Country of ref document: JP

WWE Wipo information: entry into national phase

Ref document number: 18290346

Country of ref document: US

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21940679

Country of ref document: EP

Kind code of ref document: A1