WO2023058433A1 - 学習装置、学習方法、センシングデバイス及びデータ収集方法 - Google Patents
学習装置、学習方法、センシングデバイス及びデータ収集方法 Download PDFInfo
- Publication number
- WO2023058433A1 WO2023058433A1 PCT/JP2022/034670 JP2022034670W WO2023058433A1 WO 2023058433 A1 WO2023058433 A1 WO 2023058433A1 JP 2022034670 W JP2022034670 W JP 2022034670W WO 2023058433 A1 WO2023058433 A1 WO 2023058433A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data
- learning
- model
- unit
- influence
- Prior art date
Links
- 238000000034 method Methods 0.000 title claims abstract description 114
- 238000013480 data collection Methods 0.000 title claims description 8
- 230000008569 process Effects 0.000 claims abstract description 65
- 238000004364 calculation method Methods 0.000 claims abstract description 58
- 238000010801 machine learning Methods 0.000 claims abstract description 27
- 230000006870 function Effects 0.000 claims description 63
- 238000012545 processing Methods 0.000 claims description 51
- 230000005540 biological transmission Effects 0.000 claims description 34
- 238000013523 data management Methods 0.000 claims description 18
- 230000009471 action Effects 0.000 claims description 4
- 230000010365 information processing Effects 0.000 description 137
- 238000010586 diagram Methods 0.000 description 23
- 238000013528 artificial neural network Methods 0.000 description 22
- 238000004891 communication Methods 0.000 description 19
- 238000011156 evaluation Methods 0.000 description 8
- 238000003384 imaging method Methods 0.000 description 7
- 230000009467 reduction Effects 0.000 description 7
- 238000012937 correction Methods 0.000 description 6
- 230000000694 effects Effects 0.000 description 6
- 238000012549 training Methods 0.000 description 6
- 101001111655 Homo sapiens Retinol dehydrogenase 11 Proteins 0.000 description 4
- 102100023916 Retinol dehydrogenase 11 Human genes 0.000 description 4
- 230000015654 memory Effects 0.000 description 4
- 238000012986 modification Methods 0.000 description 4
- 230000004048 modification Effects 0.000 description 4
- 239000004065 semiconductor Substances 0.000 description 4
- 238000012360 testing method Methods 0.000 description 4
- 101100366082 Saccharomyces cerevisiae (strain ATCC 204508 / S288c) SNF7 gene Proteins 0.000 description 3
- 230000008859 change Effects 0.000 description 3
- 238000004140 cleaning Methods 0.000 description 3
- 238000013527 convolutional neural network Methods 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 239000011159 matrix material Substances 0.000 description 3
- 230000003287 optical effect Effects 0.000 description 3
- 238000013526 transfer learning Methods 0.000 description 3
- 101150047375 DID2 gene Proteins 0.000 description 2
- 101100317166 Saccharomyces cerevisiae (strain ATCC 204508 / S288c) VPS24 gene Proteins 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 235000000332 black box Nutrition 0.000 description 2
- 238000013499 data model Methods 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 238000013461 design Methods 0.000 description 2
- 238000005401 electroluminescence Methods 0.000 description 2
- 238000003702 image correction Methods 0.000 description 2
- 239000004973 liquid crystal related substance Substances 0.000 description 2
- 230000004044 response Effects 0.000 description 2
- 238000012706 support-vector machine Methods 0.000 description 2
- 101150040982 DID4 gene Proteins 0.000 description 1
- MKYBYDHXWVHEJW-UHFFFAOYSA-N N-[1-oxo-1-(2,4,6,7-tetrahydrotriazolo[4,5-c]pyridin-5-yl)propan-2-yl]-2-[[3-(trifluoromethoxy)phenyl]methylamino]pyrimidine-5-carboxamide Chemical compound O=C(C(C)NC(=O)C=1C=NC(=NC=1)NCC1=CC(=CC=C1)OC(F)(F)F)N1CC2=C(CC1)NN=N2 MKYBYDHXWVHEJW-UHFFFAOYSA-N 0.000 description 1
- 230000001133 acceleration Effects 0.000 description 1
- 238000013459 approach Methods 0.000 description 1
- 210000004556 brain Anatomy 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 230000001419 dependent effect Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 229910044991 metal oxide Inorganic materials 0.000 description 1
- 150000004706 metal oxides Chemical class 0.000 description 1
- 230000001151 other effect Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 230000006403 short-term memory Effects 0.000 description 1
- 210000004243 sweat Anatomy 0.000 description 1
- 210000000225 synapse Anatomy 0.000 description 1
- 230000008685 targeting Effects 0.000 description 1
Images
Classifications
-
- 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
- G06V10/7747—Organisation of the process, e.g. bagging or boosting
-
- 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
- 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
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Definitions
- the present disclosure relates to learning devices, learning methods, sensing devices, and data collection methods.
- teacher data is created by assigning a correct label to each of a plurality of data based on whether or not an evaluation button has been pressed by the user, and when the plurality of teacher data reaches a predetermined number, the teacher data is generated. Depending on the situation, machine learning is performed using multiple teacher data.
- this disclosure proposes a learning device, a learning method, a sensing device, and a data collection method that can make a model generated using appropriate data available.
- a learning device includes a calculation unit that calculates the degree of influence that data collected by a sensing device has on learning of a model by machine learning, and calculation by the calculation unit: a learning unit that generates a learned model by a small-label learning process of learning the model using the data whose degree of influence satisfies a condition.
- FIG. 4 is a diagram illustrating an example of learning processing according to an embodiment of the present disclosure
- FIG. 1 is a diagram illustrating a configuration example of an information processing system according to an embodiment of the present disclosure
- FIG. 1 is a diagram illustrating a configuration example of a server device according to an embodiment of the present disclosure
- FIG. 4 is a diagram illustrating an example of a data information storage unit according to an embodiment of the present disclosure
- FIG. It is a figure showing an example of a model information storage part concerning an embodiment of this indication.
- 4 is a diagram illustrating an example of a threshold information storage unit according to an embodiment of the present disclosure
- FIG. 10 is a diagram of an example of a network corresponding to a model
- 1 is a diagram showing a configuration example of a sensing device according to an embodiment of the present disclosure
- FIG. It is a figure showing an example of a model information storage part concerning an embodiment of this indication.
- 4 is a flowchart showing processing of the server device according to the embodiment of the present disclosure
- FIG. 4 is a sequence diagram showing processing procedures of the information processing system according to the embodiment of the present disclosure
- 1 is a hardware configuration diagram showing an example of a computer that realizes the functions of an information processing device such as a server device and a sensing device;
- Embodiment 1-1 Outline of learning process according to embodiment of present disclosure 1-1-1. Background and Effects 1-1-2. Influence function 1-1-2-1. Examples of other methods 1-1-3. Capacity limit 1-1-4. Storage storage 1-1-5. Image correction 1-2. Configuration of information processing system according to embodiment 1-3. Configuration of learning device according to embodiment 1-3-1. Model (network) example 1-4. Configuration of Sensing Device According to Embodiment 1-5. Information processing procedure according to embodiment 1-5-1. Procedure of processing related to learning device 1-5-2. Procedure of processing related to information processing system2. Other Embodiments 2-1. Other configuration examples 2-2. Others 3. Effects of the present disclosure 4 . Hardware configuration
- FIG. 1 is a diagram illustrating an example of a processing flow of an information processing system according to an embodiment of the present disclosure.
- FIG. 2 is a diagram illustrating an example of learning processing according to an embodiment of the present disclosure.
- FIG. 2 is a diagram illustrating an example of learning processing according to the embodiment of the present disclosure.
- the learning process according to the embodiment of the present disclosure is realized by the information processing system 1 including the server device 100 and the sensing device 10, which are examples of learning devices.
- FIG. 2 outlines the learning process implemented by the information processing system 1 .
- model a discriminative model that is a deep neural network (DNN) that performs image recognition.
- DNN deep neural network
- any device included in the information processing system 1 may perform the processing described below where the information processing system 1 is the subject of the processing.
- FIG. 1 the main body of processing is described as the information processing system 1, but each processing shown in FIG.
- the information processing system 1 uses first data LDD, which is labeled data (also referred to as "labeled data") indicating the correct recognition result of the image, and second data collected by the sensing device 10.
- first data LDD is data already used as a data set.
- the label may be any information that indicates the correct recognition result for the data.
- the label may be information indicating the category to which the image belongs, information indicating the object included in the image, information indicating the area of the object in the image, or the like.
- the second data ULD is data without a label indicating the correct recognition result of the image (also referred to as "unlabeled data").
- the small-label learning process by the information processing system 1 is a learning process performed using both labeled data and unlabeled data as described above.
- the information processing system 1 learns a classifier using the first data LDD (step S1).
- the information processing system 1 uses the first data LDD to learn a classifier that classifies images.
- the information processing system 1 learns a classifier that receives an image as input and outputs information indicating the category to which the input image belongs.
- the information processing system 1 learns a classifier using the image data included in the first data LDD and the labels of the image data.
- the information processing system 1 calculates the degree of influence of each data of the second data ULD (step S2).
- the degree of influence of data is information indicating the degree of influence of the data on learning of the model, and the details will be described later.
- the information processing system 1 then compares the calculated degree of influence of the data with the threshold value S (step S3).
- step S4 the information processing system 1 deletes the data (step S4). For example, the information processing system 1 determines that data having a degree of influence equal to or less than the threshold value S among the second data ULD is data having a low degree of influence on learning (low influence data), and deletes the data. do.
- step S5 the information processing system 1 predicts the label of that data (step S5). For example, the information processing system 1 determines that data having a degree of influence greater than the threshold value S is data having a high degree of influence on learning (high-influence data), and predicts the label of the data.
- the information processing system 1 uses a classifier to predict the label of data that is unlabeled data and high-impact data in the second data ULD.
- the information processing system 1 assigns the predicted label (predicted label) to the prediction target data (prediction target data).
- the third data NLD is data added with a predictive label among the second data ULD. That is, the third data NLD is unlabeled data, but becomes labeled data by adding a predicted label.
- the information processing system 1 generates a data set NDS, which is a new data set, from the labeled third data NLD and the first data LDD.
- the information processing system 1 generates the data set NDS by adding the labeled third data NLD to the first data LDD. Note that the above is only an example. For example, when the label is used to calculate the degree of influence, the information processing system 1 predicts the label of data before calculating the degree of influence, and uses the predicted label to calculate the degree of influence. may be calculated.
- the information processing system 1 uses the data set NDS to learn the model. (Step S6). For example, the information processing system 1 learns a model using the image data included in the data set NDS and the labels of the image data.
- the information processing system 1 distributes the model to the edge device (step S7).
- the server device 100 of the information processing system 1 transmits the learned model to the sensing device 10, which is an edge device.
- the information processing system 1 repeats the processing of steps S1 to S7 using the data collected from the model distributed by the sensing device 10 as the second data ULD.
- the information processing system 1 uses the data set NDS of the previous process as the first data LDD.
- the information processing system 1 can update the model and improve the performance of the model by repeating the above-described processing (loop) at regular intervals (regularly).
- model M1 is a neural network used for image recognition.
- the server device 100 shown in FIG. 2 is an information processing device that calculates the degree of influence of data collected by the sensing device 10 on model learning by machine learning, and learns the model using data whose degree of influence satisfies a condition. be.
- the sensing device 10 shown in FIG. 2 is a device that collects image data.
- the sensing device 10 may be any device as long as it can collect desired data by sensing and transmit it to the server device 100 .
- the sensing device 10 may be a UAV (Unmanned Aerial Vehicle) such as a drone, a moving object such as a vehicle such as an automobile, a camera, or an image sensor (imager), which will be described later in detail. .
- UAV Unmanned Aerial Vehicle
- the sensing device 10 collects data by sensing (step S1).
- the sensing device 10 transmits the collected data (collected data TG) to the server device 100 (step S2).
- the collected data TG includes data DT11, data DT12, etc. collected by the sensing device 10 .
- data DT11 and data DT12 are unlabeled data (unlabeled data).
- the server device 100 Upon receiving the collected data TG from the sensing device 10, the server device 100 calculates the degree of influence of the data (also referred to as "candidate data") in the collected data TG on the learning of the model M1. For example, when each candidate data in the collected data TG is added to the data set DS1, the server device 100 calculates the degree of influence of the added candidate data on the learning of the model M1. The server device 100 calculates the degree of influence of each candidate data in the data set DS1 on the learning of the model M1 using the method of calculating the degree of influence (calculation method MT1). The degree of influence here indicates that the larger the value, the higher the degree (contribution) of the data to the learning of the model M1.
- a larger value of the degree of influence that is, a higher degree of influence, contributes to an improvement in the accuracy of identifying the model M1.
- the higher the degree of influence the more necessary the data is for learning the model M1.
- the higher the degree of influence the more useful the data is for learning the model M1.
- the smaller the value of the influence the lower the degree (contribution) of the data to the learning of the model M1.
- a smaller value of the degree of influence indicates that the model M1 does not contribute to an improvement in the accuracy of identification.
- the lower the degree of influence the more unnecessary the data is for learning the model M1.
- the lower the degree of influence the more harmful the data is to the learning of the model M1.
- FIG. 2 shows a case where the influence function (Influence functions) is used as an example of the calculation method MM1, but the influence function will be described later.
- the calculation method MM1 used by the server device 100 to calculate the degree of influence is not limited to influence function, and any method may be used as long as a value indicating the degree of influence of each data can be obtained.
- the server device 100 may calculate the degree of influence of the image data as a value larger than a predetermined threshold.
- the server device 100 may calculate the degree of influence of the image data as a value equal to or less than a predetermined threshold.
- the server device 100 may calculate the degree of influence of data using a predetermined function.
- the server device 100 may calculate the degree of influence of data using a function that outputs the degree of influence of the image data with a value indicating whether or not the recognition target is included in the image data.
- the server device 100 outputs a value larger than a predetermined threshold when the image data includes the recognition target, and outputs a value equal to or less than the predetermined threshold when the image data does not include the recognition target. may be used to calculate the degree of influence of the data.
- the server device 100 performs the following processing using the calculation method MM1 using the data DT11 and the data DT12 in the collected data TG as candidate data.
- the server device 100 calculates the degree of influence of the data DT11 in the collected data TG on the learning of the model M1 (step S3).
- the server device 100 calculates the degree of influence of the data DT11 on the learning of the model M1 as the degree of influence IV11, as shown in the calculation result RS1.
- the degree of influence IV11 is assumed to be a specific value (for example, 0.3).
- the server device 100 predicts the label of the data DT11 using a classifier.
- the server device 100 predicts the label of the data DT11 using the classifier model M2.
- the server device 100 calculates the influence IV11 of the data DT11 using the data set DS1 to which the predicted labeled data DT11 is added.
- the server device 100 determines the data DT11 based on the degree of influence IV11 of the data DT11 (step S4).
- Server device 100 determines whether data DT11 is necessary for learning model M1 based on influence IV11 of data DT11 and threshold TH1.
- the server device 100 uses the threshold TH1 stored in the threshold information storage unit 123 (see FIG. 7) to determine whether the data DT11 is necessary for learning the model M1.
- the server device 100 compares the degree of influence IV11 of the data DT11 and the threshold TH1, and determines that the data DT11 is unnecessary for learning the model M1 when the degree of influence IV11 is equal to or less than the threshold TH1.
- the server apparatus 100 determines that the data DT11 is unnecessary for learning the model M1 because the influence IV11 of the data DT11 is equal to or less than the threshold TH1. Therefore, as shown in the determination information DR1, the server device 100 determines that the degree of influence of the data DT11 on the learning of the model M1 is low, and deletes the data DT11 (step S5).
- the server device 100 also calculates the degree of influence of the data DT12 in the collected data TG on the learning of the model M1 (step S6).
- the server apparatus 100 calculates the degree of influence of the data DT12 on the learning of the model M1 as the degree of influence IV12, as shown in the calculation result RS2.
- the degree of influence IV12 is assumed to be a specific value (for example, 0.8).
- the server device 100 predicts the label of the data DT12 using a classifier.
- the server device 100 predicts the label of the data DT12 using the model M2, which is a classifier.
- the server device 100 calculates the degree of influence IV12 of the data DT12 using the data set DS1 to which the predicted labeled data DT12 is added.
- the server device 100 determines the data DT12 based on the degree of influence IV12 of the data DT12 (step S7).
- Server device 100 determines whether data DT12 is necessary for learning model M1 based on degree of influence IV12 of data DT12. For example, the server device 100 determines whether the data DT12 is necessary for learning the model M1 based on the influence IV12 of the data DT12 and the threshold TH1.
- the server device 100 compares the degree of influence IV12 of the data DT12 and the threshold TH1, and determines that the data DT12 is necessary for learning the model M1 when the degree of influence IV12 is greater than the threshold TH1. In FIG. 2, the server device 100 determines that the data DT12 is necessary for learning the model M1 because the influence IV12 of the data DT12 is greater than the threshold TH1.
- the server device 100 determines that the degree of influence of the data DT12 on the learning of the model M1 is high, as indicated by the determination information DR2, and adds the data DT12 to the data set DS1 (step S8).
- the server device 100 adds the labeled data DT12 predicted using the model M2 to the data set DS1.
- the server device 100 generates a model using the data set DS1 (step S9).
- the server device 100 generates the model M1 using the data set DS1 including the data DT12 and the like with the prediction label.
- the server device 100 generates the model M1 by small-label learning using the data set DS1 including the data DT12 that was originally unlabeled data in addition to the labeled data.
- the server device 100 designs the structure of a network (neural network, etc.) corresponding to the model M1 stored in the model information storage unit 122 (see FIG. 6).
- the server device 100 designs the network structure (network structure) of the model M1 used for image recognition.
- the server device 100 may generate the network structure of the model M1 used for image recognition based on the information about the network structure corresponding to each application stored in advance in the storage unit 120 (see FIG. 4). .
- the server device 100 may acquire network structure information of the model M1 used for image recognition from an external device.
- the server device 100 learns the model M1 using a data set DS1 in which each data (image) is associated with a label (correct label) indicating the presence or absence of a person.
- the server device 100 uses the data set DS1 to perform learning processing so as to minimize the set loss function (loss function) to learn the model M1.
- the server device 100 learns the model M1 by updating parameters such as weights and biases so that the output layer has correct values for data input. For example, in the error backpropagation method, a loss function that indicates how far the value of the output layer is from the correct state (correct label) is used for the neural network. The weights and biases are updated so that is minimized.
- the server device 100 gives an input value (data) to a neural network (model M1), and the neural network (model M1) calculates a predicted value based on the input value. are compared to evaluate the error.
- the server device 100 executes learning and construction of the model M1 by successively correcting the values of connection weights (synapse coefficients) in the neural network (model M1) based on the obtained error.
- the server apparatus 100 may perform the learning process of the model M1 by various methods.
- the server device 100 may use the model generated at the time of calculating the degree of influence in step S6 as the model M1 learned from the data set DS1 including the data DT12.
- the server device 100 transmits the generated model M1 to the sensing device 10 (step S10).
- the information processing system 1 repeats data collection and model update by repeating the processes of steps S1 to S10.
- the sensing device 10 collects data by sensing using the model M1 received from the server device 100 (step S11).
- the sensing device 10 performs sensing (recognition of a person, etc.) using the model M1, and performs processing such as automatic driving.
- the sensing device 10 transmits data collected by sensing using the model M1 to the server device 100 .
- the server device 100 generates a trained model through a small-label learning process using unlabeled data whose influence is greater than the threshold TH1. That is, the low-label learning process is learning performed using a data group that is not all pre-labeled.
- the small-label learning process uses unlabeled data collected by the sensing device 10 to predict the label of the unlabeled data, attaches the predicted label to the unlabeled data, and treats the unlabeled data as labeled data. This is the learning process to be used.
- the information processing system 1 generates a model using data whose degree of influence is greater than the threshold among the data collected by the sensing device 10 .
- the server device 100 can generate a model using appropriate data by using data with a high degree of influence. Therefore, server device 100 can make available a model generated using appropriate data.
- the information processing system 1 uploads data collected by a sensor device such as the sensing device 10 to the server device 100 .
- the information processing system 1 learns the model and distributes the learned model to sensor devices such as the sensing device 10 .
- the sensor device such as the sensing device 10 performs sensing using the updated trained model.
- the information processing system 1 can update the model and improve the performance of the model by repeating the above-described processing loop at regular intervals.
- the workflow collects new data in the operating environment of the edge device. Upload the data to the server and re-learn. At that time, without learning all the data, the data having a large influence on the model is extracted and learned. Also, since it is based on small-label learning, labels are not required for the data.
- the original model is updated with additional data by transfer learning and delivered to the edge device.
- the information processing system 1 can learn only a small amount of data necessary for learning without labels.
- models that are efficiently calculated and re-learned by the server device 100 are distributed to edge devices such as the sensing device 10 and can be used immediately.
- the model can be automatically grown by repeating the loop at regular time intervals.
- the information processing system 1 can automatically update the trained model. For example, the information processing system 1 uploads images collected by sensing to the server device 100, calculates the degree of influence of the data, and extracts data with a high degree of influence. The information processing system 1 performs transfer learning using the data to update the model. After that, the information processing system 1 distributes the model to the sensing device 10 and updates the learned model. In addition, the information processing system 1 can perform learning in the server device 100 without labels by small-label learning.
- the information processing system 1 calculates the degree of influence of data and learns only data having a high degree of influence on the model.
- the information processing system 1 does not require labels for data due to small-label learning.
- the information processing system 1 can learn only a small amount of data necessary for learning without labels.
- the information processing system 1 uses an influence function (influence function) to formulate the influence of the presence or absence of certain (learning) data on the accuracy (output result) of the model. For example, the information processing system 1 calculates the degree of influence of the added data on learning using a model trained using a data set to which data whose influence is to be calculated is added. Below, the calculation of the degree of influence using the influence function will be described using formulas.
- Influence functions are also used, for example, as a way to explain black-box models of machine learning.
- the influence function is disclosed, for example, in the following documents. ⁇ Understanding Black-box Predictions via Influence Functions, Pang Wei Kho and Percy Liang ⁇ https://arxiv.org/abs/1703.04730>
- the information processing system 1 can calculate the degree of influence of data on machine learning, and can calculate (know) how much positive or negative influence certain data has. can. For example, the information processing system 1 calculates (measures) the degree of impact using an algorithm, data, or the like, as described below. A case where an image is used as input data will be described below as an example.
- each image is labeled, that is, the image is associated with the correct label.
- each labeled image z (sometimes simply referred to as "image z") can be expressed by the following equation ( 1).
- the information processing system 1 uses Equation (3) to calculate a parameter ((the left side of Equation (3))) that minimizes the loss.
- the empirical loss is assumed to be second-differentiable and convex with respect to the parameter ⁇ .
- parameters (variables) with a letter above them such as the parameters (variables) with a hat above " ⁇ " shown on the left side of Equation (3) indicates a predicted value, for example.
- a parameter (variable) with " ⁇ " above " ⁇ ” shown on the left side of Equation (3) in a sentence " ⁇ ”.
- the information processing system 1 calculates parameters (the left side of Equation (4)) when learning is performed using Equation (4) without using certain learning data (image z).
- the impact is the difference (difference) between when training point z (image z) is removed and when there are all data points including training point z. This difference is shown as the following equation (5).
- the information processing system 1 uses the influence functions to perform calculations for the case where the image z is removed by effective approximation, as shown below.
- Equation (7) shows an influence function corresponding to a certain image z.
- equation (7) expresses the amount of change in the parameter with respect to minute ⁇ .
- the information processing system 1 can calculate (determine) the degree of influence when the data point z (image z) is removed.
- the information processing system 1 uses Equations (10-1) to (10-3) below to calculate (determine) the impact on the loss at a test point z test .
- the information processing system 1 can calculate (determine) the degree of influence of data in the machine learning model by this calculation.
- the right side of equation (10-3) consists of a gradient for a certain data loss (loss), an inverse Hessian matrix, a certain learning data loss gradient, and the like.
- the influence of certain data on the prediction (loss) of the model can be obtained by equation (10-3). Note that the above is just an example, and the information processing system 1 may perform various calculations as appropriate to calculate the degree of influence each image has on learning.
- the information processing system 1 may calculate the degree of impact using a technique related to stochastic gradient descent (SGD).
- SGD stochastic gradient descent
- the information processing system 1 may calculate the degree of impact using various methods related to stochastic gradient descent (SGD) disclosed in the following document.
- the information processing system 1 may also calculate the degree of impact using the method disclosed in the following literature.
- the information processing system 1 may also calculate the degree of impact using the method disclosed in the following literature.
- the information processing system 1 may calculate the degree of influence by any method as long as the degree of influence can be calculated.
- the information processing system 1 may reduce the amount of data by arbitrarily adopting a cache reduction method.
- the information processing system 1 may perform cache reduction based on the technique disclosed in the following document. ⁇ Data Cleansing for Deep Neural Networks with Storage-efficient Approximation of Influence Functions, Kenji Suzuki, Yoshiyuki Kobayashi, and Takuya Narihira Wei Kho and Percy Liang ⁇ https://arxiv.org/abs/2103.11807>
- the information processing system 1 calculates the degree of impact of data within the limited HDD capacity. It should be noted that although the system configuration can be achieved without using this cache reduction technique, the amount that can be calculated is limited. Therefore, in this cache reduction, a reduction of 1/1,000 or more can be achieved, and many data influence degrees can be calculated at the time of practical implementation. In the cache reduction method described above, after the calculation, the cache files are reduced, and the data influence degree is calculated one after another. In other words, the above cache reduction technique can calculate more data impact. As a result, the information processing system 1 can efficiently use the finite HDD capacity and calculate the degree of influence of more data.
- the information processing system 1 records learning data logs in the server device 100 . Specifically, the information processing system 1 uses data determined to be necessary for learning after calculating the degree of influence, and stores the data in the server device 100 . In addition, the information processing system 1 also records in the server device 100 which update date and time was used for learning.
- the information processing system 1 adjusts the brightness, contrast, chromaticity, etc. of the image in the edge device in the learning process in the server device 100 .
- the information processing system 1 is provided with a GUI (Graphical User Interface) switch or the like in the device so that the image can be adjusted. Then, the information processing system 1 can generate a more optimized model by calculating the degree of influence of the processed data and performing re-learning by transfer learning.
- GUI Graphic User Interface
- the information processing system 1 shown in FIG. 3 will be described.
- the information processing system 1 is an information processing system that implements adjustment processing for adjusting learning data.
- the information processing system 1 includes a server device 100 and multiple sensing devices 10a, 10b, 10c, and 10d.
- the sensing devices 10a, 10b, 10c, 10d, etc. may be referred to as the sensing device 10 when not distinguished.
- 3 shows four sensing devices 10a, 10b, 10c, and 10d, the information processing system 1 includes more than four sensing devices 10 (for example, 20 or 100 or more).
- the sensing device 10 and the server device 100 are communicably connected by wire or wirelessly via a predetermined communication network (network N).
- FIG. 3 is a diagram illustrating a configuration example of an information processing system according to the embodiment; Note that the information processing system 1 shown in FIG. 3 may include a plurality of server devices 100 .
- the server device 100 is an information processing device (learning) that calculates the degree of influence that data contained in a data set used for learning a model by machine learning has on learning, and learns a model using data whose degree of influence satisfies a condition. equipment).
- the server device 100 also provides the sensing device 10 with a model.
- the sensing device 10 is a computer that provides data to the server device 100 .
- the sensing device 10a is a moving object such as a UAV such as a drone or a vehicle such as an automobile.
- the sensing device 10 a may have a function of communicating with the server device 100 and may move in response to a request from the server device 100 .
- the sensing device 10a has an imaging function such as an image sensor (imager), moves to a position according to a request from the server device 100, captures an image or moving image at that position, and transmits the captured image or moving image to the server device. Send to 100.
- the sensing device 10a and the moving body may be separate bodies.
- the sensing device 10a may be a device mounted on a moving object such as a UAV such as a drone or a vehicle such as an automobile.
- the sensing device 10b is a camera having an imaging function.
- the sensing device 10b is a camera that captures moving images and images and stores captured data.
- the sensing device 10c is an image sensor (imager) having an imaging function.
- the sensing device 10 c has a function of communicating with the server device 100 and has a function of transmitting captured images and moving images to the server device 100 .
- the sensing device 10 c captures an image or moving image in response to a request from the server device 100 and transmits the captured image or moving image to the server device 100 .
- the sensing device 10d is a moving object such as a UAV such as a drone or a vehicle such as an automobile, like the sensing device 10a.
- the information processing system 1 may include a plurality of sensing devices 10 of the same type.
- the information processing system 1 may generate a model based on data collected for each sensing device 10 and provide the model for each sensing device 10 .
- the information processing system 1 may also generate a common model for multiple sensing devices 10 of the same type and provide the common model to multiple sensing devices 10 of the same type.
- the communication function, configuration, etc. of the sensing device 10d are the same as those of the sensing device 10a, so description thereof will be omitted.
- the sensing device 10 may be any device as long as it can implement the processing in the embodiment.
- the sensing device 10 may be, for example, a smartphone, a tablet terminal, a notebook PC (Personal Computer), a desktop PC, a mobile phone, a PDA (Personal Digital Assistant), or other device.
- the sensing device 10 may be a wearable device worn by the user, or the like.
- the sensing device 10 may be a wristwatch-type terminal, a glasses-type terminal, or the like.
- the sensing device 10 may be a so-called home appliance such as a television or a refrigerator.
- the sensing device 10 may be a robot that interacts with humans (users), such as smart speakers, entertainment robots, and household robots.
- the sensing device 10 may be a device such as a digital signage that is placed at a predetermined position.
- FIG. 4 is a diagram showing a configuration example of the server device 100 according to the embodiment of the present disclosure.
- the server device 100 has a communication section 110, a storage section 120, and a control section .
- the server device 100 has an input unit (for example, a keyboard, a mouse, etc.) for receiving various operations from the administrator of the server device 100, and a display unit (for example, a liquid crystal display, etc.) for displaying various information.
- an input unit for example, a keyboard, a mouse, etc.
- a display unit for example, a liquid crystal display, etc.
- the communication unit 110 is implemented by, for example, a NIC (Network Interface Card) or the like.
- the communication unit 110 is connected to the network N (see FIG. 3) by wire or wirelessly, and transmits and receives information to and from other information processing devices such as the sensing device 10 . Also, the communication unit 110 may transmit and receive information to and from the sensing device 10 .
- the storage unit 120 is implemented by, for example, a semiconductor memory device such as RAM (Random Access Memory) or flash memory, or a storage device such as a hard disk or optical disk.
- the storage unit 120 according to the embodiment has a data information storage unit 121, a model information storage unit 122, a threshold information storage unit 123, and a knowledge information storage unit 125, as shown in FIG.
- the data information storage unit 121 stores various information related to data used for learning.
- the data information storage unit 121 stores data sets used for learning.
- FIG. 5 is a diagram illustrating an example of a data information storage unit according to an embodiment of the present disclosure;
- the data information storage unit 121 stores various information related to various data such as learning data used for learning and evaluation data used for accuracy evaluation (calculation).
- FIG. 5 shows an example of the data information storage unit 121 according to the embodiment.
- the data information storage unit 121 includes items such as "data set ID", "data ID”, "data”, “label”, and "date and time”.
- Dataset ID indicates identification information for identifying a dataset.
- Data ID indicates identification information for identifying data.
- Data indicates data identified by a data ID.
- Label indicates the label (correct label) attached to the corresponding data.
- the “label” may be information (correct answer information) indicating the classification (category) of the corresponding data.
- the “label” is correct information (correct label) indicating what kind of object is included in the data (image).
- the label is stored in association with the data.
- a label predicted for that data is stored in association with that data.
- labels shown in parentheses are predicted labels
- labels LB4 to LB8 are predicted labels.
- the server device 100 predicts labels not only for a small number of labeled data, but also for a large number of unlabeled data, and applies the predicted labels to the data. By attaching it, it is used for learning as labeled data.
- “Date and time” indicates the time (date and time) related to the corresponding data.
- “DA1” or the like is shown, but the "date and time” may be a specific date and time such as "15:22:35 on August 1, 2021".
- Information indicating from which model learning the data started to be used may be stored, such as "use started from model learning of version XX”.
- the data set (data set DS1) identified by the data set ID "DS1" includes a plurality of data identified by the data IDs "DID1", “DID2", “DID3”, etc. indicates that For example, each data (learning data) identified by data IDs "DID1", “DID2", “DID3”, etc. is image information or the like used for model learning.
- the data DT1 identified by the data ID "DID1” is labeled data with the label LB1 attached, and indicates that use started from model learning at date and time DA1.
- data DT4 identified by the data ID "DID4" is data collected as unlabeled data and attached with the label LB4, which is a prediction label, and is used starting from model learning at date and time DA4. indicates that the
- the data information storage unit 121 may store various types of information, not limited to the above, depending on the purpose.
- the data information storage unit 121 may store data such as whether each data is learning data or evaluation data so as to be identifiable.
- the data information storage unit 121 stores learning data and evaluation data in a distinguishable manner.
- the data information storage unit 121 may store information identifying whether each data is learning data or evaluation data.
- the server device 100 learns a model based on each data used as learning data and the correct answer information.
- the server device 100 calculates the accuracy of the model based on each data used as the evaluation data and the correct answer information.
- the server device 100 calculates the accuracy of the model by collecting the result of comparing the output result output by the model when the evaluation data is input with the correct answer information.
- the model information storage unit 122 stores information about models.
- the model information storage unit 122 stores information (model data) indicating the structure of a model (network).
- FIG. 6 is a diagram illustrating an example of a model information storage unit according to an embodiment of the present disclosure; FIG. 6 shows an example of the model information storage unit 122 according to the embodiment.
- the model information storage unit 122 includes items such as "model ID", "usage", and "model data”.
- Model ID indicates identification information for identifying a model.
- User indicates the use of the corresponding model.
- Model data indicates model data.
- FIG. 6 shows an example in which conceptual information such as “MDT1" is stored in “model data”, but in reality, various types of information that make up the model, such as network information and functions included in the model, are stored. included.
- model M1 identified by the model ID "M1" indicates that the application is "image recognition”.
- Model M1 indicates that it is a model used for image recognition. It also indicates that the model data of the model M1 is the model data MDT1.
- model M2 identified by the model ID "M2" indicates that the application is "label prediction”.
- Model M2 indicates that it is a model used for label prediction.
- model M2 is a classifier to predict labels for unlabeled data. It also indicates that the model data of the model M2 is the model data MDT2.
- model information storage unit 122 may store various types of information, not limited to the above, depending on the purpose.
- the model information storage unit 122 stores parameter information of the model learned (generated) by the learning process.
- the threshold information storage unit 123 stores various information regarding thresholds.
- the threshold information storage unit 123 stores various information regarding thresholds used for comparison with scores.
- 7 is a diagram illustrating an example of a threshold information storage unit according to the embodiment; FIG.
- the threshold information storage unit 123 shown in FIG. 7 includes items such as "threshold ID", "usage”, and "threshold”.
- Threshold ID indicates identification information for identifying the threshold.
- User indicates the usage of the threshold.
- Threshold indicates a specific value of the threshold identified by the corresponding threshold ID.
- the threshold (threshold TH1) identified by the threshold ID "TH1" is stored in association with information indicating that it is used to determine the degree of impact. That is, the threshold TH1 is used to determine whether or not the target data has a high degree of influence. Data determined to have a high degree of influence are added to the data set, and data determined to have a low degree of influence are deleted. Also, the value of the threshold TH1 indicates that it is "VL1". In the example of FIG. 7, the value of the threshold TH1 is a specific numerical value (for example, 0.6), although it is indicated by an abstract code such as "VL1".
- the threshold information storage unit 123 may store various types of information, not limited to the above, depending on the purpose.
- the control unit 130 stores a program (for example, an information processing program such as a learning processing program according to the present disclosure) stored inside the server device 100 by, for example, a CPU (Central Processing Unit) or an MPU (Micro Processing Unit). (Random Access Memory) etc. are executed as a work area. Also, the control unit 130 is a controller, and is implemented by an integrated circuit such as an ASIC (Application Specific Integrated Circuit) or an FPGA (Field Programmable Gate Array).
- ASIC Application Specific Integrated Circuit
- FPGA Field Programmable Gate Array
- the control unit 130 has an acquisition unit 131, a calculation unit 132, a data management unit 133, a compensation unit 134, a prediction unit 135, a learning unit 136, and a transmission unit 137. , implements or performs the information processing functions and actions described below.
- the internal configuration of the control unit 130 is not limited to the configuration shown in FIG. 4, and may be another configuration as long as it performs information processing to be described later.
- the connection relationship of each processing unit of the control unit 130 is not limited to the connection relationship shown in FIG. 4, and may be another connection relationship.
- the acquisition unit 131 acquires various types of information. Acquisition unit 131 acquires various types of information from storage unit 120 . The acquisition unit 131 acquires various types of information from the data information storage unit 121 , the model information storage unit 122 and the threshold information storage unit 123 .
- the acquisition unit 131 receives various information from an external information processing device.
- the acquisition unit 131 receives various information from the sensing device 10 .
- the acquisition unit 131 acquires various information calculated by the calculation unit 132 .
- the acquisition unit 131 acquires various types of information corrected by the correction unit 134 .
- the acquisition unit 131 acquires various information predicted by the prediction unit 135 .
- the acquisition unit 131 acquires various information learned by the learning unit 136 .
- the calculation unit 132 calculates various processes.
- the calculation unit 132 calculates the degree of influence of learning data used for learning of the neural network on learning.
- the calculator 132 calculates various processes based on information from an external information processing device.
- the calculation unit 132 calculates various processes based on information stored in the storage unit 120 .
- the calculation unit 132 calculates various processes based on information stored in the data information storage unit 121, the model information storage unit 122, and the threshold information storage unit 123.
- FIG. The calculator 132 generates various types of information by calculating the processing.
- the calculation unit 132 calculates various processes based on various information acquired by the acquisition unit 131 .
- the calculation unit 132 calculates the degree of influence of the data collected by the sensing device 10 on model learning by machine learning.
- the calculator 132 calculates the degree of impact based on the loss function.
- the calculation unit 132 calculates the degree of influence using the influence function.
- the calculator 132 calculates the degree of influence of image data collected by the image sensor.
- the calculation unit 132 calculates the degree of influence of data collected by the sensing device 10, which is an external device, using the learned model.
- the data management unit 133 executes various processes related to data management.
- the data management unit 133 determines data.
- the data management unit 133 makes judgments on data collected by the sensing device 10 .
- the data management unit 133 determines the necessity of each data based on the degree of influence of each data.
- the data management unit 133 deletes data whose degree of influence does not satisfy the conditions.
- the data management unit 133 stores the data whose degree of influence satisfies the condition as a log in the storage unit 120 .
- the data management unit 133 adds data to the data set based on the calculation results of the calculation unit 132.
- the data management unit 133 adds data determined to have a high degree of influence to the data set.
- the data management unit 133 adds the target data with the prediction labels predicted by the prediction unit 135 to the data set, with the data determined to have a high degree of influence as the target data.
- the data management unit 133 associates the data determined to have a high degree of influence with the predicted label of the data and adds them to the data set.
- the compensation unit 134 compensates various data.
- the compensator 134 compensates the data collected by the sensing device 10 .
- the compensator 134 compensates the image data collected by the sensing device 10 .
- the compensation unit 134 compensates the image data by adjusting the brightness of the image data.
- the compensation unit 134 compensates the image data by adjusting the contrast of the image data.
- the compensation unit 134 compensates the image data by adjusting the chromaticity of the image data.
- the compensator 134 compensates the image data according to the image sensor of the sensing device 10 .
- the correction unit 134 corrects the image data according to the correction content corresponding to the image sensor of the sensing device 10 in the list information, using the list information indicating the correction content for each image sensor.
- the prediction unit 135 predicts various information.
- the prediction unit 135 predicts labels of data.
- the prediction unit 135 predicts the label of unlabeled data, which is unlabeled data.
- the prediction unit 135 predicts the predicted label of the unlabeled data using a classifier that has been trained on the labeled data set.
- the prediction unit 135 predicts the predicted label of data using the model M2, which is a classifier used for label prediction.
- the prediction unit 135 inputs data to be predicted (prediction target data) to the model M2, and uses the output of the model M2 to predict the prediction label of the prediction target data.
- the prediction unit 135 predicts the classification result of the prediction target data output by the model M2 as the prediction label of the prediction target data.
- the learning unit 136 learns various types of information.
- the learning unit 136 learns various types of information based on information from an external information processing device and information stored in the storage unit 120 .
- the learning section 136 learns various types of information based on the information stored in the data information storage section 121 .
- the learning unit 136 stores the model generated by learning in the model information storage unit 122 .
- the learning unit 136 stores the model updated by learning in the model information storage unit 122 .
- the learning unit 136 performs learning processing.
- the learning unit 136 performs various types of learning.
- the learning unit 136 learns various types of information based on the information acquired by the acquisition unit 131 .
- the learning unit 136 learns (generates) a model.
- the learning unit 136 learns various information such as models.
- the learning unit 136 generates a model through learning.
- the learning unit 136 learns the model using various machine learning techniques. For example, the learning unit 136 learns model (network) parameters.
- the learning unit 136 learns the model using various machine learning techniques.
- the learning unit 136 generates a model M1. Also, the learning unit 136 generates a model M2. The learning unit 136 learns network parameters. For example, the learning unit 136 learns network parameters of the model M1. Also, the learning unit 136 learns parameters of the network of the model M2.
- the learning unit 136 performs learning processing based on the learning data (teacher data) stored in the data information storage unit 121.
- the learning unit 136 generates the model M1 by performing learning processing using the learning data stored in the data information storage unit 121 .
- the learning unit 136 generates models used for image recognition.
- the learning unit 136 generates the model M1 by learning the parameters of the network of the model M1.
- the learning unit 136 generates the model M2 by learning the network parameters of the model M2.
- the method of learning by the learning unit 136 is not particularly limited. You can learn. Also, for example, a technique based on DNN (Deep Neural Network) such as CNN (Convolutional Neural Network) and 3D-CNN may be used.
- the learning unit 136 uses a recurrent neural network (RNN) or LSTM (Long Short-Term Memory units), which is an extension of RNN, when targeting time-series data such as moving images (moving images) such as videos. You may use the method based on.
- RNN recurrent neural network
- LSTM Long Short-Term Memory units
- the learning unit 136 generates a trained model through a small-label learning process in which the model is learned using data whose degree of influence calculated by the calculation unit 132 satisfies the conditions.
- the learning unit 136 performs a small-label learning process using data whose degree of influence is greater than a predetermined threshold.
- the learning unit 136 performs a small-label learning process using the predicted label predicted by the prediction unit 135 and the target data, with the unlabeled data whose influence degree satisfies the condition as the target data.
- the learning unit 136 generates a trained model using the data set to which the predicted labeled target data is added.
- the learning unit 136 executes learning processing using the data set.
- the learning unit 136 performs small-label learning processing using image data whose degree of influence satisfies the conditions.
- the learning unit 136 performs a small-label learning process using corrected image data obtained by correcting the image data whose degree of influence satisfies the condition.
- the learning unit 136 updates the learned model using data whose degree of influence calculated by the calculation unit 132 satisfies the condition.
- the transmission unit 137 transmits various types of information.
- the transmission unit 137 transmits various types of information to an external information processing device.
- the transmission unit 137 provides various types of information to an external information processing device.
- the transmission unit 137 transmits various information to other information processing apparatuses such as the sensing device 10 .
- the transmitter 137 provides information stored in the storage 120 .
- Transmitter 137 transmits information stored in storage 120 .
- the transmission unit 137 provides various information based on information from other information processing devices such as the sensing device 10 .
- the transmitting section 137 provides various information based on the information stored in the storage section 120 .
- the transmission unit 137 provides various information based on the information stored in the data information storage unit 121, the model information storage unit 122, and the threshold information storage unit 123. FIG.
- the transmission unit 137 transmits the trained model generated by the learning unit 136 to the sensing device 10.
- the transmission unit 137 transmits the model M ⁇ b>1 that is the generated trained model to the sensing device 10 .
- the transmission unit 137 transmits the trained model updated by the learning unit 136 to the sensing device 10 .
- the transmitter 137 transmits the updated model M1 to the sensing device 10 .
- the server device 100 may use a model (network) in the form of a neural network (NN) such as a deep neural network (DNN).
- NN neural network
- DNN deep neural network
- the server device 100 is not limited to a neural network, and may use various types of models (functions) such as regression models such as SVM (Support Vector Machine).
- the server device 100 may use any type of model (function).
- the server device 100 may use various regression models such as a nonlinear regression model and a linear regression model.
- FIG. 8 is a diagram of an example of a network corresponding to the model.
- a network NW1 shown in FIG. 8 represents a neural network including a plurality of (multilayer) intermediate layers between an input layer INL and an output layer OUTL.
- a network NW1 shown in FIG. 8 corresponds to the neural network NN in FIG.
- the server device 100 may learn the parameters of the network NW1 shown in FIG.
- a network NW1 shown in FIG. 8 corresponds to the network of model M1 and is a conceptual diagram showing a neural network (model) used for image recognition. For example, when an image is input from the input layer INL side, the network NW1 outputs the recognition result from the output layer OUTL. For example, the server device 100 outputs recognition results corresponding to the input from the output layer OUTL by inputting information to the input layer INL in the network NW1.
- FIG. 8 shows the network NW1 as an example of a model (network)
- the network NW1 may be of various types depending on the application.
- the server device 100 learns the model M1 by learning the parameters (weights) of the model M1 having the structure of the network NW1 shown in FIG.
- FIG. 9 is a diagram illustrating a configuration example of a sensing device according to an embodiment of the present disclosure.
- the sensing device 10 has a communication section 11, an input section 12, an output section 13, a storage section 14, a control section 15, and a sensor section 16.
- the sensing device 10 may have any device configuration as long as it can collect data and provide it to the server device 100 .
- the sensing device 10 may have any other configuration as long as it has a communication unit 11 that communicates with the server device 100 and a control unit 15 that performs processing for collecting data.
- the sensing device 10 may not have any of the input section 12 , the output section 13 , the storage section 14 , and the sensor section 16 .
- the sensing device 10 may be configured to have only the communication section 11, the control section 15, and the sensor section 16.
- an imaging device used in an image sensor (imager) is a CMOS (Complementary Metal Oxide Semiconductor).
- the imaging element used in the image sensor (imager) is not limited to CMOS, and may be various imaging elements such as CCD (Charge Coupled Device).
- the sensing device 10 when the sensing device 10 is a data server, the sensing device 10 may be configured to have only the communication unit 11 , the storage unit 14 and the control unit 15 .
- the sensing device 10 when the sensing device 10 is a moving body, the sensing device 10 may be configured to have a mechanism for realizing movement, such as a drive section (motor).
- the communication unit 11 is implemented by, for example, a NIC, a communication circuit, or the like.
- the communication unit 11 is connected to a network N (Internet or the like) by wire or wirelessly, and transmits and receives information to and from other devices such as the server device 100 via the network N.
- a network N Internet or the like
- the input unit 12 accepts various inputs.
- the input unit 12 receives a user's operation.
- the input unit 12 may accept an operation (user operation) to the sensing device 10 used by the user as an operation input by the user.
- the input unit 12 may receive, via the communication unit 11, information regarding user operations using a remote controller (remote controller).
- the input unit 12 may also have buttons provided on the sensing device 10 and a keyboard and mouse connected to the sensing device 10 .
- the input unit 12 may have a touch panel capable of realizing functions equivalent to those of a remote controller, keyboard, or mouse.
- various information is input to the input unit 12 via the display (output unit 13).
- the input unit 12 receives various operations from the user via the display screen using a touch panel function realized by various sensors. That is, the input unit 12 receives various operations from the user via the display (output unit 13) of the sensing device 10.
- FIG. For example, the input unit 12 receives a user's operation via the display (output unit 13) of the sensing device 10.
- the output unit 13 outputs various information.
- the output unit 13 has a function of displaying information.
- the output unit 13 is provided in the sensing device 10 and displays various information.
- the output unit 13 is realized by, for example, a liquid crystal display or an organic EL (Electro-Luminescence) display.
- the output unit 13 may have a function of outputting sound.
- the output unit 13 has a speaker that outputs audio.
- the storage unit 14 is implemented, for example, by a semiconductor memory device such as a RAM or flash memory, or a storage device such as a hard disk or optical disk.
- the storage unit 14 stores various information necessary for collecting data.
- the storage unit 14 has a model information storage unit 141 .
- the model information storage unit 141 stores information (model data) indicating the structure of the model (network).
- FIG. 10 is a diagram illustrating an example of a model information storage unit according to an embodiment of the present disclosure; FIG. 10 shows an example of the model information storage unit 141 according to the embodiment.
- the model information storage unit 141 includes items such as "model ID", "usage", and "model data”.
- Model ID indicates identification information for identifying a model.
- User indicates the use of the corresponding model.
- Model data indicates model data.
- FIG. 10 shows an example in which conceptual information such as “MDT1" is stored in “model data”, but in reality, various types of information that make up the model, such as network information and functions included in the model, are stored. included.
- model M1 identified by the model ID "M1" indicates that the application is "image recognition”.
- Model M1 indicates that it is a model used for image recognition. It also indicates that the model data of the model M1 is the model data MDT1.
- model information storage unit 141 may store various types of information, not limited to the above, depending on the purpose.
- the model information storage unit 141 stores parameter information of the model learned (generated) by the learning process.
- the control unit 15 executes, for example, a program stored inside the sensing device 10 (for example, an information processing program such as a data providing program according to the present disclosure) using a RAM or the like as a work area by a CPU, an MPU, or the like. Realized. Also, the control unit 15 is a controller, and may be realized by an integrated circuit such as an ASIC or FPGA, for example.
- control unit 15 has a receiving unit 151, a collecting unit 152, and a transmitting unit 153, and implements or executes the information processing functions and actions described below.
- the internal configuration of the control unit 15 is not limited to the configuration shown in FIG. 9, and may be another configuration as long as it performs the information processing described later.
- the receiving unit 151 receives various information.
- the receiving unit 151 receives various information from an external information processing device.
- the receiving unit 151 receives various information from other information processing devices such as the server device 100 .
- the receiving unit 151 receives from the server device 100 the trained model learned by the server device 100 .
- the receiving unit 151 receives from the server device 100 a learned model updated using data collected by the sensing device 10 through sensing using the learned model.
- the receiving unit 151 receives from the server device 100 a trained model that has been trained by the server device 100 using image data.
- the collection unit 152 collects various types of information.
- the collection unit 152 determines collection of various information.
- the collection unit 152 collects various types of information based on information from an external information processing device.
- the collection unit 152 collects various types of information based on the information stored in the storage unit 14 .
- the collection unit 152 collects data by sensing using the model M1 stored in the model information storage unit 141 .
- the collection unit 152 collects data by sensing using the trained model.
- the collection unit 152 collects data by sensing using the trained model updated by the server device 100 .
- the collection unit 152 collects image data detected by the sensor unit 16 .
- the collection unit 152 collects image data by sensing using the learned model.
- the transmission unit 153 transmits various types of information to an external information processing device. For example, the transmission unit 153 transmits various information to other information processing devices such as the server device 100 . The transmission unit 153 transmits information stored in the storage unit 14 .
- the transmission unit 153 transmits various information based on information from other information processing devices such as the server device 100 .
- the transmission unit 153 transmits various information based on the information stored in the storage unit 14 .
- the transmission unit 153 generates a trained model by a small label learning process of learning a model using data collected by sensing when the degree of influence of the data on model learning by machine learning satisfies a condition. It transmits to the server device 100 .
- the transmission unit 153 transmits data collected by the collection unit 152 through sensing using the learned model to the server device 100 .
- the transmission unit 153 transmits image data collected by sensing to the server device 100 .
- the transmission unit 153 transmits image data detected by the image sensor (image sensor) of the sensor unit 16 to the server device 100 .
- the sensor unit 16 detects various sensor information.
- the sensor unit 16 has a function as an imaging unit that captures an image.
- the sensor unit 16 has the function of an image sensor and detects image information.
- the sensor unit 16 functions as an image input unit that receives an image as an input.
- the sensor unit 16 is not limited to the above, and may have various sensors.
- the sensor unit 16 includes a sound sensor, a position sensor, an acceleration sensor, a gyro sensor, a temperature sensor, a humidity sensor, an illuminance sensor, a pressure sensor, a proximity sensor, and sensors for receiving biological information such as odor, sweat, heartbeat, pulse, and brain waves. It may have various sensors such as sensors. Further, the sensors for detecting the above various information in the sensor unit 16 may be a common sensor, or may be implemented by different sensors.
- FIG. 11 is a flow chart showing processing of the learning device according to the embodiment of the present disclosure. Specifically, FIG. 11 is a flowchart showing the procedure of information processing by the server device 100, which is an example of the learning device.
- the server device 100 performs processing using the data ULD collected by the sensing device 10 .
- the server device 100 receives data ULD from the sensing device 10 .
- the server device 100 calculates the degree of influence of data (step S101). For example, the server device 100 calculates the degree of influence of each piece of data in the data ULD.
- the server device 100 generates a trained model through a small-label learning process using data with a high degree of data influence (step S102). For example, the server device 100 generates a trained model by a small label learning process using data having a high degree of data influence among the data ULD. The server device 100 predicts the label of data with a high degree of influence, and uses the predicted label and the data with a high degree of influence to generate a trained model.
- FIG. 12 is a sequence diagram showing processing procedures of the information processing system according to the embodiment of the present disclosure.
- the sensing device 10 collects data by sensing (step S201). The sensing device 10 then transmits the collected data to the server device 100 (step S202).
- the server device 100 calculates the degree of influence of each data collected by the sensing device 10 (step S203).
- the server device 100 deletes data with a low degree of influence (step S204). For example, the server device 100 deletes data whose degree of influence is equal to or less than a threshold among the data collected by the sensing device 10 and does not store the data in the storage unit 120 .
- the server device 100 adds data with a high degree of influence to the data set (step S205).
- the server device 100 adds data with a degree of influence greater than the threshold to the data set used for learning.
- the server device 100 generates a model through a small-label learning process using a data set to which data with a high degree of influence have been added. (Step S206). For example, the server device 100 predicts a label for data with a high degree of influence as target data, attaches the predicted label to the target data, and generates a model using a data set to which the target data is added.
- the server device 100 transmits the generated model to the sensing device 10 (step S207). Then, the sensing device 10 updates the model in its own device to the model received from the server device 100 (step S208).
- the sensing device 10 collects data by sensing using the updated model (step S209).
- the sensing device 10 then transmits the collected data to the server device 100 (step S210).
- the information processing system 1 repeats data collection and model update by repeating the processes of steps S203 to S210.
- the server device 100 calculates the degree of impact of collected data using a model updated by the sensing device 10 .
- the server device 100 updates the model using data whose degree of influence satisfies the condition.
- Server device 100 transmits the updated model to sensing device 10 .
- the sensing device 10 then collects data by sensing using the updated model.
- the server device 100 and the sensing device 10 are separate units, that is, the learning device for learning the model and the device for sensing data are separate units.
- the sensing device 10 may be a learning device (information processing device) having a function of collecting data by sensing and a function of learning a model.
- the sensing device 10 has various configurations (for example, the calculation unit 132, the learning unit 136, etc.) for learning the model of the server device 100 described above, and generates the model using the data collected by the device itself.
- the sensing device 10 may be a camera, smartphone, television, automobile, drone, robot, or the like. In this way, the sensing device 10 may be a terminal device (computer) that autonomously collects highly influential learning data and generates a model.
- each component of each device illustrated is functionally conceptual and does not necessarily need to be physically configured as illustrated.
- the specific form of distribution and integration of each device is not limited to the one shown in the figure, and all or part of them can be functionally or physically distributed and integrated in arbitrary units according to various loads and usage conditions. Can be integrated and configured.
- the learning device (the server device 100 in the embodiment) according to the present disclosure includes the calculation unit (the calculation unit 132 in the embodiment) and the learning unit (the learning unit 136 in the embodiment).
- the calculation unit calculates the degree of influence of data collected by the sensing device (sensing device 10 in the embodiment) on model learning by machine learning.
- the learning unit generates a trained model through a small-label learning process of learning the model using data whose degree of influence calculated by the calculation unit satisfies a condition.
- the learning device executes the small-label learning process using the data that satisfies the condition of the degree of influence on model learning among the data collected by the sensing device, and generates the model.
- the learning device can generate a model using appropriate data by using data whose degree of influence satisfies the condition. Therefore, the learning device can make available the model generated using the appropriate data.
- the calculation unit performs small-label learning processing using data whose degree of influence is greater than a predetermined threshold.
- the learning device executes the small-label learning process using data with a degree of influence greater than a predetermined threshold, ie, data with a high degree of influence, and generates a model.
- the learning device can generate a model using appropriate data by using data with a high degree of influence.
- the calculation unit calculates the degree of impact based on the loss function. In this way, the learning device can accurately calculate the influence of each data by calculating the influence based on the loss function. Therefore, the learning device can generate a model using appropriate data.
- the calculation unit calculates the degree of influence using the influence function. In this way, the learning device can accurately calculate the influence of each data by calculating the influence using the influence function. Therefore, the learning device can generate a model using appropriate data.
- the learning device includes a prediction unit (the prediction unit 135 in the embodiment).
- the predictor predicts the label of unlabeled data, which is unlabeled data.
- the learning unit performs a small-label learning process using the predicted label predicted by the prediction unit and the target data, with the unlabeled data whose influence degree satisfies the condition as the target data.
- the learning device uses unlabeled data whose influence degree satisfies the condition as target data, and performs the low-label learning process using the target data and predicted labels predicted, thereby using unlabeled data as well. can be used to generate the model.
- the prediction unit predicts the predicted label of the target data using a classifier learned from the labeled data set.
- the learning unit generates a trained model using the data set to which the predicted labeled target data is added. In this way, the learning device adds the data set to which the target data with the prediction label is added, and generates the trained model using the data, thereby generating the model using appropriate data. can be done.
- the learning device includes a data management unit (data management unit 133 in the embodiment).
- the data management unit deletes data whose influence degree does not satisfy the condition, and stores data whose influence degree satisfies the condition as a log in the storage unit. In this way, the learning device can reduce the amount of data to be stored in the storage unit by deleting data whose degree of influence does not satisfy the condition.
- the learning device manages the data used for learning by storing the data whose degree of influence satisfies the conditions as a log in the storage unit (storage unit 120 in the embodiment), and uses the data for model generation as necessary. It can be possible to explain the model, such as presenting the data obtained.
- the calculation unit calculates the degree of influence of the image data collected by the image sensor.
- the learning unit performs a small-label learning process using image data whose degree of influence satisfies a condition.
- the learning device executes the small-label learning process using image data that satisfies the condition of the degree of influence on model learning among the image data collected by the sensing device, and generates a model.
- the learning device can generate a model using appropriate image data by using image data whose degree of influence satisfies the condition.
- the learning unit performs a small-label learning process using the corrected image data obtained by correcting the image data whose degree of influence satisfies the condition.
- the learning device can generate a model using appropriate image data by generating a model using the corrected image data.
- the learning device includes a transmission unit (transmission unit 137 in the embodiment).
- the transmission unit transmits the trained model generated by the learning unit to an external device (sensing device 10 in the embodiment). In this way, the learning device can make available the model generated using appropriate data by transmitting the generated model to the external device.
- the calculation unit calculates the degree of influence of the data collected by the sensing device, which is an external device, using the trained model.
- the learning unit updates the learned model using data whose degree of influence calculated by the calculation unit satisfies the condition. In this way, the learning device can appropriately update the model using the data collected using the generated model. As a result, the learning device can update the model and improve the accuracy (performance) of the model by repeating this loop at regular intervals.
- the transmission unit transmits the learned model updated by the learning unit to the sensing device.
- the learning device can cause the sensing device to perform processing using the updated model by transmitting the updated model to the sensing device. Therefore, the learning device can make available the model generated using the appropriate data.
- the learning device is a server device that provides a model to the sensing device.
- the learning device can make available a model generated using appropriate data. can.
- the sensing device (the sensing device 10 in the embodiment) according to the present disclosure includes a transmitter (the transmitter 153 in the embodiment), a receiver (the receiver 151 in the embodiment), and a collector (the Then, a collection unit 152) is provided.
- the transmission unit uses the data collected by sensing to generate a trained model through a small-label learning process that trains a model using the data when the degree of influence of the data on the learning of the model by machine learning satisfies the conditions. It transmits to the device (the server device 100 in the embodiment).
- the receiving unit receives the trained model trained by the learning device from the learning device.
- the collecting unit collects data by sensing using the trained model.
- the sensing device transmits collected data to the learning device, performs the small-label learning process using the data that satisfies the condition of the degree of influence exerted by the learning device on the model learning, and generates the generated data. receive the model. Then, the sensing device collects data by sensing using the model. Thereby, the sensing device can collect data using a model generated using the data collected by the sensing device. Thus, the sensing device can make available the model generated with the appropriate data.
- the transmission unit transmits data collected by the collection unit through sensing using the trained model to the learning device.
- the sensing device provides the learning device with data collected using the model generated by the learning device, thereby enabling the learning device to update the model using the data. .
- the sensing device can make available the model generated with the appropriate data.
- the receiving unit receives from the learning device the learned model updated using the data collected by the sensing device through sensing using the learned model.
- the collecting unit collects data by sensing using the trained model updated by the learning device.
- the sensing device can collect data using a model updated using the data collected by the sensing device.
- the sensing device can make available the model generated with the appropriate data.
- the collection unit also collects image data detected by the sensor unit (sensor unit 16 in the embodiment).
- the sensing device can collect image data to enable the learning device to update the model using the image data.
- the sensing device can make available the model generated with the appropriate data.
- the transmission unit transmits image data collected by sensing to the learning device.
- the receiving unit receives, from the learning device, a trained model trained by the learning device using the image data.
- the collection unit collects image data by sensing using the trained model.
- the sensing device transmits collected image data to the learning device and receives a model generated by the learning device using the image data. Then, the sensing device collects image data by sensing using the model. Thereby, the sensing device can collect image data using a model generated using the image data collected by the sensing device. Thus, the sensing device can make available the model generated with the appropriate data.
- FIG. 13 is a hardware configuration diagram showing an example of a computer 1000 that realizes functions of information processing apparatuses such as the server apparatus 100 and the sensing device 10.
- the computer 1000 has a CPU 1100 , a RAM 1200 , a ROM (Read Only Memory) 1300 , a HDD (Hard Disk Drive) 1400 , a communication interface 1500 and an input/output interface 1600 .
- Each part of computer 1000 is connected by bus 1050 .
- the CPU 1100 operates based on programs stored in the ROM 1300 or HDD 1400 and controls each section. For example, the CPU 1100 loads programs stored in the ROM 1300 or HDD 1400 into the RAM 1200 and executes processes corresponding to various programs.
- the ROM 1300 stores a boot program such as BIOS (Basic Input Output System) executed by the CPU 1100 when the computer 1000 is started, and programs dependent on the hardware of the computer 1000.
- BIOS Basic Input Output System
- the HDD 1400 is a computer-readable recording medium that non-temporarily records programs executed by the CPU 1100 and data used by such programs.
- HDD 1400 is a recording medium that records an information processing program according to the present disclosure, which is an example of program data 1450 .
- a communication interface 1500 is an interface for connecting the computer 1000 to an external network 1550 (for example, the Internet).
- CPU 1100 receives data from another device via communication interface 1500, and transmits data generated by CPU 1100 to another device.
- the input/output interface 1600 is an interface for connecting the input/output device 1650 and the computer 1000 .
- the CPU 1100 receives data from input devices such as a keyboard and mouse via the input/output interface 1600 .
- the CPU 1100 also transmits data to an output device such as a display, speaker, or printer via the input/output interface 1600 .
- the input/output interface 1600 may function as a media interface for reading a program or the like recorded on a predetermined recording medium.
- Media include, for example, optical recording media such as DVD (Digital Versatile Disc) and PD (Phase change rewritable disk), magneto-optical recording media such as MO (Magneto-Optical disk), tape media, magnetic recording media, semiconductor memories, etc. is.
- the CPU 1100 of the computer 1000 implements the functions of the control unit 130 and the like by executing the information processing program loaded on the RAM 1200.
- the HDD 1400 also stores an information processing program according to the present disclosure and data in the storage unit 120 .
- CPU 1100 reads and executes program data 1450 from HDD 1400 , as another example, these programs may be obtained from another device via external network 1550 .
- the present technology can also take the following configuration.
- a calculation unit that calculates the degree of influence of the data collected by the sensing device on model learning by machine learning; a learning unit that generates a trained model by a small-label learning process of learning the model using data that satisfies the condition of the degree of influence calculated by the calculation unit;
- a learning device with (2) The learning unit The learning device according to (1), wherein the low-label learning process is performed using data having the degree of influence greater than a predetermined threshold.
- the calculation unit The learning device according to (1) or (2), wherein the degree of influence is calculated based on a loss function.
- the calculation unit The learning device according to any one of (1) to (3), wherein the degree of influence is calculated by influence functions.
- a prediction unit that predicts the label of unlabeled data that is unlabeled data; further comprising The learning unit any one of (1) to (4), wherein the small label learning process is performed using the predicted labels predicted by the prediction unit and the target data, with unlabeled data satisfying the influence degree satisfying the condition as target data; A learning device as described.
- the prediction unit Predicting the predicted label of the target data using a classifier learned from a dataset of labeled data that has been labeled;
- the learning unit The learning device according to (5), wherein the trained model is generated using a data set to which the target data with the prediction label is added.
- a data management unit that deletes data whose influence degree does not satisfy the condition and stores data whose influence degree satisfies the condition as a log in a storage unit;
- the learning device according to any one of (1) to (6), further comprising: (8) The calculation unit calculating the degree of influence of image data collected by an image sensor; The learning unit The learning device according to any one of (1) to (7), wherein the low-label learning process is performed using image data whose degree of influence satisfies a condition. (9) The learning unit The learning device according to (8), wherein the low-label learning process is performed using corrected image data obtained by correcting the image data satisfying the influence degree condition.
- (10) a transmitting unit that transmits the trained model generated by the learning unit to an external device;
- the learning device according to any one of (1) to (9), further comprising: (11) The calculation unit calculating the degree of influence of data collected by the sensing device, which is the external device, using the trained model; The learning unit The learning device according to (10), wherein the learned model is updated using data whose degree of influence calculated by the calculator satisfies a condition. (12) The transmission unit (11), wherein the learned model updated by the learning unit is transmitted to the sensing device. (13) The learning device according to (11) or (12), which is a server device that provides a model to the sensing device.
- a transmitter for transmitting a receiving unit that receives the trained model learned by the learning device from the learning device; a collection unit that collects data by sensing using the trained model; Sensing device with (16) The transmission unit (15) The sensing device according to (15), wherein the collection unit transmits data collected by sensing using the trained model to the learning device. (17) The receiving unit the sensing device receives the learned model updated using data collected by sensing using the learned model from the learning device; The collection unit is (16), wherein data is collected by sensing using the learned model updated by the learning device. (18) The collection unit is The sensing device according to any one of (15) to (17), which collects image data detected by the sensor unit.
- the transmission unit transmits image data collected by sensing to the learning device, The receiving unit receiving the trained model trained by the learning device using image data from the learning device; The collection unit is (18), wherein image data is collected by sensing using the trained model.
- a learning device that generates a trained model by a small label learning process that uses data collected by sensing to learn the model when the degree of influence of the data on the learning of the model by machine learning satisfies a condition. send and receiving the learned model learned by the learning device from the learning device; collecting data by sensing using the trained model; The data collection method that performs the processing.
- 1 information processing system 100 server device (learning device) 110 communication unit 120 storage unit 121 data information storage unit 122 model information storage unit 123 threshold information storage unit 130 control unit 131 acquisition unit 132 calculation unit 133 data management unit 134 correction unit 136 learning unit 137 transmission unit 10 sensing device 11 communication unit 12 Input unit 13 Output unit 14 Storage unit 141 Model information storage unit 15 Control unit 151 Reception unit 152 Collection unit 153 Transmission unit 16 Sensor unit
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Software Systems (AREA)
- Medical Informatics (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Databases & Information Systems (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
Description
1.実施形態
1-1.本開示の実施形態に係る学習処理の概要
1-1-1.背景及び効果等
1-1-2.Influence function(影響関数)
1-1-2-1.その他の手法例
1-1-3.容量制限
1-1-4.保管ストレージ
1-1-5.画像補整
1-2.実施形態に係る情報処理システムの構成
1-3.実施形態に係る学習装置の構成
1-3-1.モデル(ネットワーク)例
1-4.実施形態に係るセンシングデバイスの構成
1-5.実施形態に係る情報処理の手順
1-5-1.学習装置に係る処理の手順
1-5-2.情報処理システムに係る処理の手順
2.その他の実施形態
2-1.その他の構成例
2-2.その他
3.本開示に係る効果
4.ハードウェア構成
[1-1.本開示の実施形態に係る学習処理の概要]
以下、図1及び図2を用いて、情報処理システム1が行う処理の概要について説明する。図1は、本開示の実施形態に係る情報処理システムの処理フローの一例を示す図である。図2は、本開示の実施形態に係る学習処理の一例を示す図である。また、図2は、本開示の実施形態に係る学習処理の一例を示す図である。本開示の実施形態に係る学習処理は、学習装置の一例であるサーバ装置100やセンシングデバイス10を含む情報処理システム1によって実現される。図2では、情報処理システム1によって実現される学習処理の概要を説明する。
ここで、上述した情報処理システム1の背景や効果等について説明する。ディープラーニングの進化により、人間を超える物体認識が実現されている。しかしながら、エッジデバイスで利用されているモデルは、開発者が学習したモデルを利用しており、現実的な世界でのデータの変化に対応していない。大量なデータを必要とするディープラーニングでは、状況の変化に応じた新たなデータを必要とする。だが、現状では、開発者が作製したモデルを利用し続けているのが現状である。
ここから、情報処理システム1における各手法について記載する。まず、Influence functionについて記載する。情報処理システム1は、Influence functionにより、データが生成するモデル(パラメータ)に与える影響を定量的に解析する。
・Understanding Black-box Predictions via Influence Functions, Pang Wei Kho and Percy Liang <https://arxiv.org/abs/1703.04730>
・Residuals and Influence in Regression, Cook, R.D. and Weisberg, S <https://conservancy.umn.edu/handle/11299/37076>
上述したInfluence functionは一例に過ぎず、影響度の算出に用いる手法は、Influence functionに限られない。この点についての例示を以下記載する。
次に、データ量に関する点について記載する。データの影響度の計算において、膨大なHDDキャッシュ容量を必要とする。仮に無限大にHDDがあれば、問題ないが、現実的には有限の容量である。
・Data Cleansing for Deep Neural Networks with Storage-efficient Approximation of Influence Functions, Kenji Suzuki, Yoshiyuki Kobayashi, and Takuya Narihira Wei Kho and Percy Liang <https://arxiv.org/abs/2103.11807>
次に、データの保管ストレージに関する点について記載する。自動アップデートを繰り返していると、大量のデータが次々と計算される。どのデータを使ったのかが分からなくなる問題が生じ、AI(Artificial Intelligence)の透明性が確保できない。
次に、画像データの補整に関する点について記載する。センシングデバイス10等のエッジデバイスでの現実的世界から集めたデータは、最初のモデルの学習時に利用したデータとの画像の明暗、コントラスト、色度などが異なることがある。これは、カメラや撮影条件などで差異が生じることによる。その場合、学習モデルは、適切なパフォーマンスが発揮できない場合がある。
図3に示す情報処理システム1について説明する。情報処理システム1は、学習データを調整する調整処理を実現する情報処理システムである。図3に示すように、情報処理システム1は、サーバ装置100と、複数のセンシングデバイス10a、10b、10c、10dとが含まれる。なお、センシングデバイス10a、10b、10c、10d等を区別しない場合、センシングデバイス10と記載する場合がある。また、図3では、4個のセンシングデバイス10a、10b、10c、10dを図示するが、情報処理システム1には、4個より多い数(例えば20個や100個以上)のセンシングデバイス10が含まれてもよい。センシングデバイス10と、サーバ装置100とは所定の通信網(ネットワークN)を介して、有線または無線により通信可能に接続される。図3は、実施形態に係る情報処理システムの構成例を示す図である。なお、図3に示した情報処理システム1には、複数台のサーバ装置100が含まれてもよい。
次に、実施形態に係る学習処理を実行する学習装置の一例であるサーバ装置100の構成について説明する。図4は、本開示の実施形態に係るサーバ装置100の構成例を示す図である。
上述したように、サーバ装置100は、ディープニューラルネットワーク(DNN)等のニューラルネットワーク(NN)の形式のモデル(ネットワーク)を用いてもよい。なお、サーバ装置100は、ニューラルネットワークに限らず、SVM(Support Vector Machine)等の回帰モデルや等の種々の形式のモデル(関数)を用いてもよい。このように、サーバ装置100は、任意の形式のモデル(関数)を用いてもよい。サーバ装置100は、非線形の回帰モデルや線形の回帰モデル等、種々の回帰モデルを用いてもよい。
次に、実施形態に係る情報処理を実行するセンシングデバイスの一例であるセンシングデバイス10の構成について説明する。図9は、本開示の実施形態に係るセンシングデバイスの構成例を示す図である。
次に、図11及び図12を用いて、実施形態に係る各種情報処理の手順について説明する。
まず、図11を用いて、本開示の実施形態に係る学習装置に係る処理の流れについて説明する。図11は、本開示の実施形態に係る学習装置の処理を示すフローチャートである。具体的には、図11は、学習装置の一例であるサーバ装置100による情報処理の手順を示すフローチャートである。
次に、図12を用いて、情報処理システムに係る具体的な処理の一例について説明する。図12は、本開示の実施形態に係る情報処理システムの処理手順を示すシーケンス図である。
上述した各実施形態に係る処理は、上記各実施形態や変形例以外にも種々の異なる形態(変形例)にて実施されてよい。
なお、上記の例では、サーバ装置100とセンシングデバイス10とが別体、すなわちモデルを学習する学習装置とデータをセンシングする装置が別体である場合を示したが、これらの装置は一体であってもよい。例えば、センシングデバイス10は、センシングによりデータを収集する機能と、モデルを学習する機能とを有する学習装置(情報処理装置)であってもよい。この場合、センシングデバイス10は、上述したサーバ装置100のモデルを学習するための各種構成(例えば算出部132、学習部136等)を有し、自装置で収集したデータを用いてモデルを生成する。センシングデバイス10は、カメラ、スマホ、テレビ、自動車、ドローン、ロボット等であってもよい。このように、センシングデバイス10は、自律的に影響度の高い学習データを収集し、モデルを生成する端末装置(コンピュータ)であってもよい。
また、上記各実施形態において説明した各処理のうち、自動的に行われるものとして説明した処理の全部または一部を手動的に行うこともでき、あるいは、手動的に行われるものとして説明した処理の全部または一部を公知の方法で自動的に行うこともできる。この他、上記文書中や図面中で示した処理手順、具体的名称、各種のデータやパラメータを含む情報については、特記する場合を除いて任意に変更することができる。例えば、各図に示した各種情報は、図示した情報に限られない。
上述のように、本開示に係る学習装置(実施形態ではサーバ装置100)は、算出部(実施形態では算出部132)と、学習部(実施形態では学習部136)とを備える。算出部は、センシングデバイス(実施形態ではセンシングデバイス10)により収集されたデータが機械学習によるモデルの学習に与える影響度を算出する。学習部は、算出部により算出された影響度が条件を満たすデータを用いてモデルを学習する少ラベル学習処理により、学習済みモデルを生成する。
上述してきた各実施形態や変形例に係るサーバ装置100やセンシングデバイス10等の情報機器は、例えば図13に示すような構成のコンピュータ1000によって実現される。図13は、サーバ装置100やセンシングデバイス10等の情報処理装置の機能を実現するコンピュータ1000の一例を示すハードウェア構成図である。以下、実施形態に係るサーバ装置100を例に挙げて説明する。コンピュータ1000は、CPU1100、RAM1200、ROM(Read Only Memory)1300、HDD(Hard Disk Drive)1400、通信インターフェイス1500、及び入出力インターフェイス1600を有する。コンピュータ1000の各部は、バス1050によって接続される。
(1)
センシングデバイスにより収集されたデータが機械学習によるモデルの学習に与える影響度を算出する算出部と、
前記算出部により算出された前記影響度が条件を満たすデータを用いて前記モデルを学習する少ラベル学習処理により、学習済みモデルを生成する学習部と、
を備える学習装置。
(2)
前記学習部は、
前記影響度が所定の閾値よりも大きいデータを用いて前記少ラベル学習処理を行う
(1)に記載の学習装置。
(3)
前記算出部は、
損失関数に基づいて前記影響度を算出する
(1)または(2)に記載の学習装置。
(4)
前記算出部は、
Influence functionsにより前記影響度を算出する
(1)~(3)のいずれか1つに記載の学習装置。
(5)
ラベルが付されていないデータであるラベル無しデータのラベルを予測する予測部、
をさらに備え、
前記学習部は、
前記影響度が条件を満たすラベル無しデータを対象データとして前記予測部が予測した予測ラベルと前記対象データとを用いて前記少ラベル学習処理を行う
(1)~(4)のいずれか1つに記載の学習装置。
(6)
前記予測部は、
ラベルが付されたラベル有りデータのデータセットにより学習された分類器を用いて、前記対象データの前記予測ラベルを予測し、
前記学習部は、
前記予測ラベルを付した前記対象データが追加されたデータセットを用いて前記学習済みモデルを生成する
(5)に記載の学習装置。
(7)
前記影響度が条件を満たさないデータを削除し、前記影響度が条件を満たすデータをログとして記憶部に格納するデータ管理部、
をさらに備える(1)~(6)のいずれか1つに記載の学習装置。
(8)
前記算出部は、
イメージセンサにより収集した画像データの前記影響度を算出し、
前記学習部は、
前記影響度が条件を満たす画像データを用いて前記少ラベル学習処理を行う
(1)~(7)のいずれか1つに記載の学習装置。
(9)
前記学習部は、
前記影響度が条件を満たす画像データが補整された補整後の画像データを用いて前記少ラベル学習処理を行う
(8)に記載の学習装置。
(10)
前記学習部により生成された前記学習済みモデルを外部装置へ送信する送信部、
をさらに備える(1)~(9)のいずれか1つに記載の学習装置。
(11)
前記算出部は、
前記外部装置であるセンシングデバイスが前記学習済みモデルを用いて収集したデータの前記影響度を算出し、
前記学習部は、
前記算出部により算出された前記影響度が条件を満たすデータを用いて前記学習済みモデルを更新する
(10)に記載の学習装置。
(12)
前記送信部は、
前記学習部により更新された前記学習済みモデルを前記センシングデバイスへ送信する
(11)に記載の学習装置。
(13)
前記センシングデバイスにモデルを提供するサーバ装置である
(11)または(12)に記載の学習装置。
(14)
センシングデバイスにより収集されたデータが機械学習によるモデルの学習に与える影響度を算出し、
算出した前記影響度が条件を満たすデータを用いて前記モデルを学習する少ラベル学習処理により、学習済みモデルを生成する、
処理を実行する学習方法。
(15)
センシングにより収集したデータを、当該データが機械学習によるモデルの学習に与える影響度が条件を満たす場合に当該データを用いて前記モデルを学習する少ラベル学習処理により学習済みモデルを生成する学習装置に送信する送信部と、
前記学習装置が学習した前記学習済みモデルを前記学習装置から受信する受信部と、
前記学習済みモデルを用いたセンシングによりデータを収集する収集部と、
を備えるセンシングデバイス。
(16)
前記送信部は、
前記収集部が前記学習済みモデルを用いたセンシングにより収集したデータを前記学習装置へ送信する
(15)に記載のセンシングデバイス。
(17)
前記受信部は、
前記センシングデバイスが前記学習済みモデルを用いたセンシングにより収集したデータを用いて更新された前記学習済みモデルを前記学習装置から受信し、
前記収集部は、
前記学習装置により更新された前記学習済みモデルを用いたセンシングによりデータを収集する
(16)に記載のセンシングデバイス。
(18)
前記収集部は、
センサ部により検知された画像データを収集する
(15)~(17)のいずれか1つに記載のセンシングデバイス。
(19)
前記送信部は
センシングにより収集した画像データを、前記学習装置に送信し、
前記受信部は、
前記学習装置が画像データを用いて学習した前記学習済みモデルを前記学習装置から受信し、
前記収集部は、
前記学習済みモデルを用いたセンシングにより画像データを収集する
(18)に記載のセンシングデバイス。
(20)
センシングにより収集したデータを、当該データが機械学習によるモデルの学習に与える影響度が条件を満たす場合に当該データを用いて前記モデルを学習する少ラベル学習処理により学習済みモデルを生成する学習装置に送信し、
前記学習装置が学習した前記学習済みモデルを前記学習装置から受信し、
前記学習済みモデルを用いたセンシングによりデータを収集する、
処理を実行するデータ収集方法。
100 サーバ装置(学習装置)
110 通信部
120 記憶部
121 データ情報記憶部
122 モデル情報記憶部
123 閾値情報記憶部
130 制御部
131 取得部
132 算出部
133 データ管理部
134 補整部
136 学習部
137 送信部
10 センシングデバイス
11 通信部
12 入力部
13 出力部
14 記憶部
141 モデル情報記憶部
15 制御部
151 受信部
152 収集部
153 送信部
16 センサ部
Claims (20)
- センシングデバイスにより収集されたデータが機械学習によるモデルの学習に与える影響度を算出する算出部と、
前記算出部により算出された前記影響度が条件を満たすデータを用いて前記モデルを学習する少ラベル学習処理により、学習済みモデルを生成する学習部と、
を備える学習装置。 - 前記学習部は、
前記影響度が所定の閾値よりも大きいデータを用いて前記少ラベル学習処理を行う
請求項1に記載の学習装置。 - 前記算出部は、
損失関数に基づいて前記影響度を算出する
請求項1に記載の学習装置。 - 前記算出部は、
Influence functionsにより前記影響度を算出する
請求項1に記載の学習装置。 - ラベルが付されていないデータであるラベル無しデータのラベルを予測する予測部、
をさらに備え、
前記学習部は、
前記影響度が条件を満たすラベル無しデータを対象データとして前記予測部が予測した予測ラベルと前記対象データとを用いて前記少ラベル学習処理を行う
請求項1に記載の学習装置。 - 前記予測部は、
ラベルが付されたラベル有りデータのデータセットにより学習された分類器を用いて、前記対象データの前記予測ラベルを予測し、
前記学習部は、
前記予測ラベルを付した前記対象データが追加されたデータセットを用いて前記学習済みモデルを生成する
請求項5に記載の学習装置。 - 前記影響度が条件を満たさないデータを削除し、前記影響度が条件を満たすデータをログとして記憶部に格納するデータ管理部、
をさらに備える請求項1に記載の学習装置。 - 前記算出部は、
イメージセンサにより収集した画像データの前記影響度を算出し、
前記学習部は、
前記影響度が条件を満たす画像データを用いて前記少ラベル学習処理を行う
請求項1に記載の学習装置。 - 前記学習部は、
前記影響度が条件を満たす画像データが補整された補整後の画像データを用いて前記少ラベル学習処理を行う
請求項8に記載の学習装置。 - 前記学習部により生成された前記学習済みモデルを外部装置へ送信する送信部、
をさらに備える請求項1に記載の学習装置。 - 前記算出部は、
前記外部装置であるセンシングデバイスが前記学習済みモデルを用いて収集したデータの前記影響度を算出し、
前記学習部は、
前記算出部により算出された前記影響度が条件を満たすデータを用いて前記学習済みモデルを更新する
請求項10に記載の学習装置。 - 前記送信部は、
前記学習部により更新された前記学習済みモデルを前記センシングデバイスへ送信する
請求項11に記載の学習装置。 - 前記センシングデバイスにモデルを提供するサーバ装置である
請求項11に記載の学習装置。 - センシングデバイスにより収集されたデータが機械学習によるモデルの学習に与える影響度を算出し、
算出した前記影響度が条件を満たすデータを用いて前記モデルを学習する少ラベル学習処理により、学習済みモデルを生成する、
処理を実行する学習方法。 - センシングにより収集したデータを、当該データが機械学習によるモデルの学習に与える影響度が条件を満たす場合に当該データを用いて前記モデルを学習する少ラベル学習処理により学習済みモデルを生成する学習装置に送信する送信部と、
前記学習装置が学習した前記学習済みモデルを前記学習装置から受信する受信部と、
前記学習済みモデルを用いたセンシングによりデータを収集する収集部と、
を備えるセンシングデバイス。 - 前記送信部は、
前記収集部が前記学習済みモデルを用いたセンシングにより収集したデータを前記学習装置へ送信する
請求項15に記載のセンシングデバイス。 - 前記受信部は、
前記センシングデバイスが前記学習済みモデルを用いたセンシングにより収集したデータを用いて更新された前記学習済みモデルを前記学習装置から受信し、
前記収集部は、
前記学習装置により更新された前記学習済みモデルを用いたセンシングによりデータを収集する
請求項16に記載のセンシングデバイス。 - 前記収集部は、
センサ部により検知された画像データを収集する
請求項15に記載のセンシングデバイス。 - 前記送信部は
センシングにより収集した画像データを、前記学習装置に送信し、
前記受信部は、
前記学習装置が画像データを用いて学習した前記学習済みモデルを前記学習装置から受信し、
前記収集部は、
前記学習済みモデルを用いたセンシングにより画像データを収集する
請求項18に記載のセンシングデバイス。 - センシングにより収集したデータを、当該データが機械学習によるモデルの学習に与える影響度が条件を満たす場合に当該データを用いて前記モデルを学習する少ラベル学習処理により学習済みモデルを生成する学習装置に送信し、
前記学習装置が学習した前記学習済みモデルを前記学習装置から受信し、
前記学習済みモデルを用いたセンシングによりデータを収集する、
処理を実行するデータ収集方法。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP22878304.9A EP4414900A1 (en) | 2021-10-06 | 2022-09-16 | Learning device, learning method, sensing device, and data collection method |
CN202280066530.9A CN118043828A (zh) | 2021-10-06 | 2022-09-16 | 学习装置、学习方法、感测装置及数据收集方法 |
JP2023552781A JPWO2023058433A1 (ja) | 2021-10-06 | 2022-09-16 |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2021-164984 | 2021-10-06 | ||
JP2021164984 | 2021-10-06 |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2023058433A1 true WO2023058433A1 (ja) | 2023-04-13 |
Family
ID=85804202
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2022/034670 WO2023058433A1 (ja) | 2021-10-06 | 2022-09-16 | 学習装置、学習方法、センシングデバイス及びデータ収集方法 |
Country Status (4)
Country | Link |
---|---|
EP (1) | EP4414900A1 (ja) |
JP (1) | JPWO2023058433A1 (ja) |
CN (1) | CN118043828A (ja) |
WO (1) | WO2023058433A1 (ja) |
Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018513490A (ja) * | 2015-04-16 | 2018-05-24 | シーメンス アクチエンゲゼルシヤフトSiemens Aktiengesellschaft | 自動化システムを動作させる方法及び装置 |
JP2020009141A (ja) * | 2018-07-06 | 2020-01-16 | 株式会社 日立産業制御ソリューションズ | 機械学習装置及び方法 |
JP2020177344A (ja) * | 2019-04-16 | 2020-10-29 | 富士通株式会社 | 学習方法、学習プログラムおよび学習装置 |
JP2020204800A (ja) * | 2019-06-14 | 2020-12-24 | Awl株式会社 | 学習用データセット生成システム、学習サーバ、及び学習用データセット生成プログラム |
JP2021026505A (ja) | 2019-08-05 | 2021-02-22 | ギリア株式会社 | 情報処理装置および情報処理用プログラム |
JP2021196921A (ja) * | 2020-06-16 | 2021-12-27 | 株式会社日立製作所 | モデル運用支援システム及び方法 |
-
2022
- 2022-09-16 WO PCT/JP2022/034670 patent/WO2023058433A1/ja active Application Filing
- 2022-09-16 CN CN202280066530.9A patent/CN118043828A/zh active Pending
- 2022-09-16 JP JP2023552781A patent/JPWO2023058433A1/ja active Pending
- 2022-09-16 EP EP22878304.9A patent/EP4414900A1/en active Pending
Patent Citations (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2018513490A (ja) * | 2015-04-16 | 2018-05-24 | シーメンス アクチエンゲゼルシヤフトSiemens Aktiengesellschaft | 自動化システムを動作させる方法及び装置 |
JP2020009141A (ja) * | 2018-07-06 | 2020-01-16 | 株式会社 日立産業制御ソリューションズ | 機械学習装置及び方法 |
JP2020177344A (ja) * | 2019-04-16 | 2020-10-29 | 富士通株式会社 | 学習方法、学習プログラムおよび学習装置 |
JP2020204800A (ja) * | 2019-06-14 | 2020-12-24 | Awl株式会社 | 学習用データセット生成システム、学習サーバ、及び学習用データセット生成プログラム |
JP2021026505A (ja) | 2019-08-05 | 2021-02-22 | ギリア株式会社 | 情報処理装置および情報処理用プログラム |
JP2021196921A (ja) * | 2020-06-16 | 2021-12-27 | 株式会社日立製作所 | モデル運用支援システム及び方法 |
Non-Patent Citations (1)
Title |
---|
SOKI TAGUCHI: "AS-1-3 Proposal of XAI Utilization in DNN Lightweight Implementation", PROCEEDINGS OF THE IEICE ENGINEERING SCIENCES SOCIETY/NOLTA SOCIETY CONFERENCE, IEICE, JP, vol. 2021, 31 August 2021 (2021-08-31) - 17 September 2021 (2021-09-17), JP , pages S5 - S6, XP009546267, ISSN: 2189-700X * |
Also Published As
Publication number | Publication date |
---|---|
JPWO2023058433A1 (ja) | 2023-04-13 |
CN118043828A (zh) | 2024-05-14 |
EP4414900A1 (en) | 2024-08-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
Hossain et al. | Active learning enabled activity recognition | |
CN108182394B (zh) | 卷积神经网络的训练方法、人脸识别方法及装置 | |
KR102557956B1 (ko) | 딥 러닝 모델들의 분산 훈련을 위한 시스템들 및 방법들 | |
EP3555815B1 (en) | Unsupervised learning techniques for temporal difference models | |
CN110826453B (zh) | 一种通过提取人体关节点坐标的行为识别方法 | |
US9111375B2 (en) | Evaluation of three-dimensional scenes using two-dimensional representations | |
US10671895B2 (en) | Automated selection of subjectively best image frames from burst captured image sequences | |
JP6972389B2 (ja) | システム状態を解析及び修正するための時系列の検索 | |
US20230115987A1 (en) | Data adjustment system, data adjustment device, data adjustment method, terminal device, and information processing apparatus | |
CN113284142A (zh) | 图像检测方法、装置、计算机可读存储介质及计算机设备 | |
WO2019097784A1 (ja) | 情報処理装置、情報処理方法、およびプログラム | |
WO2019018533A1 (en) | NEURO-BAYESIAN ARCHITECTURE FOR THE IMPLEMENTATION OF GENERAL ARTIFICIAL INTELLIGENCE | |
US20240070449A1 (en) | Systems and methods for expert guided semi-supervision with contrastive loss for machine learning models | |
JP2018528511A (ja) | 生産システムにおける出力効率の最適化 | |
US11989939B2 (en) | System and method for enhancing machine learning model for audio/video understanding using gated multi-level attention and temporal adversarial training | |
Deng et al. | Few-shot human activity recognition on noisy wearable sensor data | |
JP6981428B2 (ja) | 情報処理装置および情報処理方法 | |
Nikpour et al. | Deep Reinforcement Learning in Human Activity Recognition: A Survey and Outlook | |
Nikpour et al. | Deep reinforcement learning in human activity recognition: A survey | |
WO2023058433A1 (ja) | 学習装置、学習方法、センシングデバイス及びデータ収集方法 | |
Jiang et al. | CMCI: A Robust Multimodal Fusion Method for Spiking Neural Networks | |
CN116958041A (zh) | 一种产品缺陷检测方法、装置、电子设备和存储介质 | |
KR101893290B1 (ko) | 딥 러닝 기반 교육용 비디오 학습 및 평가 시스템 | |
JP2024507765A (ja) | エッジデバイスに展開されたモデルによる機械学習への自己教師あり共同アプローチ | |
Manolova et al. | Human activity recognition with semantically guided graph-convolutional 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: 22878304 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2023552781 Country of ref document: JP |
|
WWE | Wipo information: entry into national phase |
Ref document number: 18692848 Country of ref document: US |
|
WWE | Wipo information: entry into national phase |
Ref document number: 202280066530.9 Country of ref document: CN |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2022878304 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: 2022878304 Country of ref document: EP Effective date: 20240506 |