Information processing device, information processing method, computer program, imaging device, vehicle device, and medical robot device

Abstract

Provided is an information processing device that generates training data for making unbiased decisions on each item of input data by artificially increasing the data of minor attributes.　The information processing device comprises: a data retaining unit for retaining first training data that is used in training a machine learning model; an acquisition unit for acquiring information related to the bias of the training data; a data generation unit for generating second training data using the data included in the training data, on the basis of the information related to the bias; and a training unit for training the machine learning model using the first training data and the second training data.

Description

[Technical Field]

[0001] The technology disclosed in this specification (hereinafter referred to as "the present disclosure") relates to an information processing device and information processing method for processing learning data, a computer program, an imaging device, a vehicle device, and a medical robot device. [Background technology]

[0002] Artificial intelligence can analyze and make inferences from huge amounts of data, and is used, for example, in image recognition, voice recognition, and natural language processing. Artificial intelligence is realized by training a machine learning model composed of a neural network or the like. For example, an imaging device has been proposed that includes a recognition unit that performs recognition processing using a trained model and outputs recognition results for pixel signals (see Patent Document 1).

[0003] Deep learning using massive training datasets can produce artificial intelligence capable of inference (such as facial and object recognition) that exceeds human capabilities. However, bias in the training data raises concerns that the AI ​​may be biased and make unfair judgments. For example, when training a machine learning model to classify images of people, if the model is trained using data that is biased in attributes such as age, gender, race, and ethnicity, there is a risk that the model will not be able to make fair judgments, such as outputting unfavorable results for minority attributes. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent No. 6635221 [Patent Document 2] Japanese Patent Application Publication No. 2018-72893 [Non-patent literature]

[0005] [Non-Patent Document 1] Goodfellow et al."Explaining and Harnessing Adversarial Examples."(2015) ICLR. Summary of the Invention [Problem to be solved by the invention]

[0006] An object of the present disclosure is to provide an information processing device and information processing method, a computer program, an imaging device, a vehicle device, and a medical robot device that processes data to learn a model so that fair judgments can be made for each input data. [Means for solving the problem]

[0007] The present disclosure has been made in consideration of the above problems, and a first aspect thereof is: a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about the bias of the learning data; a data generation unit that generates second training data using data included in the training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; The information processing device is provided with:

[0008] The acquisition unit acquires information indicating an attribute that is a minority of the first training data, and the data generation unit generates second training data having the same attribute from data of the minority attribute included in the first training data.

[0009] The data generation unit generates Adversarial Examples to serve as second training data from the data of the minority attributes included in the training data. For example, the data generation unit generates Adversarial Examples based on a Fast Gradient Sign Method.

[0010] Furthermore, a second aspect of the present disclosure is inputting first learning data to be used for learning the machine learning model; acquiring information about bias in the learning data; generating second training data using data included in the training data based on information about the bias; training the machine learning model using the first training data and the second training data; The information processing method has the following features.

[0011] Furthermore, a third aspect of the present disclosure is a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the learning data; a data generation unit that generates second training data using data included in the training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; It is a computer program written in a computer-readable format to cause a computer to function as a

[0012] A computer program according to a third aspect of the present disclosure defines a computer program written in a computer-readable format to execute predetermined processing on a computer. In other words, by installing the computer program according to the claims of the present application on a computer, a cooperative action is exerted on the computer, and the same effects as those of the information processing device according to the first aspect of the present disclosure can be obtained.

[0013] Furthermore, a fourth aspect of the present disclosure is an imaging unit that captures an image; a recognition unit that recognizes the captured image using a machine learning model; Equipped with generating training data using data included in the training data based on information about bias in the training data for training the machine learning model, and training the machine learning model using the generated training data; It is an imaging device.

[0014] Furthermore, a fifth aspect of the present disclosure is an imaging unit that captures an image of the surroundings of the vehicle; a recognition unit that recognizes the captured image using a machine learning model; Equipped with an imaging device that generates training data using data included in the training data based on information about bias in the training data for training the machine learning model, and trains the machine learning model using the generated training data; The vehicle device is equipped with the above.

[0015] A sixth aspect of the present disclosure is: an imaging unit that captures an image of the area around the surgical site; a recognition unit that recognizes the captured image using a machine learning model; Equipped with an imaging device that generates training data using data included in the training data based on information about bias in the training data for training the machine learning model, and trains the machine learning model using the generated training data It is a medical robot device equipped with the above. [Effects of the Invention]

[0016] According to the present disclosure, it is possible to provide an information processing device and information processing method, a computer program, an imaging device, a vehicle device, and a medical robot device that generate learning data for making fair judgments on each input data by artificially increasing data with minority attributes.

[0017] It should be noted that the effects described in this specification are merely examples, and the effects brought about by the present disclosure are not limited to these. Furthermore, the present disclosure may also bring about additional effects in addition to the effects described above.

[0018] Further objects, features, and advantages of the present disclosure will become apparent from the following detailed description based on the embodiments and accompanying drawings. [Brief explanation of the drawings]

[0019] [Figure 1] FIG. 1 is a diagram showing an example of the functional configuration of a learning system 100. As shown in FIG. [Figure 2] Figure 2 shows how Adversarial Examples are generated. [Figure 3] FIG. 3 is a diagram showing a mechanism for adding learning data. [Figure 4] Figure 4 shows how inequality between groups arises based on differences in sensitive attributes. [Figure 5] FIG. 5 is a diagram showing how unfairness occurs between individuals. [Figure 6] FIG. 6 is a flowchart showing an example of the operation of learning system 100 in the learning phase. [Figure 7] FIG. 7 is a flowchart showing another example of the operation of learning system 100 in the learning phase. [Figure 8] FIG. 8 is a diagram showing an example of the functional configuration of an imaging device 800. As shown in FIG. [Figure 9] FIG. 9 is a diagram showing an example of hardware implementation of the imaging device 800. As shown in FIG. [Figure 10] FIG. 10 is a diagram showing another example of hardware implementation of the imaging device 800. In FIG. [Figure 11] FIG. 11 is a diagram showing an example in which the semiconductor chip of the imaging device 800 is formed as a stacked image sensor 1100 having a two-layer structure. [Figure 12]FIG. 12 is a diagram showing an example in which the semiconductor chip of the imaging device 800 is formed as a stacked image sensor 1200 having a three-layer structure. [Figure 13] FIG. 13 is a diagram showing an example of the configuration of the sensor unit 802. As shown in FIG. [Figure 14] FIG. 14 is a flowchart showing an example of the operation of the image capturing device 800 with a recognition function in the learning phase. [Figure 15] FIG. 15 is a diagram showing an example of the functional configuration of an in-vehicle camera 1500. As shown in FIG. [Figure 16] FIG. 16 is a diagram showing an example of the internal configuration of the image sensor 1502 and the signal processing unit 1503. [Figure 17] FIG. 17 is a flowchart showing an example of the operation of the recognition unit 1504 of the vehicle-mounted camera 1500 in the learning phase. [Figure 18] FIG. 18 is a diagram showing an example of the configuration of a medical robot device 1800. [Figure 19] FIG. 19 is a flowchart showing an example of the operation of the image recognizer 1821 in the learning phase. DETAILED DESCRIPTION OF THE INVENTION

[0020] The present disclosure will be described below in the following order with reference to the drawings.

[0021] A. Overview B. Learning System C. Adding training data D. Fairness E. Example of operation F. Application to imaging devices G. Application to in-vehicle cameras H. Applications to Healthcare Systems

[0022] A. Overview Artificial intelligence consists of models using, for example, neural networks, support vector regression, Gaussian process regression, and other types of models. For convenience, this specification focuses on neural network-type models, but the present disclosure is not limited to a specific model type and is equally applicable to models other than neural networks. The use of artificial intelligence consists of a "learning phase" in which a model is trained, and an "inference phase" in which inference is made using the trained model. Inference includes recognition processes such as image recognition and speech recognition, and prediction processes that estimate or predict events. The present disclosure is particularly applicable to models that determine classification problems such as image classification.

[0023] In the learning phase of artificial intelligence, a model is trained using a learning algorithm such as backpropagation, using a dataset consisting of a combination of data input to the model (hereinafter also referred to as "input data") and labels that the model is to estimate for the input data, so that the correct label corresponding to each input data can be output. Then, in the inference phase of artificial intelligence, the model that has been trained in the learning phase (hereinafter also referred to as "trained model") outputs an appropriate label for the input data.

[0024] Deep learning using massive training datasets can produce artificial intelligence that can make inferences that exceed human capabilities. However, if the training data is biased across categories, there is a risk that the learning model will be biased and unable to make fair judgments in some categories.

[0025] For example, unfairness in judgments between categories can be corrected by increasing the amount of training data for categories with few attributes, which makes fair judgments impossible, and training the model accordingly, or by reducing the amount of training data for categories with many attributes, which allows fair judgments. However, manually adjusting the large amount of training data required for deep learning is an extremely difficult task. Furthermore, when the training data involves personal information, such as facial images, obtaining the data requires the individual's consent, which is a significant hurdle. In short, increasing the amount of training data is not realistic. On the other hand, reducing the training data for other categories results in a problem of reduced model accuracy for those categories.

[0026] Therefore, the present disclosure artificially generates training data with a minority attribute from original training data. By artificially increasing the training data with a minority attribute, it is possible to eliminate imbalances in the training data for each attribute and train a machine learning model to make unbiased and fair judgments. Furthermore, because additional training data is increased by computational processing from already acquired original training data, this is an easier task than manual adjustment, and hurdles such as obtaining individual consent are also lowered.

[0027] For example, additional training data can be artificially generated from some or all of the training data with minority attributes included in the original training data using adversarial examples (see, for example, Non-Patent Document 1). Adversarial examples are images that are imperceptible to humans but affect machine learning. For example, adversarial examples can be generated easily (i.e., with a small computational load) using a method based on fast gradient signs (FGSM).

[0028] B. Learning System FIG. 1 shows an example of the functional configuration of a learning system 100 to which the present disclosure is applied. The illustrated learning system 100 is mounted on, for example, an edge device for use, but some or all of the functions of the learning system 100 may be implemented on the cloud or a computing device capable of large-scale computing. The learning system 100 includes a learning data storage unit 101, a learning unit 102, a model parameter storage unit 103, an analysis unit 104, a data generation unit 105, an inference unit 111, a data input unit 112, and an input data processing unit 113. Note that the above-mentioned functional modules 101 to 105 may all be located within a single device, or may be distributed across two or more physically independent devices.

[0029] The learning data holding unit 101 accumulates a data set that is used for model learning by the learning unit 102. When the learning unit 102 performs deep learning, a huge amount of data set is accumulated in the learning data holding unit 101.

[0030] A dataset generally consists of a combination (x, y) of data x to be input to a model to be trained and a correct label y that is the correct answer for the data x. The explanatory variables of the data x include sensitive attributes, which may cause issues such as unfairness in the inference results of a trained model, and other non-sensitive attributes. For example, race, gender, age, and other explanatory variables are sensitive attributes. In this embodiment, all explanatory variables other than sensitive attributes are further treated as a dataset by adding an attribute s of the input data x. For example, in the case of a dataset for image classification, the input data x is a face image of a person, and the attribute s is the person's age, gender, race, ethnicity, etc.

[0031] The learning unit 102 sequentially reads data sets from the learning data storage unit 101 and performs model learning. The model is configured, for example, by a neural network, but may also be a model using support vector regression, Gaussian process regression, or other types of model. The learning unit 102 then stores model parameters obtained as a result of learning in the model parameter storage unit 103. The model parameters are variable elements that define the model, such as coefficients or weighting coefficients assigned to each neuron in the neural network model.

[0032] The inference unit 111, the data input unit 112, and the input data processing unit 113 perform the inference phase of the trained model. The data input unit 112 inputs sensor information acquired by a sensor provided in the edge device. The input data processing unit 113 processes the data input from the data input unit 112 into a data format that can be input to a model (e.g., a neural network model), and inputs the data to the inference unit 111. The inference unit 111 outputs a label inferred from the input data using a model in which model parameters read from the model parameter holding unit 103 are set, i.e., a trained model.

[0033] The analysis unit 104 analyzes the bias of data in the dataset used by the learning unit 102 for model learning, and acquires information about the bias of the learning data. Data bias means that the dataset used for learning is concentrated on some attributes, resulting in a dataset with a few attributes and a dataset with a majority of attributes. The analysis unit 104 may acquire information about the bias of data in the dataset by means other than analysis.

[0034] The method by which the analysis unit 104 analyzes the bias in the training dataset is not particularly limited. The analysis unit 104 may analyze the explanatory variables of the dataset stored in the training data storage unit 101, may analyze the explanatory variables of the dataset read by the learning unit 102 from the training data storage unit 101, may analyze the model learned by the learning unit 102, or may analyze the fairness of the results of inference by the inference unit 111 using the trained model. Furthermore, the analysis unit 104 may analyze the bias in the training dataset based on techniques such as XAI (eXplainable AI), confidence score calculation of training data, influence function calculation, Bayesian DNN (Deep Neural Network), etc.

[0035] The data generation unit 105 generates training data that is found to have minority attributes from the analysis results by the analysis unit 104, and adds and stores the data in the training data storage unit 101. The analysis unit 104 reads out a dataset having the corresponding attribute from the training data storage unit 101, and artificially generates a training dataset with the minority attribute from the original dataset that has been read out.

[0036] Therefore, by artificially increasing the amount of training data for minority attributes, the learning system 100 can eliminate imbalances in the training data for each attribute and train a machine learning model to make unbiased and fair judgments. Furthermore, because additional training data is generated from the original training data already acquired through computational processing, this is an easier task than manual adjustment, and hurdles such as obtaining consent from the individual are also lowered.

[0037] C. Adding training data A model trained using a dataset (x, y) can be considered as a function f that calculates the objective variable y from the explanatory variable x (y = f(x)). For example, the explanatory variable is a face image, and the objective variable is person detection or face identification. The explanatory variables x(s) are broadly divided into explanatory variables x(s = 0) for sensitive attributes (s = 0) related to the unfairness of the objective variable, such as race, gender, and age, and all explanatory variables x(s = 1) other than sensitive attributes (s = 1). A method is needed to train a model that is not influenced by explanatory variables x(s = 0) related to sensitive attributes and can make fair judgments without bias.

[0038] Therefore, in the present disclosure, additional training data is artificially generated from training data with a minority attribute included in the original training data using an adversarial example, and training data is added. According to the present disclosure, since additional training data is artificially generated from the original training data, there is no need to supplement actual data, and this can be said to be a realistic method.

[0039] Adversarial examples are images that are imperceptible to humans but affect machine learning. Figure 2 shows how an adversarial example 203 is generated by superimposing minute noise 202 on an original image 201. For example, an adversarial example can be easily generated using a fast gradient sign method (FGSM). With FGSM, as shown on the right side of equation (1) below, an adversarial example on the left side can be generated by using an update formula that adds (or subtracts) an appropriate value to (x) the original image in the direction that increases the loss.

[0040]

number

[0041] In the above equation (1), x is the input data (image vector), y is the correct label, ε is an appropriately small value, J is the loss function, and θ is the model parameter. According to the above equation (1), by applying a perturbation represented by the second term on the right-hand side to the original data x in the first term on the right-hand side, an Adversarial Example that increases the loss with the correct label is generated.

[0042] The second term on the left side of the above equation (1) is ε(∇ x J(θ,x,y) is noise that humans cannot distinguish. If the noise in the second term on the left side is added to the original data x in the first term on the left side, the classifier for x will mistakenly classify it as y' instead of y, for example.

[0043] Considering the attribute s of data x, the Adversarial Example generation formula shown in formula (1) above can be expressed as formula (2) below. However, s in formula (2) below takes on the value s=0 or s=1 depending on the attribute, as shown in formula (3) below. The attribute value s=0 for sensitive data attributes (minority attributes), and the attribute value s=1 for other non-sensitive data attributes (majority attributes).

[0044]

number

[0045]

number

[0046] Using the above equation (2), two Adversarial Examples, x(s=0) and x(s=1), can be generated from the original data x(s), as shown in the following equations (4) and (5).

[0047]

number

[0048]

number

[0049] FIG. 3 illustrates a mechanism for adding training data in the present disclosure. Of the two Adversarial Examples that can be generated from the original data x(s), x(s=0) is added to the training data to eliminate data bias, while x(s=1) is treated as noise and is not used. Then, a model is trained using data in which an Adversarial Example has been added to the original dataset. Therefore, according to the present disclosure, by adding a dataset with a minority attribute and performing training, it is possible to train an unbiased model that is less susceptible to explanatory variables related to sensitive attributes.

[0050] D. Fairness Regarding fairness, there are issues of group fairness and individual fairness.

[0051] The former issue of fairness between groups is caused by differences in sensitive attributes between different groups. Figure 4 shows how unfairness between groups occurs based on differences in sensitive attributes. To solve this issue, group-independent control is required.

[0052] On the other hand, the latter, inter-individual fairness, refers to cases where unfairness between individuals occurs. Figure 5 shows how unfairness between individuals occurs. There are cases where unfair differences arise between individuals with the same ability. The solution to this is to make adjustments so that there are no differences in results between individuals.

[0053] The present disclosure provides a technique for mitigating data bias in inequality between groups.

[0054] E. Example of operation FIG. 6 shows, in the form of a flowchart, an example of the operation of the learning system 100 shown in FIG. 1 in the learning phase.

[0055] The original data set stored in the training data holding unit 101 is input to the data generation unit 105. The data generation unit 105 generates additional data using an Adversarial Example in accordance with the above formula (2) or (4) (step S601). The additional data set made up of the generated data is stored in the training data holding unit 101.

[0056] Then, the learning unit 102 performs model learning using the original data set and the additional data set stored in the learning data storage unit 101 (step S602). The learning unit 102 stores model parameters obtained as a result of learning in the model parameter storage unit 103.

[0057] In the inference phase, the inference unit 111 outputs a label inferred from the data input to the data input unit 112 using a model in which model parameters read from the model parameter holding unit 103 are set, i.e., a trained model.

[0058] The datasets used by the learning unit 102 for model training are acquired by external purchase or web crawling. Regardless of the acquisition route, the collected datasets contain data with minority attributes, making it difficult to train the model while ensuring fairness.

[0059] Increasing the amount of data with minority attributes requires collecting the data again, which is practically difficult. In contrast, in the processing procedure shown in Fig. 6, the data generation unit 105 generates adversarial examples related to the data with minority attributes, adds them to the dataset, and performs learning, thereby reducing the bias in the original dataset.

[0060] For example, when the learning system 100 trains a person detection or face identification model, by adding an adversarial example generated from the original data to the dataset, it becomes possible to train an improved person detection or face identification model that does not make unfair judgments against minority attributes.

[0061] FIG. 7 shows, in the form of a flowchart, another example of the operation of the learning system 100 shown in FIG. 1 in the learning phase.

[0062] The learning system 100 waits until an event occurs that requires the generation of additional data (No in step S701). During the wait, the inference unit 111 may perform inference on the input data using the learned model parameters and output a label.

[0063] Here, the event for which additional data should be generated is not particularly limited. For example, the event may be when the analysis unit 104 analyzes a data set stored in the learning data holding unit 101 and detects a minority attribute. Alternatively, the event may be when the analysis unit 104 analyzes a trained model and outputs an unfair inference result for input data with a minority attribute. Alternatively, the event may be when a user who has seen the output label of the inference unit 111 points out unfairness.

[0064] When an event occurs that requires the generation of additional data (Yes in step S701), the analysis unit 104 analyzes the attributes of the data to be added (step S702).

[0065] The data generation unit 105 reads a data set having attributes to be added from the training data storage unit 101 based on the analysis results by the analysis unit 104 (step S703), and generates additional data using an Adversarial Example in accordance with the above formula (2) or (4) (step S704). The additional data set made up of the generated data is stored in the training data storage unit 101.

[0066] Then, the learning unit 102 performs model learning using the original data set and the additional data set stored in the learning data storage unit 101 (step S705). The learning unit 102 stores model parameters obtained as a result of learning in the model parameter storage unit 103.

[0067] 7 also allows the data generation unit 105 to generate adversarial examples related to data with minority attributes and add them to the dataset, thereby reducing bias in the original dataset. For example, when the learning system 100 trains a person detection or face identification model, adding adversarial examples generated from the original data to the dataset makes it possible to train an improved person detection or face identification model that does not make unfair decisions regarding minority attributes.

[0068] F. Application to imaging devices The present disclosure can be applied to various devices that use machine learning models. FIG. 8 shows an example of the functional configuration of an imaging device 800 to which the present disclosure can be applied. The illustrated imaging device 800 includes an optical unit 801, a sensor unit 802, a sensor control unit 803, a recognition processing unit 804, a memory 805, a visual recognition processing unit 806, an output control unit 807, and a display unit 808. For example, a CMOS (Complementary Metal Oxide Semiconductor) can be used to integrate the sensor unit 802, the sensor control unit 803, the recognition processing unit 804, and the memory 805 into a CMOS image sensor. However, the imaging device 800 may be an infrared light sensor that captures images using infrared light, or another type of light sensor.

[0069] The optical unit 801 includes, for example, a plurality of optical lenses for focusing light from the subject onto the light-receiving surface of the sensor unit 802, an aperture mechanism for adjusting the size of an aperture for incident light, and a focus mechanism for adjusting the focus of light irradiated onto the light-receiving surface. The optical unit 801 may further include a shutter mechanism for adjusting the time for which light is irradiated onto the light-receiving surface. The aperture mechanism, focus mechanism, and shutter mechanism included in the optical unit are configured to be controlled by, for example, a sensor control unit 803. The optical unit 801 may be configured integrally with the imaging device 800 or may be configured separately from the imaging device 800.

[0070] The sensor unit 802 includes a pixel array in which a plurality of pixels are arranged in a matrix. Each pixel includes a photoelectric conversion element, and the pixels arranged in a matrix form a light-receiving surface. The optical unit 801 forms an image of incident light on the light-receiving surface, and each pixel of the sensor unit 802 outputs a pixel signal corresponding to the irradiated light. The sensor unit 802 further includes a drive circuit for driving each pixel included in the pixel array, and a signal processing circuit for performing predetermined signal processing on signals read from each pixel and outputting the result as a pixel signal for each pixel. The sensor unit 802 outputs the pixel signal of each pixel included in the pixel area as image data in digital format.

[0071] The sensor control unit 803 is configured by, for example, a microprocessor, controls the reading of pixel data from the sensor unit 802, and outputs image data based on each pixel signal read from each pixel. The pixel data output from the sensor control unit 803 is passed to the recognition processing unit 804 and the visual recognition processing unit 806.

[0072] The sensor control unit 803 also generates an imaging control signal for controlling imaging in the sensor unit 802 and supplies it to the sensor unit 802. The imaging control signal includes information indicating exposure and analog gain when imaging in the sensor unit 802. The imaging control signal further includes control signals for performing imaging operations of the sensor unit 802, such as a vertical synchronization signal and a horizontal synchronization signal.

[0073] The recognition processing unit 804 performs recognition processing (person detection, face identification, image classification, etc.) of objects included in an image made of pixel data based on the pixel data passed from the sensor control unit 803. However, the recognition processing unit 804 may perform recognition processing using image data after visual recognition processing by the visual recognition processing unit 806. The recognition result by the recognition processing unit 804 is passed to the output control unit 807.

[0074] In this embodiment, the recognition processing unit 804 performs recognition processing using a machine learning model. Model parameters obtained by prior model training are stored in the memory 805, and the recognition processing unit 804 performs recognition processing using a model set with model parameters read from the memory 805. Furthermore, if the model parameters used by the recognition processing unit 804 cannot guarantee fairness in recognition results for pixel data or image data with a minority attribute, additional model training may be performed using Adversarial Examples generated from existing (or original) data with a minority attribute.

[0075] The visual recognition processing unit 806 processes the pixel data passed from the sensor control unit 803 to obtain an image suitable for human visual recognition, and outputs, for example, image data consisting of a group of pixel data. For example, if a color filter is provided for each pixel included in the sensor unit 802 and each pixel data has color information of R (red), G (green), or B (blue), the visual recognition processing unit 806 performs demosaic processing, white balance processing, etc. Furthermore, the visual recognition processing unit 806 can instruct the sensor control unit 803 to read pixel data required for the visual recognition processing from the sensor unit 802. The visual recognition processing unit 806 passes the image data resulting from the pixel data processing to the output control unit 807. For example, the above-mentioned functions of the visual recognition processing unit 806 are realized by an image signal processor executing a program pre-stored in a local memory (not shown).

[0076] The output control unit 807 is configured with, for example, a microprocessor. The output control unit 807 receives the recognition result of an object included in an image from the recognition processing unit 804 and image data as a visual recognition processing result from the visual recognition processing unit 806, and outputs one or both of them to the outside of the imaging device 800. The output control unit 807 also outputs the image data to the display unit 808. The user can visually recognize the image displayed on the display unit 808. The display unit 808 may be built in the imaging device 800 or may be externally connected to the imaging device 800.

[0077] Fig. 9 shows an example of hardware implementation of an imaging device 800. In the example shown in Fig. 9, a sensor unit 802, a sensor control unit 803, a recognition processing unit 804, a memory 805, a visual recognition processing unit 806, and an output control unit 807 are mounted on a single chip 900. However, in Fig. 9, the memory 805 and the output control unit 807 are omitted from illustration to avoid confusion in the drawing.

[0078] 9, the recognition result by the recognition processing unit 804 is output to the outside of the chip 900 via the output control unit 807. In addition, the recognition processing unit 804 can acquire pixel data or image data to be used for recognition from the sensor control unit 803 via an interface inside the chip 900.

[0079] Fig. 10 shows another example of hardware implementation of the imaging device 800. In the example shown in Fig. 10, a sensor unit 802, a sensor control unit 803, a visual recognition processing unit 806, and an output control unit 807 are mounted on a single chip 1000, but a recognition processing unit 804 and a memory 805 are arranged outside the chip 1000. However, in Fig. 10 as well, the memory 805 and the output control unit 807 are omitted from illustration to avoid confusion in the drawing.

[0080] 10, the recognition processing unit 804 acquires pixel data or image data to be used for recognition from the output control unit 807 via the inter-chip communication interface. The recognition processing unit 804 also directly outputs the recognition result to the outside. Of course, the recognition result by the recognition processing unit 804 can also be configured to be returned to the output control unit 807 in the chip 1000 via the inter-chip communication interface and output from the output control unit 807 to the outside of the chip 1000.

[0081] In the configuration example shown in Fig. 9, the recognition processing unit 804 and the sensor control unit 803 are both mounted on the same chip 900, so communication between the recognition processing unit 804 and the sensor control unit 803 can be performed at high speed via an interface within the chip 900. On the other hand, in the configuration example shown in Fig. 10, the recognition processing unit 804 is arranged outside the chip 1000, so the recognition processing unit 804 can be easily replaced, and the learning model can be replaced by replacing it. However, communication between the recognition processing unit 804 and the sensor control unit 803 must be performed via an interface between the chips, which is slow.

[0082] 11 shows an example in which semiconductor chips 900 (or 1000) of an imaging device 800 are stacked in two layers to form a two-layer stacked image sensor 1100. In the structure shown, a pixel section 1111 is formed in a first-layer semiconductor chip 1101, and a memory and logic section 1112 is formed in a second-layer semiconductor chip 1102.

[0083] The pixel unit 1111 includes at least a pixel array in the sensor unit 802. The memory and logic unit 1112 includes, for example, a sensor control unit 803, a recognition processing unit 804, a memory 805, a visual recognition processing unit 806, an output control unit 807, and an interface for communicating between the imaging device 800 and the outside. The memory and logic unit 1112 also includes a part or all of a drive circuit that drives the pixel array in the sensor unit 802. Although not shown in FIG. 11 , the memory and logic unit 1112 may further include, for example, a memory used by the visual recognition processing unit 806 to process image data.

[0084] As shown on the right side of FIG. 11, a first layer semiconductor chip 1101 and a second layer semiconductor chip 1102 are bonded together while being in electrical contact with each other, thereby forming an imaging device 800 as a single solid-state imaging element.

[0085] 12 shows an example in which semiconductor chips 900 (or 1000) of an imaging device 800 are stacked in three layers to form a two-layer stacked image sensor 1200. In the structure shown, a pixel section 1211 is formed in a first-layer semiconductor chip 1201, a memory section 1212 is formed in a second-layer semiconductor chip 1202, and a logic section 1213 is formed in a third-layer semiconductor chip 1203.

[0086] The pixel unit 1211 includes at least the pixel array in the sensor unit 802. The logic unit 1213 includes, for example, a sensor control unit 803, a recognition processing unit 804, a visual recognition processing unit 806, an output control unit 807, and an interface for communicating between the imaging device 800 and the outside. The logic unit 1213 also includes a part or all of a drive circuit that drives the pixel array in the sensor unit 802. The memory unit 1212 may further include, in addition to the memory 805, a memory that the visual recognition processing unit 806 uses to process image data, for example.

[0087] As shown on the right side of Figure 12, the imaging device 800 is constructed as a single solid-state imaging element by bonding together a first layer semiconductor chip 1201, a second layer semiconductor chip 1202, and a third layer semiconductor chip 1203 while maintaining electrical contact between them.

[0088] 13 shows an example of the configuration of the sensor unit 802. The sensor unit 802 shown in the figure includes a pixel array unit 1301, a vertical scanning unit 1302, an AD (Analog to Digital) conversion unit 1303, a horizontal scanning unit 1304, pixel signal lines 1305, vertical signal lines VSL, a control unit 1306, and a signal processing unit 1307. Note that the control unit 1306 and the signal processing unit 1307 in FIG. 13 may be included in, for example, the sensor control unit 803 in FIG. 8.

[0089] The pixel array unit 1301 is composed of a plurality of pixel circuits 1310, each of which includes a photoelectric conversion element that performs photoelectric conversion on received light and a circuit that reads out electric charges from the photoelectric conversion element. The plurality of pixel circuits 1310 are arranged in a matrix array in the horizontal direction (row direction) and vertical direction (column direction). The arrangement of pixel circuits 1310 in the row direction forms a line. For example, when one frame of image is formed with 1920 pixels x 1080 lines, the pixel array unit 1301 forms one frame of image using pixel signals read out from 1080 lines of lines each consisting of 1920 pixel circuits 1310.

[0090] In the pixel array unit 1301, pixel signal lines 1305 are connected to the rows and columns of each pixel circuit 1310, and vertical signal lines VSL are connected to the columns. The ends of each pixel signal line 1305 that are not connected to the pixel array unit 1301 are connected to a vertical scanning unit 1302. Under the control of a control unit 1306, the vertical scanning unit 1302 transmits control signals such as drive pulses used to read pixel signals from pixels to the pixel array unit 1301 via the pixel signal lines 1305. The ends of the vertical signal lines VSL that are not connected to the pixel array unit 1301 are connected to an AD conversion unit 1303. The pixel signals read from the pixels are transmitted to the AD conversion unit 1303 via the vertical scanning lines VSL.

[0091] A pixel signal is read from the pixel circuit 1310 by transferring charges accumulated in a photoelectric conversion element due to exposure to a floating diffusion layer (FD) and converting the transferred charges into a voltage in the floating diffusion layer. The voltage converted from the charges in the floating diffusion layer is output to a vertical signal line VSL via an amplifier.

[0092] The AD conversion unit 1303 includes a column AD converter (ADC) 1311 provided for each vertical signal line VSL, a reference signal generation unit 1312, and a horizontal scanning unit 1304. The column AD converter 1311 is a column AD converter that performs AD conversion processing for each column of the pixel array unit 1301, and performs AD conversion processing on pixel signals supplied from the pixel circuits 1310 via the vertical signal lines VSL to generate two digital values ​​for correlated double sampling (CDS) processing that reduces noise and outputs them to the signal processing unit 1307.

[0093] The reference signal generating unit 1312 generates a ramp signal as a reference signal, which is used by each column AD converter 1311 to convert a pixel signal into two digital values, based on a control signal from the control unit 1306, and supplies the ramp signal to each column AD converter 1311. The ramp signal is a signal whose voltage level decreases at a constant slope over time, or a signal whose voltage level decreases in a step-like manner.

[0094] Within the column AD converter 1311, when a ramp signal is supplied, a counter starts counting in accordance with a clock signal, compares the voltage of the pixel signal supplied from the vertical signal line VSL with the voltage of the ramp signal, stops counting by the counter when the voltage of the ramp signal crosses the voltage of the pixel signal, and outputs a value corresponding to the count value at that time, thereby converting the pixel signal, which is an analog signal, into a digital value.

[0095] The signal processing unit 1307 performs CDS processing based on the two digital values ​​generated by the column AD converter 1311, generates a digital pixel signal (pixel data), and outputs it to the outside of the sensor control unit 803.

[0096] Under the control of the control unit 1306, the horizontal scanning unit 1304 performs a selection operation to select each column AD converter 1311 in a predetermined order, thereby causing each column AD converter 1311 to sequentially output the digital values ​​temporarily held therein to the signal processing unit 1307. The horizontal scanning unit 1304 is configured using, for example, a shift register, an address decoder, etc.

[0097] Based on the imaging control signal supplied from the sensor control unit 803, the control unit 1306 generates drive signals for controlling the driving of the vertical scanning unit 1302, the AD conversion unit 1303, the reference signal generation unit 1312, the horizontal scanning unit 1304, etc., and outputs the drive signals to each unit. For example, based on the vertical synchronization signal and the horizontal synchronization signal included in the imaging control signal, the control unit 1306 generates control signals for the vertical scanning unit 1302 to supply to each pixel circuit 1310 via the pixel signal line 1305, and supplies the control signals to the vertical scanning unit 1302. The control unit 1306 also passes information indicating an analog gain included in the imaging control signal to the AD conversion unit 1303. Based on this information indicating the analog gain, the AD conversion unit 1303 controls the gain of the pixel signal input to each column AD converter 1311 via the vertical signal line VSL.

[0098] The vertical scanning unit 1302 supplies various signals, including drive pulses, to the pixel signal lines 1305 of selected pixels in the pixel array unit 1301 for each line to each pixel circuit 1310, based on a control signal supplied from the control unit 1306, causing each pixel circuit 1310 to output a pixel signal to a vertical signal line VSL. The vertical scanning unit 1302 is configured using, for example, a shift register, an address decoder, etc. Furthermore, the vertical scanning unit 1302 controls the exposure of each pixel circuit 1310 based on information indicating exposure supplied from the control unit 1306.

[0099] The sensor unit 802 configured as shown in FIG. 13 is a column AD type image sensor in which column AD converters 1311 are arranged for each column.

[0100] The configuration of the imaging device 800 with an image recognition function has been described with reference to FIGS. 8 to 13. When the imaging device 800 is applied to, for example, an image classification service, a model used by the recognition processing unit 804 is trained in advance using a training dataset. Here, it is expected that sufficient training cannot be performed on data with a minority attribute included in the original dataset, resulting in a lower recognition rate compared to data with other attributes. In response to this, in the present disclosure, adversarial examples related to data with a minority attribute are generated and added to the dataset for training, thereby making it possible to alleviate bias in the original dataset.

[0101] FIG. 14 shows, in the form of a flowchart, an example of the operation in the learning phase for applying the image capturing device 800 with a recognition function to an image classification service.

[0102] First, the original data set is input, and for data with a minority attribute, additional data is generated using an Adversarial Example according to the above formula (2) or (4) (step S1401).

[0103] Then, the recognition processing unit 804 uses the original data set and the additional data set to train a model (step S1402). The model trained in this way can improve the recognition rate even for minority attributes, and can perform image classification while ensuring fairness.

[0104] G. Application to in-vehicle cameras When driving a vehicle, it is necessary to detect various objects such as surrounding vehicles, pedestrians, lanes, etc. For this reason, technologies have been developed that equip in-vehicle cameras with object recognition functions, notify the driver of the recognition results, and control the vehicle based on the recognition results (see, for example, Patent Document 2).

[0105] 15 schematically shows an example of the functional configuration of an in-vehicle camera 1500. The in-vehicle camera 1500 shown in the figure includes a lens 1501, an image sensor 1502, a signal processing unit 1503, a recognition unit 1504, and a control unit 1505.

[0106] The image sensor 1502 is configured using elements such as a CMOS, and captures an image formed on an imaging surface by the lens 1501. The signal processing unit 1503 performs signal processing on the RAW data output from the image sensor 1502. The signal processing performed by the signal processing unit 1503 corresponds to, for example, demosaic, noise reduction, white balance adjustment, gamma correction, sensor spectral correction, YC conversion, and the like.

[0107] The recognition unit 1504 recognizes objects included in the captured image after processing by the signal processing unit 1503. The recognition unit 1504 recognizes various objects, such as motorbikes, bicycles, pedestrians, road signs, traffic lights, lanes, medians, guardrails, road trees, and streetlights. The recognition unit 1504 performs object recognition processing using a trained model configured by a neural network or the like.

[0108] The control unit 1505 comprehensively controls the operation of each unit in the in-vehicle camera 1500. The control unit 1505 controls, for example, the image capturing operation in the image sensor 1502 and the signal processing in the signal processing unit 1503. The control unit 1505 may also add, delete, or change the objects to be recognized by the recognition unit 1504.

[0109] A vehicle control system 1510 at the subsequent stage controls the operation of the vehicle under the control of a control unit 1505, based on images captured by an image sensor 1502 and developed by a signal processing unit 1503, and on the recognition results of a recognition unit 1504. The vehicle control referred to here includes, for example, vehicle control for autonomous driving or ADAS, such as automatic distance control (ACC), lane departure warning (LDW), lane keep assist (LKA), automatic emergency braking (AEB), and blind spot detection (BSD), as well as drive control of each drive unit, such as active cornering lights (ACL), brake actuators (BRK), and steering devices (STR).

[0110] FIG. 16 shows an example of the internal configuration of the image sensor 1502 and the signal processing unit 1503.

[0111] The image sensor 1502 includes a shutter 1601, an element unit 1602, and an analog gain processing unit 1603. Light collected by the lens 1501 passes through the shutter 1601 and reaches the imaging surface of the element unit 1602. The element unit 1602 is made up of a two-dimensional pixel array, and each pixel outputs a pixel signal according to the amount of received light. Each pixel signal is amplified in the analog domain by the analog gain processing unit 1603, then converted to a digital signal and output to the signal processing unit 1603.

[0112] The signal processing unit 1603 includes a development processing unit 1604, a detection unit 1605, and a comparison unit 1606. The development processing unit 1604 performs development processing, including digital gain processing and gamma processing, on the digital pixel signals output from the image sensor 1502. The detection unit 1605 detects the brightness (luminance) of the screen by performing OPD (Optical Detection) detection on the entire screen captured by the image sensor 1502. The comparison unit 1606 then compares the brightness of the entire screen detected by the detection unit 1605 with a predetermined reference value (Ref).

[0113] Based on the difference between the screen brightness output from the comparison unit 1606 and the reference value, the control unit 1505 controls the opening and closing timing of the shutter 1601 (i.e., exposure time), adjusts the analog gain of the analog gain processing unit 303, and adjusts the digital gain and other development parameters in the development processing unit 1604, thereby controlling the image captured by the image sensor 1502 to have appropriate brightness.

[0114] Note that the vehicle-mounted camera 1500 is not necessarily intended to capture images for observation by a user (such as a driver), but is primarily intended to acquire image information that can be used by a downstream vehicle control system 1510. For this reason, the development processing performed by the signal processing unit 1503 of the vehicle-mounted camera 1500 does not need to be the same as that of the imaging device 800.

[0115] The in-vehicle camera 1500 is equipped with an object recognition function and recognizes various objects, such as motorcycles, bicycles, pedestrians, road signs, traffic lights, lanes, medians, guardrails, roadside trees, and streetlights. The model used by the recognition unit 1504 is trained in advance using a training dataset. It is expected that sufficient training cannot be performed on data with a small number of attributes included in the original dataset, resulting in a lower recognition rate compared to data with other attributes. For example, the model used by the recognition unit 1504 is trained to recognize humans, but there is an imbalance in the amount of data between images of children and adults, and the recognition rate for children, for which there is a small amount of training data, is expected to be low. In contrast, the present disclosure generates images that the model used by the recognition unit 1504 recognizes as children using adversarial examples and adds these images to the image analysis, thereby enabling training of a model with an improved child recognition rate.

[0116] FIG. 17 shows, in the form of a flowchart, an example of the operation for training the recognition unit 1504 of the vehicle-mounted camera 1500 to recognize humans in the learning phase.

[0117] First, the original data set is input, and for images of children with a small number of samples, additional data is generated using Adversarial Examples according to the above formula (2) or (4) (step S1701).

[0118] The recognition unit 1504 then uses the original data set and the additional data set to train a model (step S1702). The model trained in this way can improve the recognition rate even for images of children, and can perform recognition while ensuring fairness for all humans.

[0119] H. Applications to Healthcare Systems There are high expectations for the use of artificial intelligence in the medical field. For example, trained models can recognize and process endoscopic images to assist in surgery. Based on the recognition results of trained models, doctors can appropriately proceed with surgery or control the operation of surgical robots.

[0120] 18 shows an example of the configuration of a medical robot device 1800 using a robot arm. The medical robot device 1800 comprises a robot arm 1810 and a control device 1820 that controls the operation of the robot arm 1810.

[0121] The robot arm 1810 includes one or more robot arms each having a multi-link structure in which multiple links are connected by joint axes. To simplify the drawing, only one robot arm is shown in Fig. 18. Medical surgical tools such as an endoscope, forceps, an insufflation tube, an energy treatment device, a surgical excisor, and a retractor are mounted on the tip of the robot arm.

[0122] The control device 1820 includes an image recognizer 1821 and a movement predictor 1822. The image recognizer 1821 performs image recognition on an image captured by the endoscope. The movement predictor 1822 predicts the movement of the robot arm 1810 in accordance with the recognition result of the image recognizer 1821.

[0123] The control device 1820 receives as input an image of the surgical site captured by the endoscope, and robot arm motion information and robot arm sensor information from the robot arm 1810. The robot arm motion information includes information on the position, speed, and acceleration of a medical instrument such as an endoscope supported at the tip of the robot arm, and the posture of each joint of the robot arm (joint angle measured by an encoder attached to the rotation axis of the joint). The robot arm sensor information also includes information such as acceleration measured by an IMU (Inertial Measurement Unit) mounted on the robot arm 1810, torque information acting on each joint, and information on external forces acting on the medical instrument supported at the tip of the robot arm.

[0124] The image recognizer 1821 uses a model trained to perform image recognition to perform image recognition of medical instruments included in images captured by the endoscope and the environment within the field of view of the endoscope, and outputs instrument recognition information and environment recognition information. In this embodiment, the image recognizer 1821 performs on-site user-specific model learning using images captured by the endoscope, and further performs image recognition specialized to the user's needs using the trained model.

[0125] The image recognizer 1821 recognizes, as instrument recognition information, the types of medical instruments recognized within the field of view of the endoscope (e.g., forceps, insufflation tube, energy treatment instrument, pliers, retractor, etc.), the position and posture of each instrument, and the operation state (e.g., the open / closed state for forceps, and the energy output state for energy treatment instruments). Furthermore, the image recognizer 1821 recognizes, as environment recognition information, depth information (including the shapes of organs and instruments) of organs and medical instruments included in captured images within the field of view of the endoscope, an environmental map of the surgical site (e.g., an environmental map created using SLAM (Simultaneous Localization and Mapping) technology), the types of organs, the types of medical instruments, and the materials of each object included in the captured images. Furthermore, the image recognizer 1821 recognizes, as environment recognition information, for example, each object, such as organs and medical instruments, included in images of the surgical site, and their materials, the depth information of each object, and the environmental map.

[0126] The motion predictor 1822 uses a model trained to predict the motion of the robot arm from the image recognition results to predict and output target command-related information for the robot arm 1810 based on the recognition information of the instrument recognition information and the environment recognition information. The motion predictor 1822 predicts various target command values ​​as the target command-related information, such as the target position, posture, speed, acceleration, gaze point, line-of-sight vector (object position, distance, vector posture), and electronic crop position and distance of the captured image. The motion predictor 1822 also predicts the target position, posture, speed, acceleration, and operating force of the instrument as the target command-related information. The control device 1820 then calculates the target joint angle, joint angular velocity, and joint angular acceleration of each joint of the robot arm by inverse kinematics calculation based on the target position, posture, speed, and acceleration information of the medical instrument, such as an endoscope, supported at the tip of the robot arm, predicted by the motion predictor 1822, and outputs command values ​​for the robot arm 1810.

[0127] The model used by the image recognizer 1821 is trained in advance using a training dataset so that it can recognize instrument information and environmental information from endoscopic images. Here, it is expected that sufficient training cannot be performed on data with a small number of attributes included in the original dataset, resulting in a lower recognition rate compared to data with other attributes. For example, many images of successful examples are provided as training data, but few images of unsuccessful examples. In contrast, in the present disclosure, images that the image recognizer 1821 recognizes as unsuccessful examples are generated using adversarial examples, and these images are added to the training data, thereby enabling training of a model with improved recognition accuracy.

[0128] FIG. 19 shows, in the form of a flowchart, an example of the operation for training the image recognizer 1821 to recognize endoscopic images in the learning phase.

[0129] First, the original data set is input, and for images of failure examples with a small number of samples, additional data is generated using Adversarial Examples according to the above formula (2) or (4) (step S1901).

[0130] Then, the image recognizer 1821 uses the original data set and the additional data set to train a model (step S1902). The model trained in this way improves the recognition rate even for images of failure cases, and can perform recognition while ensuring fairness for all endoscopic images. [Industrial Applicability]

[0131] Although the present disclosure has been described in detail above with reference to specific embodiments, it is obvious that those skilled in the art can make modifications or substitutions to the embodiments without departing from the spirit and scope of the present disclosure.

[0132] The present disclosure can be applied to learning machine learning models that mainly perform image classification, and a machine learning model to which the present disclosure is applied can be installed in, for example, an imaging device. For example, a machine learning model to which the present disclosure is applied can be used as a recognizer for an in-vehicle camera or for image recognition of an operating site in the medical field.

[0133] In short, the present disclosure has been described in the form of examples, and the contents of the specification should not be interpreted as limiting. To determine the gist of the present disclosure, the claims should be taken into consideration.

[0134] The present disclosure may also be configured as follows.

[0135] (1) an acquisition unit that acquires information about bias in training data used for model training; a generation unit that generates additional training data from data included in the training data based on information about the bias; a learning unit that learns the model using the learning data and the additional learning data; An information processing device comprising:

[0136] (2) the acquiring unit acquires information indicating a minority attribute of the first learning data; the data generation unit generates second learning data having the same attribute from the data having the minority attribute included in the first learning data. The information processing device according to (1) above.

[0137] (3) the data generation unit generates Adversarial Examples to be used as second training data from the data of the minority attributes included in the training data; The information processing device according to any one of (1) and (2) above.

[0138] (4) The data generation unit generates Adversarial Examples based on a Fast Gradient Sign Method. The information processing device according to (3) above.

[0139] (5) The data generation unit generates second training data by superimposing noise on data included in the training data. An information processing device according to any one of (1) to (4) above.

[0140] (6) inputting first learning data to be used for learning the machine learning model; acquiring information about bias in the learning data; generating second training data using data included in the training data based on information about the bias; training the machine learning model using the first training data and the second training data; An information processing method comprising:

[0141] (7) a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the learning data; a data generation unit that generates second training data using data included in the training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; A computer program written in a computer-readable format to cause a computer to function as a

[0142] (8) an imaging unit that captures an image; a recognition unit that recognizes the captured image using a machine learning model; Equipped with generating training data using data included in the training data based on information about bias in the training data for training the machine learning model, and training the machine learning model using the generated training data; Imaging device.

[0143] (9) generating training data using a small number of image data from the training data according to the field to which the imaging device is applied, and training the machine learning model using the generated training data; The imaging device according to (8) above.

[0144] (10) a recognition unit that recognizes the captured image using a machine learning model; Equipped with an imaging device that generates training data using data included in the training data based on information about bias in the training data for training the machine learning model, and trains the machine learning model using the generated training data; A vehicle device comprising:

[0145] (11) An imaging unit that captures an image of the area around the surgical site; a recognition unit that recognizes the captured image using a machine learning model; Equipped with an imaging device that generates training data using data included in the training data based on information about bias in the training data for training the machine learning model, and trains the machine learning model using the generated training data; A medical robotic device comprising: [Explanation of symbols]

[0146] 100... learning system, 101... learning data storage unit 102... learning unit, 103... model parameter holding unit 104...analysis unit, 105...data generation unit, 111...inference unit 112...data input unit, 113...input data processing unit 800...imaging device, 801...optical section, 802...sensor section 803: Sensor control unit, 804: Recognition processing unit, 805: Memory 806...visibility processing unit, 807...output control unit, 808...display unit 1301: pixel array section, 1302: vertical scanning section 1303...AD conversion unit, 1304...horizontal scanning unit 1305: pixel signal line; 1306: control unit 1307...signal processing unit, 1310...pixel circuit 1311... column AD converter, 1312... reference signal generating unit 1500...In-vehicle camera, 1501...Lens 1502...image sensor, 1503...signal processing unit 1504...recognition unit, 1505...control unit 1510...Vehicle control system 1601...shutter, 1602...element part 1603...analog gain processing unit, 1604...development processing unit 1605...detection section, 1606...comparison section 1800...Medical robotic device, 1810...Robot arm 1820...controller, 1821...image recognizer, 1822...motion predictor

Claims

1. a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the first learning data; a data generation unit that generates second training data using data included in the first training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; Equipped with the acquiring unit acquires information indicating attributes in which sensitive attributes are in a minority among the first learning data, the data generation unit generates second learning data having the same sensitive attribute from data having the minority sensitive attribute included in the first learning data; Information processing device.

2. The data generation unit generates Adversarial Examples to be used as second learning data from data of attributes in which the number of sensitive attributes included in the first learning data is the small number. The information processing device according to claim 1 .

3. The data generation unit generates an Adversarial Example based on a Fast Gradient Sign Method. The information processing device according to claim 2 .

4. The data generation unit generates second training data by superimposing noise on data included in the first training data. The information processing device according to claim 1 .

5. A step of inputting first training data to be used for training a machine learning model; an acquisition step of acquiring information about bias in the first learning data; a data generating step of generating second training data using data included in the first training data based on information about the bias; a learning step of learning the machine learning model using the first learning data and the second learning data; and In the acquiring step, information indicating attributes in which the number of sensitive attributes is small among the first learning data is acquired, In the data generating step, second learning data having the same sensitive attribute is generated from data having the same sensitive attribute as the first learning data, the data having the same sensitive attribute as the first learning data. Information processing methods.

6. A data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the first learning data; a data generation unit that generates second training data using data included in the first training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; Make the computer function as the acquiring unit acquires information indicating attributes in which sensitive attributes are in a minority among the first learning data, the data generation unit generates second learning data having the same sensitive attribute from data having the minority sensitive attribute included in the first learning data; A computer program written in a computer-readable form so as to

7. An imaging unit that captures an image; a recognition unit that recognizes the captured image using a machine learning model; a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the first learning data; a data generation unit that generates second training data using data included in the first training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; Equipped with the acquiring unit acquires information indicating attributes in which sensitive attributes are in a minority among the first learning data, the data generation unit generates second learning data having the same sensitive attribute from data having the minority sensitive attribute included in the first learning data; Imaging device.

8. An imaging unit that captures an image of the area around the vehicle; a recognition unit that recognizes the captured image using a machine learning model; a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the first learning data; a data generation unit that generates second training data using data included in the first training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; an imaging device comprising: the acquiring unit acquires information indicating attributes in which sensitive attributes are in a minority among the first learning data, the data generation unit generates second learning data having the same sensitive attribute from data having the minority sensitive attribute included in the first learning data; Vehicle equipment.

9. An imaging unit that captures an image of the area around the surgical site; a recognition unit that recognizes the captured image using a machine learning model; a data storage unit that stores first learning data used for learning the machine learning model; an acquisition unit that acquires information about bias in the first learning data; a data generation unit that generates second training data using data included in the first training data based on information about the bias; a learning unit that learns the machine learning model using the first learning data and the second learning data; an imaging device comprising: the acquiring unit acquires information indicating attributes in which sensitive attributes are in a minority among the first learning data, the data generation unit generates second learning data having the same sensitive attribute from data having the minority sensitive attribute included in the first learning data; A medical robotic device comprising:

Images