Information processing device and information processing method

Abstract

An information processing device according to the present disclosure is provided with a machine learning model in which parameters for performing inference have been incorporated, and incorporates, into the machine learning model, other parameters that have been estimated by another machine learning model on the basis of training data that is desired to be invalid in relation to a training data set that includes training data used for training the machine learning model.

Description

Information processing device and information processing method

[0001] The present disclosure relates to an information processing device and an information processing method.

[0002] In inference processing using an AI (Artificial Intelligence) model, the accuracy of inference can be improved by training the AI ​​model using a training data set that includes more training data.

[0003] On the other hand, when performing inference using a trained AI model, there may be cases where it is not desirable to reflect, for example, part of the training data in the inference results.

[0004] As one example, in an AI model applied to a camera that performs face recognition to focus on a face, an object (such as a stuffed animal) in the background of the face may be recognized as a face, resulting in a situation where the intended face is not in focus. As another example, the rights holder of some training data (e.g., image data) included in a training dataset for training an AI model may not permit the use of that image data as training data.

[0005] Japanese Patent Application Laid-Open No. 2022-057202

[0006] Rohit Gandikota, and 3 others, "Erasing concepts from diffusion models.", [online], June 21, 2020, arXiv preprint arXiv:2303.07345 (2023), [Retrieved September 28, 2020], Internet, <https: / / arxiv.org / abs / 2303.07345>

[0007] Conventionally, when it is determined that it is undesirable to reflect, for example, a portion of the training data in the inference result, it has been necessary to delete the training data from the training dataset and retrain the AI ​​model. In particular, in the case of an AI model trained using a training dataset that includes a huge amount of training data, retraining incurs a huge cost.

[0008] Therefore, an object of the present disclosure is to provide an information processing device and an information processing method that can easily realize inference processing using an AI model in which specific learning data is invalidated.

[0009] The information processing device according to the present disclosure includes a machine learning model incorporating parameters for performing inference, and incorporates into the machine learning model other parameters estimated by other machine learning models based on training data that is to be invalidated for a training dataset that includes training data used to train the machine learning model.

[0010] 1 is a schematic diagram for generally explaining an application example of a model according to an existing technology. FIG. 1 is a schematic diagram for generally explaining an application example of a model according to a first embodiment. FIG. 2 is a flowchart of an example of a redeployment process of a model according to an existing technology. FIG. 3 is a flowchart of an example of a redeployment process of a model according to the first embodiment. FIG. 4 is a schematic diagram showing the configuration of an example of an information processing system applicable to the first embodiment. FIG. 5 is a block diagram showing the configuration of an example of a camera applicable to the first embodiment. FIG. 6 is a block diagram showing the configuration of an example of a server applicable to the first embodiment. FIG. 7 is a flowchart of an example of a procedure for training a machine learning model for parameter estimation according to the first embodiment. FIG. 8 is a schematic diagram showing the configuration of a machine learning model for face recognition before parameter replacement, applicable to the first embodiment. FIG. 9 is a schematic diagram showing the configuration of a machine learning model for face recognition after parameter replacement, applicable to the first embodiment. FIG. 10 is a flowchart of an example of a parameter update process according to the first embodiment. FIG. 11 is a schematic diagram showing the parameter update process according to the first embodiment. FIG. 12 is a schematic diagram showing the parameter update process according to the first embodiment. FIG. 13 is a schematic diagram showing an example of adding additional parameters in series to original parameters of a machine learning model for face recognition, as a modification of the first embodiment. 1 is a flowchart of an example of processing for opt-out according to an existing technology. FIG. 2 is a flowchart of an example of processing for opt-out according to a second embodiment. FIG. 3 is a schematic diagram showing a configuration of an example of an information processing system applicable to the second embodiment. FIG. 4 is a schematic diagram for explaining image generation processing using a diffusion generation model according to an existing technology. FIG. 5 is a schematic diagram for generally explaining an application example of a model according to the second embodiment. FIG. 6 is a schematic diagram for generally explaining a learning method for a diffusion generation model. FIG. 7 is a flowchart of an example of a procedure for learning a machine learning model for parameter estimation according to the second embodiment. FIG. 8 is a flowchart of an example of parameter update processing according to the second embodiment. FIG. 9 is a schematic diagram for explaining parameter update processing according to the second embodiment.FIG. 10 is a schematic diagram for explaining a parameter update process according to the second embodiment.

[0011] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the drawings. In the following embodiments, the same components are denoted by the same reference numerals, and redundant description will be omitted.

[0012] Hereinafter, embodiments of the present disclosure will be described in the following order: 1. Overview of Technology According to the Present Disclosure 2. First Embodiment of the Present Disclosure 2-1. Overview of the First Embodiment 2-2. More Specific Description of the First Embodiment 2-2-1. Configuration Applicable to the First Embodiment 2-2-2. Processing According to the First Embodiment 2-3. Modified Examples of the First Embodiment 2-4. Other Application Examples of the First Embodiment 3. Second Embodiment of the Present Disclosure 3-1. Background of the Second Embodiment 3-2. Overview of the Second Embodiment 3-3. More Specific Description of the Second Embodiment 3-3-1. Configuration Applicable to the Second Embodiment 3-3-2. Processing According to the Second Embodiment 3-3-3. Example of a System Configuration Applicable to the Second Embodiment 3-4. Modified Examples of the Second Embodiment

[0013] (1. Overview of the Technology Relating to the Present Disclosure) The present disclosure involves inputting data representing information that is to be forgotten into a pre-trained AI (Artificial Intelligence) model (hereinafter referred to as an original AI model), and training another AI model that outputs the weights (parameters) of the AI ​​model when the AI ​​model successfully forgets. By replacing the parameters of the original AI model with the output of the trained other AI model, it becomes possible to cause the original AI model to forget specified data or concepts.

[0014] According to the technology disclosed herein, there is no need to perform learning each time the need arises for the AI ​​model to forget, and the forgetting process can be achieved in a short time and at low cost.

[0015] Note that "making an AI model forget (certain information)" refers to disabling the learning of specific information for a trained AI model. In other words, "making an AI model forget" refers to disabling learning using specific information for a trained AI model that has been trained using a training dataset that includes the specific information (training data).

[0016] For example, if an AI model trained to recognize human faces from photographic images has learned that an object that is not a human face but is mistakenly recognized as a human face is a human face, the learning for that object will be invalidated (forgotten).Furthermore, in an AI model trained using training data X1 to X4, if the rights holder of training data X3, for example, prohibits the use of training data X3 for training, the learning using training data X3 will be invalidated (forgotten).

[0017] Not limited to the above, "causing the AI ​​model to forget" may mean causing a trained AI model to forget learning using specific information, or may have an effect equivalent to that, or may mean making the specific information used to train the trained AI model nonexistent, or may have an effect equivalent to that. Furthermore, "causing the AI ​​model to forget" may mean making a trained AI model that has been trained using specific information return to a state where it has been trained without using the specific information, or may have an effect equivalent to that.

[0018] Machine unlearning technology is known as a technology for causing an AI model to forget data. For example, Patent Document 1 describes a technology for training an AI model trained using specific data so that the output of the AI ​​model when the specific data is input is the same as the output of an AI model that has not been trained using the specific data. The technology described in Patent Document 1 is similar to the present disclosure in that it causes a trained AI model to forget specific data, but requires training for each piece of data to be forgotten.

[0019] Furthermore, Non-Patent Document 1 describes a technique for fine-tuning weights in a diffusion model-based generative AI model I so that images unrelated to specified text are generated. The technique described in Non-Patent Document 1 is common to the present disclosure in that it causes a trained AI model to forget specific data, but requires training for each piece of data to be forgotten.

[0020] In contrast, according to the technology disclosed herein, parameters obtained by training another AI model on the target original AI model are replaced with the parameters of the original AI model, causing the original AI model to forget (invalidate) learning related to specific information. Therefore, by applying the technology disclosed herein, it is not necessary to train the original AI model for each piece of data to be forgotten, and it is possible to achieve forgetting processing (unlearning processing) in a shorter time and at lower cost than existing technologies.

[0021] 2. First Embodiment of the Present Disclosure Next, a first embodiment of the present disclosure will be described. The first embodiment is an example in which the technology according to the present disclosure is applied to a camera.

[0022] Cameras are known that have the ability to automatically recognize and focus on human faces. One example of a malfunction of such a camera is when it recognizes an object (such as a stuffed animal) placed behind the camera as a face instead of a human face. By applying the technology disclosed herein to such a camera, it becomes possible to quickly correct such erroneous recognition.

[0023] 1 is a schematic diagram for outlining an example of an application of a model according to existing technology. As shown in Fig. 1, in a method for automatically recognizing a human face using a camera in existing technology, a face recognition machine learning model 20 (second machine learning model) is built into the camera, with a photographic image as input data 10 (second learning data) and the "probability that a face appears in the photograph" as output data 30.

[0024] In the first embodiment, a machine learning model for estimating parameters of the original machine learning model 20 for face recognition is prepared.

[0025] 2 is a schematic diagram for outlining an example of application of a model according to the first embodiment. As shown in Fig. 2, in the first embodiment, a parameter estimation machine learning model 40 (first machine learning model) that estimates parameters (weights, etc.) of the face recognition machine learning model 20 is prepared separately from the face recognition machine learning model 20 that is installed in a camera and performs face recognition.

[0026] 2, photographic images that are erroneously recognized as including a human face with a probability equal to or greater than a predetermined value are input as input data 11 (first learning data) to a parameter estimation machine learning model 40. In other words, the input data 11 are photographic images that should not be erroneously recognized as including a human face. Examples of such photographic images that should not be erroneously recognized as including a human face include stuffed toys and dolls that have facial features such as eyes, a nose, and a mouth.

[0027] The parameter estimation machine learning model 40 is trained based on the input data 11. More specifically, the parameter estimation machine learning model 40 is trained so that its output becomes "the parameter 21 (first parameter) of the face recognition machine learning model 20 when correction is made to prevent erroneous recognition while maintaining high recognition accuracy." The parameter estimation machine learning model 40 trained in this manner outputs the parameter 21. The parameter 21 output from the parameter estimation machine learning model 40 is incorporated into the face recognition machine learning model 20. For example, the parameter 21 is substituted for a parameter (second parameter) originally included in the face recognition machine learning model 20.

[0028] In the first embodiment, the parameters 21 of the face recognition machine learning model 20 mounted on the camera are replaced with the output of the parameter estimation machine learning model 40 trained using images that should not be misrecognized as training data. This allows the trained face recognition machine learning model 20 to be quickly corrected after being deployed to the target device (a camera in this example).

[0029] The model redeployment process according to the first embodiment will be compared with the model redeployment process according to existing technology using Figures 3 and 4. Figure 3 is a flowchart illustrating an example of the model redeployment process according to the existing technology. Figure 4 is a flowchart illustrating an example of the model redeployment process according to the first embodiment. Note that each process in Figures 3 and 4 is executed by, for example, an information processing device having a function for training a model.

[0030] 3, when the information processing device receives a request to correct a recognition error (step S10), it corrects (step S11a) the training data or training method used to train the face recognition machine learning model 20. The information processing device retrains the face recognition machine learning model 20 in accordance with the training data or training method corrected in step S11a (step S12a). The information processing device redeploys the retrained face recognition machine learning model 20 to the target device (a camera in this example) (step S13).

[0031] 4, when the information processing device receives a request to correct a recognition error (step S10), it estimates parameters 21 of the face recognition machine learning model 20 using the parameter estimation machine learning model 40 (step S11b). The information processing device replaces parameters 21 of the face recognition machine learning model 20 with the parameters estimated in step S11b (step S12b). The information processing device redeploys the face recognition machine learning model 20, whose parameters have been replaced, to the target device (step S13).

[0032] As described above, in the existing technology, modifying the face recognition machine learning model 20 after deployment requires re-learning of the face recognition machine learning model 20. On the other hand, in the first embodiment, re-learning is not necessary after deployment, and the face recognition machine learning model 20 can be modified simply by replacing the parameters (weights). Therefore, compared to the existing technology, it is possible to modify the model quickly and easily.

[0033] (2-2. More Specific Description of the First Embodiment) Next, the first embodiment will be described in more detail.

[0034] (2-2-1. Configuration Applicable to First Embodiment) A configuration applicable to the first embodiment will be described below. Fig. 5 is a schematic diagram showing the configuration of an example of an information processing system applicable to the first embodiment.

[0035] 5, the information processing system 1000 includes a server 1300 and a camera 1100 connected via a communication network 1002. The communication network 1002 may be the Internet or a local area network (LAN) established in a closed environment such as an in-house network. The camera 1100 is connected to the communication network 1002 via wireless communication such as Wi-Fi (Wireless Fidelity). However, the camera 1100 may also be connected to the communication network 1002 via wired communication.

[0036] The server 1300 may acquire the face recognition machine learning model 20 deployed in the camera 1100 from the camera 1100, for example, via the communication network 1002. The server 1300 may also have the above-described parameter estimation machine learning model 40, and may replace the parameters of the face recognition machine learning model 20 acquired from the camera 1100 with parameters estimated and output by the parameter estimation machine learning model 40.

[0037] 5, the server 1300 is shown as being configured in a cloud network 1003 connected to a communication network 1002. However, the server 1300 is not limited to this, and may be configured as a single computer or may be configured as a distributed server across multiple computers.

[0038] (Configuration Example of Camera) FIG. 6 is a block diagram showing the configuration of an example of a camera 1100 applicable to the first embodiment.

[0039] 6, the camera 1100 includes a sensor 1110, a signal processing unit 1120, an image processing unit 1121, a frame memory 1122, an output I / F 1123, and a memory card I / F 1124. The memory card I / F 1124 is, for example, a memory card slot into which a card-type storage medium using, for example, a flash memory can be inserted.

[0040] The camera 1100 further includes a CPU (Central Processing Unit) 1140, a ROM (Read Only Memory) 1141, a RAM (Random Access Memory) 1142, a storage device 1143, a communication I / F 1144, an AI processing unit 1130, and a UI unit 1150.

[0041] The storage device 1143 is a non-volatile storage medium such as a flash memory. The storage device 1143 may be the same as the ROM 1141. The CPU 1140 controls the overall operation of the camera 1100 in accordance with the programs stored in the ROM 1141 and the storage device 1143, using the RAM 1142 as a work memory.

[0042] The UI unit 1150 includes an operator and a display device, and provides a user interface (UI) to the user. The communication I / F 1144 controls communication with the communication network 1002 via wireless or wired communication in accordance with instructions from the CPU 1140.

[0043] The AI ​​processing unit 1130 includes a memory and a calculation unit, and executes processing using the face recognition machine learning model 20 in accordance with instructions from the CPU 1140. A DSP (Digital Signal Processor) may be applied to the AI ​​processing unit 1130. However, the functions of the AI ​​processing unit 1130 may also be realized by the CPU 1140 in accordance with a program.

[0044] The sensor 1110 includes a pixel array in which, for example, photodiodes serving as light receiving elements are arranged in an array, and a drive circuit for driving the pixel array. The sensor 1110 outputs pixel signals corresponding to light received by the pixel array via an optical system 1111 from an object. The sensor 1110 is controlled in operations such as exposure and reading of pixel signals from the light receiving elements, for example, in accordance with instructions from a signal processing unit 1120 (described later). The optical system 1111 also includes a focus adjustment mechanism and an iris adjustment mechanism.

[0045] The signal processing unit 1120 may control the timing of exposure by the sensor 1110, reading of pixel signals, etc., in accordance with instructions from the CPU 1140. The signal processing unit 1120 performs predetermined signal processing, such as noise removal processing and gain adjustment processing, on the analog pixel signals output from the sensor 1110, converts the processed analog pixel signals into digital pixel data, and outputs the digital pixel data.

[0046] Furthermore, the signal processing unit 1120 generates focus information and iris information based on, for example, pixel data, and controls the focus adjustment mechanism and iris adjustment mechanism in the optical system 1111 to achieve autofocus adjustment and autoiris adjustment. The signal processing unit 1120 may control the focus adjustment mechanism and iris adjustment mechanism in the optical system 1111 in accordance with the output of an AI processing unit 1130, which will be described later.

[0047] The image processing unit 1121 writes the pixel data output from the signal processing unit 1120 to the frame memory 1122. The image processing unit 1121 performs predetermined image processing, such as development processing, on the image data written to the frame memory 1122 to generate image data in units of frames. The image processing unit 1121 writes the image data, for example, one frame's worth of image data, after image processing, back to the frame memory 1122. The camera 1100 may output the image data written to the frame memory 1122 to the outside from the output I / F 1123 under control of the CPU 1140. Alternatively, the image data may be supplied to the memory card I / F 1124 and stored in an attached storage medium.

[0048] 7 is a block diagram showing an example of the configuration of a server 1300 applicable to the first embodiment. For the sake of explanation, it is assumed here that the server 1300 is configured as a stand-alone piece of hardware.

[0049] In FIG. 7, the server 1300 includes a CPU 1301, a ROM 1302, a RAM 1303, a storage device 1304, a learning processing unit 1305, and a communication I / F 1306, which are communicably connected to each other via a bus 1310.

[0050] The storage device 1304 is a non-volatile storage medium such as a flash memory or a hard disk drive. The CPU 1301 operates in accordance with programs stored in the storage device 1304 and the ROM 1302, using the RAM 1303 as a work memory, and controls the overall operation of the server 1300. The communication I / F 1306 controls communication via the communication network 1002.

[0051] The server 1300 may further be connected to an input device that accepts user input and a display device that displays information.

[0052] In the server 1300, the learning processing unit 1305 functions as a learning unit that trains an AI model. For example, the learning processing unit 1305 includes a parameter estimation machine learning model 40, and also acquires a face recognition machine learning model 20 from the camera 1100. The learning processing unit 1305 trains the parameter estimation machine learning model 40 based on input data 11, which is a photographic image that may be erroneously recognized. The learning processing unit 1305 replaces the parameters of the face recognition machine learning model 20 acquired from the camera 1100 with parameters 21 estimated by the parameter estimation machine learning model 40.

[0053] The learning processing unit 1305 may be realized by controlling a DSP with the CPU 1301, or by running a predetermined program on the CPU 1301. The server 1300 may further include another CPU for configuring the learning processing unit 1305.

[0054] A program for realizing the learning processing unit 1305 in the server 1300 can be acquired from an external source via the communication network 1002, for example, by communication via the communication I / F 1006, and installed in the server 1300. However, the program may be provided by being stored in a removable storage medium such as a CD (Compact Disk), a DVD (Digital Versatile Disk), or a USB (Universal Serial Bus) memory, or may be provided from an external device via a data I / F (not shown).

[0055] (2-2-2. Processing According to the First Embodiment) Next, processing according to the first embodiment will be described. FIG. 8 is a flowchart illustrating an example of a procedure for training the parameter estimation machine learning model 40 according to the first embodiment. Each process in the flowchart shown in FIG. 8 may be executed by an operator (hereinafter simply referred to as "operator") who operates the parameter estimation machine learning model 40, or some or all of the processes may be executed automatically in the server 1300, for example.

[0056] First, for example, the operator trains the original face recognition machine learning model 20 in advance (step S100).

[0057] Next, for example, an operator selects which parameters in the face recognition machine learning model 20 are to be estimated by the parameter estimation machine learning model 40 (step S101). A method called Low-Rank Adaptation (LoRA), which is one of the efficient additional learning methods, can be applied to the parameter selection in step S101. Applying LoRA can reduce the scale of the parameters to be estimated. A method called LoRA (2) proposes preparing additional parameters in parallel with the original parameters and learning only those parameters. A possible method is to add parameters using LoRA and estimate only the additional parameter portions.

[0058] 9A and 9B, a parameter selection method using LoRA according to the first embodiment will be described. Note that in Fig. 9A and 9B, only parameters (weight matrices) are shown, and activation functions are omitted.

[0059] 9A is a schematic diagram illustrating the configuration of a face recognition machine learning model 20a before parameter replacement, applicable to the first embodiment. As shown in FIG. 9A , the face recognition machine learning model 20a is configured such that, where N is an integer greater than or equal to 2, the output of the (N−1)th layer is input to the Nth layer, and the output of the Nth layer is input to the next (N+1)th layer. In this configuration, consider the case where the parameters of the Nth layer (referred to as original parameters 210) are parameters to be estimated. The original parameters 210 correspond to the parameters 21 described above.

[0060] 9B is a schematic diagram illustrating the configuration of a machine learning model for face recognition 20b after parameter replacement, which is applicable to the first embodiment. As shown in FIG. 9B , the learning processing unit 1305 trains the machine learning model for parameter estimation 40 based on input data 11, which is a photographic image that is likely to be misrecognized, and estimates additional parameters 220. In the Nth layer of the machine learning model for face recognition 20b, the learning processing unit 1305 leaves the original parameters 210 as they are and adds the additional parameters 220 estimated by the machine learning model for parameter estimation 40 so as to be processed in parallel with the original parameters 210.

[0061] In the face recognition machine learning model 20b, the output of the (N-1)th layer is processed in parallel in the Nth layer by the original parameters 210 and the additional parameters 220, and each is input to the (N+1)th layer. In this case, if the output from the (N-1)th layer is "X," the processing by the original parameters 210 is "α," and the processing by the additional parameters 220 is "β," the output from the Nth layer to the (N+1)th layer can be expressed as αX+βX. The original parameters 210 and the additional parameters 220 replace the original parameters 210 of the face recognition machine learning model 20.

[0062] The operator may choose the best performing method from among the various selection methods available.

[0063] Returning to FIG. 8 , the administrator prepares a training dataset for training the parameter estimation machine learning model 40 (step S102). More specifically, the administrator collects image data of human faces and image data of non-human faces to which class labels (dog, cat, etc.) have been assigned. The image data of human faces may be the same image data used in step S100 to train the face recognition machine learning model 20. The administrator uses the collected image data of human faces and image data of non-human faces as a training dataset for training the parameter estimation machine learning model 40.

[0064] Next, the operator causes the learning processing unit 1305 to train the parameter estimation machine learning model 40 using the training data set prepared in step S102 (step S103).

[0065] More specifically, the learning processing unit 1305 replaces the parameters 21 of the face recognition machine learning model 20 with the output when image data An other than a human face in the learning data set is input. Furthermore, the learning processing unit 1305 trains the parameter estimation machine learning model 40 based on the results when image data Bn different from the image data An in the learning data set is input to the face recognition machine learning model 20 for which the parameters 21 have been replaced.

[0066] The learning processing unit 1305 uses the parameters output by the learned parameter estimation machine learning model 40 to update the parameters 21 of the face recognition machine learning model 20. At this time, the learning processing unit 1305 does not update any parameters (original parameters 210) of the parameters 21 of the face recognition machine learning model 20 other than the parameters (additional parameters 220) selected in step S101.

[0067] 10 is a flowchart illustrating an example of the parameter update process according to the first embodiment. The flowchart in Fig. 10 shows the process of step S103 in the flowchart in Fig. 8.

[0068] 10 , the learning processing unit 1305 executes a first process (step S110), a second process (step S111), and a third process (step S112) as learning processes for the parameter estimation machine learning model 40. The learning processing unit 1305 integrates the results of the first process, the second process, and the third process to update the parameters 21 of the face recognition machine learning model 20 (step S113).

[0069] 11A, 11B, and 11C are schematic diagrams for explaining the parameter update process according to the first embodiment. The first process (step S110), the second process (step S111), and the third process (step S112) will be explained using FIG. 11A, 11B, and 11C.

[0070] 11A , in the first process (step S110), for example, the learning processing unit 1305 inputs learning data A1 based on image data other than a human face to the trained parameter estimation machine learning model 40. The learning processing unit 1305 updates the parameters 21 of the face recognition machine learning model 20 with parameters estimated by the trained parameter estimation machine learning model 40 based on the input learning data A1. At this time, the learning processing unit 1305 updates only the parameters selected in the process of step S101 in FIG. 8 out of the parameters 21 of the face recognition machine learning model 20, and does not update other parameters.

[0071] The learning processing unit 1305 inputs learning data B1 based on image data of human faces to the face recognition machine learning model 20 with updated parameters 21 and the original face recognition machine learning model 20 with unupdated parameters 21. The original face recognition machine learning model 20 outputs output data 31 indicating the probability that the model is a face. Similarly, the face recognition machine learning model 20 with updated parameters 21 also outputs output data 32 indicating the probability that the model is a face.

[0072] The learning processing unit 1305 trains the trained machine learning model for parameter estimation 40 so that the probability indicated by the output data 31 approaches the probability indicated by the output data 32 .

[0073] 11B , in the second process (step S111), similar to the example of FIG. 11A , for example, the learning processing unit 1305 inputs learning data A1 based on image data other than a human face to the trained parameter estimation machine learning model 40. The learning processing unit 1305 updates the parameters 21 of the face recognition machine learning model 20 with parameters estimated by the trained parameter estimation machine learning model 40 based on the input learning data A1. At this time, the learning processing unit 1305 updates only the parameters selected in the process of step S101 in FIG. 8 out of the parameters 21 of the face recognition machine learning model 20, and does not update other parameters.

[0074] The learning processing unit 1305 inputs learning data B2, which is image data other than a human face and is of a different class from the image data of learning data A1, to the face recognition machine learning model 20 with updated parameters 21 and the original face recognition machine learning model 20 with unupdated parameters 21. The original face recognition machine learning model 20 outputs output data 33 indicating the probability that the image is a face. Similarly, the face recognition machine learning model 20 with updated parameters 21 also outputs output data 34 indicating the probability that the image is a face.

[0075] The learning processing unit 1305 trains the trained machine learning model for parameter estimation 40 so that the probability indicated by the output data 33 approaches the probability indicated by the output data 34 .

[0076] 11C , in the third process (step S112), similar to the examples in FIGS. 11A and 11B , for example, the learning processing unit 1305 inputs training data A1 based on image data other than a human face to the trained parameter estimation machine learning model 40. The learning processing unit 1305 updates the parameters 21 of the face recognition machine learning model 20 with parameters estimated by the trained parameter estimation machine learning model 40 based on the input training data A1. At this time, the learning processing unit 1305 updates only the parameters selected in the process of step S101 in FIG. 8 out of the parameters 21 of the face recognition machine learning model 20, and does not update other parameters.

[0077] The learning processing unit 1305 inputs learning data B3, which is image data other than a human face and is of the same class as the image data of learning data A1, to the face recognition machine learning model 20 with updated parameters 21. The face recognition machine learning model 20 with updated parameters 21 outputs output data 35 indicating the probability that the data is a face.

[0078] The learning processing unit 1305 trains the trained machine learning model for parameter estimation 40 so that the probability value indicated by the output data 35 approaches 0 as much as possible.

[0079] The learning processing unit 1305 uses the function shown in the following equation (1) as an objective function and updates only the parameters of the learned machine learning model for parameter estimation 40, which has been learned by each of the above-mentioned first, second, and third processes, so as to minimize the objective function.

[0080]

[0081] In equation (1), the first term of the addition corresponds to the first process, the second term corresponds to the second process, and the third term corresponds to the third process. In equation (1), the function F represents the original face recognition machine learning model 20, and the subscripts 1, 2, and 3 represent the first process, the second process, and the third process, respectively. Furthermore, the value θ represents the parameters of the parameter estimation machine learning model 40, and the function f θ represents a function whose output is the machine learning model for face recognition 20 with updated parameters 21.

[0082] The variables λ1, λ2, and λ3 are parameters for controlling the balance between the first process, the second process, and the third process. The variables λ1, λ2, and λ3 may be adjusted by, for example, a user while observing performance.

[0083] The learning processing unit 1305 replaces the parameters 21 of the face recognition machine learning model 20 with the parameters of the updated parameter estimation machine learning model 40 obtained by minimizing using equation (1). The server 1300 deploys the face recognition machine learning model 20 with the replaced parameters 21 to, for example, the camera 1100.

[0084] In the camera 1100, the AI ​​processing unit 1130 may perform face recognition processing using the deployed face recognition machine learning model 20 on image data captured by the sensor 1110. In this case, the AI ​​processing unit 1130 can focus on a human face while avoiding focusing on an object that may be erroneously recognized as a human face, for example, by controlling the optical system 1111 via the signal processing unit 1120, for example.

[0085] As described above, in the first embodiment of the present disclosure, when modifying an AI model after deployment, the AI ​​model can be modified simply by replacing the parameters of the AI ​​model without relearning the AI ​​model, which makes it possible to modify the AI ​​model after deployment more quickly and easily.

[0086] For example, by applying the first embodiment of the present disclosure to a camera that performs focus adjustment using AI, even if the AI ​​erroneously recognizes an unintended object as the target for focus adjustment, it is possible to correct the focus adjustment error caused by the erroneous recognition simply by replacing the parameters of the AI ​​model, without having to re-learn the AI ​​model.

[0087] (2-3. Modifications of the First Embodiment) Next, modifications of the first embodiment will be described.

[0088] In the above description, an example has been described in which the camera 1100 has a function of automatically recognizing and focusing on a human face. However, the configuration according to the first embodiment can also be applied to a camera that has a function of automatically focusing on a specified person by replacing the "probability that it is a face" in the above-described implementation method with the "probability that the face of the specified person is captured in the image."

[0089] In the above description, the object to be focused on is a "human face," but this is not limited to this example. For example, in the above-described implementation method, by replacing the "human face" with another object, the configuration according to the first embodiment can be applied to a model that recognizes the presence of an object other than a human face.

[0090] The above describes a method for "making it difficult to recognize all images of a specified class as faces." However, the first embodiment is not limited to this, and can also be used to "make it difficult to recognize only specified images as faces" by regarding each image as having a different class label assigned to it.

[0091] In the above, a method for adding additional parameters 220 in parallel to original parameters 210 by LoRA has been described as a method for selecting which parameters of face recognition machine learning model 20 to estimate using parameter estimation machine learning model 40 (see FIG. 9B ). This is not limited to this example, and additional parameters 220 that are added in series to original parameters 210 can also be estimated by LoRA.

[0092] 12 is a schematic diagram illustrating an example of adding additional parameters 220 in series to the original parameters 210 of the face recognition machine learning model 20c, as a modification of the first embodiment. The parameter estimation machine learning model 40 is trained based on input data 11, which is a photographic image that may be misrecognized, and estimates the additional parameters 220. The learning processing unit 1305 estimates the additional parameters 220 from the Nth layer 200c of the face recognition machine learning model 20c. N In this step, the original parameters 210 are left as they are, and the additional parameters 220 estimated by the parameter estimation machine learning model 40 are added to the original parameters 210 so as to be processed in series.

[0093] In the face recognition machine learning model 20b, the output of the (N-1)th layer is processed by the original parameters 210 in the Nth layer, and the processing results of the original parameters 210 are further processed by the additional parameters 220. The output by the additional parameters 220 in the Nth layer is input to the (N+1)th layer. In this case, if the output from the (N-1)th layer is "X," the processing by the original parameters 210 is "α," and the processing by the additional parameters 220 is "β," the output from the Nth layer to the (N+1)th layer can be expressed as βαX. The original parameters 210 and the additional parameters 220 replace the original parameters 210 of the face recognition machine learning model 20.

[0094] In the above, as a method for invalidating specific learning data, f θ (x A1 )(x B3 ) to make the target data unrecognizable. The method for invalidating specific training data is not limited to this example.

[0095] For example, a model that outputs a probability that is as close to 0 as possible for each class may be created, and the parameters of the model may be used as the correct label for the parameter estimation machine learning model 40 to train the parameter estimation machine learning model 40.

[0096] Furthermore, for example, the influence can be calculated using a technique for calculating the influence of each piece of data on a parameter (e.g., an influence function), and the parameter can be modified so that the influence of a specified class becomes 0, thereby obtaining a correct label for the parameter estimation machine learning model 40. The correct label may be used to train the parameter estimation machine learning model 40.

[0097] (2-4. Other Application Examples of the First Embodiment) Next, other application examples of the first embodiment will be described. Although an example in which the technology according to the present disclosure is applied to a camera has been described above as the first embodiment, the technology according to the present disclosure is not limited to application to cameras.

[0098] Techniques for generating images using generative models with AI are known. The techniques disclosed herein can also be used to prevent undesirable images from being generated by the generative models.

[0099] A technique for classifying spam emails from a large number of received emails using an AI spam email classification model is known. The technique according to the present disclosure can also be used to correct misclassifications in the spam email classification model.

[0100] A technology is known for making recommendations to users on electronic commerce (EC) sites based on the user's interests. The technology disclosed herein can also be used to correct erroneous recommendations on EC sites (for example, cases in which information that is completely uninteresting to the user is recommended).

[0101] Pet robots whose behavior is controlled by AI are known. Some pet robots can be trained by training an AI model to perform actions in response to a user's actions, making it possible to train the pet robot. The technology disclosed herein can be used to train a pet robot to prevent undesirable behavior. For example, a pet robot may engage in undesirable behavior, such as taking photos at undesirable times using an automatic photo-taking function or going in an undesirable direction. By applying the technology disclosed herein to the training of the pet robot's AI model, it is possible to suppress undesirable behavior by the pet robot.

[0102] (3. Second Embodiment of the Present Disclosure) Next, a second embodiment of the present disclosure will be described. The second embodiment relates to opting out of the use of learning data.

[0103] For example, in order to train a generative AI model in an image generation AI that generates images based on text data, etc., a large amount of image data must be used. For example, the developer of the generative AI model can collect image data to be used as training data from image data publicly available on the Internet, for example. On the other hand, a rights holder who holds the rights to the image data, such as the creator of the image data collected as training data, may not consent to the use of the image data to which they hold the rights as training data. In this case, the rights holder of the image data can take an action to indicate their intention not to consent to the developer of the generative AI model that uses the image data as training data. This action of indicating their intention not to consent is called opting out.

[0104] (3-1. Background of the Second Embodiment) The background of the second embodiment will be described.

[0105] The European AI Bill (EU Bill 2023), passed by the European Parliament in May 2023, also regulates foundational models and generative AI. Article 28b4 of the EU AI Bill imposes the following obligations on providers of foundational models used in AI systems specifically intended to generate complex text, images, audio, video, or other content with varying levels of autonomy (generative AI), as well as providers of specializing foundational models for generative AI systems:

[0106] The obligation is to "document and make publicly available, without prejudice to national or Union law on copyright, a sufficiently detailed summary of the use of copyright-protected training data" (Article 28b4(c) of the EU AI Bill). Providers of foundation models and generative AI must comply with this obligation by making public a sufficiently detailed summary of the use of training data.

[0107] In Europe, there is the Digital Single Market Copyright Directive (hereinafter referred to as the Copyright Directive), which states the following about the use of copyrighted works for AI learning (Digital Single Market Copyright Directive):

[0108] Article 3 of the Copyright Directive does not allow research institutions to opt out of AI training data for scientific research. On the other hand, Article 4 of the Copyright Directive explicitly allows rights holders to opt out of the commercial use of AI. Providers of foundational models and generative AI for commercial use are required to accept opt-out requests under the Copyright Directive, in accordance with the obligation to disclose copyright information under the EU AI Bill (Article 28b, paragraph 4(c) of the EU AI Bill).

[0109] Furthermore, such disclosure may also allow for opt-out based on personal information. The "right to be forgotten" under the General Data Protection Regulation (GDPR) is defined as the erasure of data (Article 17 of the GDPR). Data subjects have the right to delete their personal data under certain conditions. In other words, the disclosure of learning data suggests the possibility of opting out under the Copyright Directive and the GDPR.

[0110] (3-2. Overview of the Second Embodiment) When a provider of a base model and a generative AI receives an opt-out request for AI training data, the provider must delete the corresponding AI training data and rebuild the base model or the generative AI model. Re-training the base model incurs significant computational costs. Depending on the size of the base model, this may take several months. In other words, rebuilding the base model or the generative AI model requires financial and time investments as economic costs. There is a need for the development of a method for quickly responding to opt-out requests while minimizing computational costs.

[0111] Efficient opt-out requires a technique for technically changing a model without re-learning. In recent years, the aforementioned machine unlearning technique has been attracting attention as a method for making a model forget specific data (invalidate specific data). For more information on machine unlearning, see Non-Patent Document 2, "Thanh Tam Nguyen and five others, 'A Survey of Machine Unlearning,' [online], September 6, 2022, arXiv:2209.02299, [searched October 26, 2023], Internet, <https: / / arxiv.org / abs / 2209.02299>."

[0112] Because the model stores data that corresponds to opt-out, deleting such data is not easy in the context of AI. Non-Patent Document 2 aims to comprehensively examine the concepts, scenarios, methods, and application examples of machine unlearning. Specifically, Non-Patent Document 2 introduces cutting-edge research on machine unlearning, its formulation, design criteria, deletion requirements, algorithms, and applications.

[0113] Fig. 13 is a flowchart showing an example of a process for opting out according to existing technology. In existing technology, as shown in Fig. 13, when a right holder of certain learning data requests an opt-out, the process is to consolidate the requests to some extent and then perform re-learning.

[0114] For example, a rights holder who wishes to submit an opt-out application creates application data by entering necessary information, such as the name of the data for which the opt-out is being requested, into a specified form, and then transmits the created application data to an information processing system related to training of a generative AI model. Upon receiving the opt-out application (step S200), the information processing system determines in step S201 whether or not a predetermined number of pieces of application data have been collected. If the information processing system determines that no more than a predetermined number of pieces of application data have been collected (step S201, "No"), the process returns to step S200. On the other hand, if the information processing system determines that no more than a predetermined number of pieces of application data have been collected (step S201, "Yes"), the process proceeds to step S202, where the collected application data are compiled.

[0115] In the next step S203, the information processing system deletes the opt-out target data requested by each application data from the training dataset used to train the generative AI model, and reconstructs the training dataset. In the next step S204, the information processing system re-trains the generative AI model using the reconstructed training dataset, and in the next step S205, deploys the re-trained generative AI model to the target device.

[0116] FIG. 14 is a flowchart illustrating an example of a process for opting out according to the second embodiment.

[0117] As described above, a rights holder who wishes to make an opt-out request creates application data according to a predetermined form and sends it to an information processing system related to training of a generative AI model. Upon receiving the opt-out request (step S210), the information processing system performs unlearning processing on the target data for which the opt-out request was made in the next step S211.

[0118] This unlearning process invalidates learning using the target data in the generated AI model. In the next step S212, the information processing system deploys the unlearned generated AI model to the target device.

[0119] As an example, a rights holder who wishes to submit an opt-out application creates application data using a UI provided by the information processing system, and operates, for example, a submit button provided on the UI to submit the created application data to the information processing system. In response to the operation of the submit button, the information processing system automatically starts unlearning processing for the target data in accordance with the submitted application data, and when the unlearning processing is complete, deploys the unlearned generated AI model to a device.

[0120] In this way, in the second embodiment, in response to the submission of application data for an opt-out application, it is possible to consistently perform the unlearning process of the requested target data and the deployment of the unlearned generated AI model to a device.

[0121] In contrast, the opt-out process according to the existing technology requires manual intervention, such as deleting data and relearning, which requires a great deal of time and computational resources. The second embodiment is configured as a real-time opt-out device that can solve this problem.

[0122] By incorporating the learning by machine unlearning according to the second embodiment into the system of an AI provider that provides generative AI, re-learning in response to an opt-out request is no longer necessary, and computational costs can be significantly reduced. As a result, the process from opt-out request to deployment of the generative AI model can be carried out cost- and time-efficiently. Applying the second embodiment has significant economic benefits, particularly in large-scale models such as foundation models.

[0123] Specific services to which the second embodiment can be applied include large-scale language models provided online and models that generate images from text. Applying the second embodiment to these models eliminates the need for costly re-learning and enables automatic online model updates. Furthermore, as described above, applying the second embodiment to a system for training an AI model enables a rights holder to cause the AI ​​model to forget (invalidate) the data that is the subject of the opt-out request by simply clicking once on the UI of the request form, for example, in the opt-out request.

[0124] (3-3. More Specific Description of the Second Embodiment) Next, the second embodiment will be described in more detail.

[0125] (3-3-1. Configuration Applicable to Second Embodiment) A configuration applicable to the second embodiment will be described below. Fig. 15 is a schematic diagram showing the configuration of an example of an information processing system applicable to the second embodiment.

[0126] 15, an information processing system 1004 includes a server 1600, a user terminal 1400, and a rights holder terminal 1500, which are connected via a communication network 1002. The communication network 1002 may be the Internet or a LAN constructed in a closed environment such as an in-house network.

[0127] 15, the server 1600 is shown as being configured in a cloud network 1003 connected to a communication network 1002. However, the server 1600 is not limited to this, and may be configured as a single computer or may be configured as a distributed server across multiple computers.

[0128] The server 1600 has a generative AI model and, for example, collects data (e.g., image data) via the communication network 1002 as training data to be used for training the generative AI model, and creates a training dataset. The server 1600 may provide a service to generate, for example, image data to, for example, a user terminal 1400 connected to the communication network 2, using the generative AI model trained using the training dataset. Without being limited to this, the server 1600 may deploy part or all of the generative AI model to the user terminal 1400 via the communication network 1002.

[0129] The rights holder terminal 1500 is, for example, a terminal device used by a rights holder who has rights to some of the training data included in the training data set created by the server 1600 for training the generation AI model. For example, the server 1600 may provide the rights holder terminal 1500 with a UI including an application form for submitting an opt-out application for the training data to which the rights holder has rights.

[0130] Note that a general computer configuration can be applied to user terminal 1400 and right holder terminal 1500, and therefore a description thereof will be omitted here. Also, a configuration equivalent to that of server 1300 described using FIG. 7 can be applied to server 1600, and therefore a description thereof will be omitted here.

[0131] (3-3-2. Processing According to the Second Embodiment) In the second embodiment, an example of application of the technology of the present disclosure to a diffusion generation model that generates images from text will be described. In such a diffusion generation model, as described above, it is necessary to comply with the Copyright Directive and the GDPR opt-out.

[0132] Such generative diffusion models are often very large in scale, and removing the target image from the training data and restarting training from scratch to accommodate opt-out is costly and unrealistic. As a way to accommodate opt-out without restarting training from scratch, technology has been proposed that prevents the generation of images similar to opt-out target images, but the issue of requiring training every time an opt-out request is made remains.

[0133] By applying the second embodiment to responding to opt-out for a diffusion generation model, the influence of the opt-out target data can be quickly removed from the target model without re-learning upon acceptance of an opt-out request, and the model can lose its ability to generate opt-out target data.

[0134] FIG. 16 is a schematic diagram for explaining image generation processing using a diffusion generation model according to existing technology.

[0135] In a diffusion generation model that generates an image from text data 50 (second text data, second learning data), when generating an image, information from the text data 50 is used, and noise is gradually removed from input image data 60 containing original random noise using a noise removal machine learning model 310 (second machine learning model), thereby generating a natural image as output image data 70.

[0136] Generally, text data 50 cannot be used as is as input to a machine learning model. Therefore, in the diffusive generation model, a dedicated machine learning model for text information conversion 300 is used to convert the text data 50 into a vector (a vector representing text information) in a format that is easier for the machine learning model to handle. The machine learning model for noise removal 310 uses this vector to convert the image being generated into an image with some noise removed. The diffusive generation model starts with a random noise image and repeats noise removal on the image multiple times, ultimately obtaining output image data 70 that is a natural image that matches the text data 50.

[0137] In the second embodiment, a machine learning model is prepared for estimating parameters of a diffusion generation model according to existing technology, and a method is proposed for preventing the generation of images that are subject to opt-out.

[0138] Here, we consider a case where we want to prevent the generation of an image when a person's name is used to generate an image using a diffusion generation model, resulting in the generation of a facial image of that person, or when a copyrighted work's name is used to generate an image using a diffusion generation model, resulting in the generation of an image similar to that copyrighted work.

[0139] 17 is a schematic diagram for outlining an example of application of a model according to the second embodiment. As shown in Fig. 17 , in the second embodiment, a machine learning model for parameter estimation 320 is prepared, which estimates parameters 311 of the machine learning model for noise removal 310, in addition to a machine learning model for noise removal 310 for removing noise based on vectors representing text information from input image data 60 in a diffusion generative model.

[0140] In the second embodiment, the parameter estimation machine learning model 320 (first machine learning model) receives as input text data 51 (such as a person's name or the name of a copyrighted work) representing an opt-out target image, i.e., an image that the user does not want to be generated. The opt-out target image is not limited to an image that the user does not want to be generated. For example, the opt-out target image may be an image that the user does not want to be used or recognized.

[0141] The parameter estimation machine learning model 320 estimates and outputs parameters 311 (first parameters) of the noise removal machine learning model 310 based on the input text data 51 (first text data, first learning data). At this time, the parameter estimation machine learning model 320 is trained so that the output parameters 311 are "the case where correction is made so that only the undesired images are not generated, without losing the ability to generate images unrelated to the undesired images."

[0142] The parameters (second parameters) of the noise removal machine learning model 310 are replaced using the parameters 311 output from the parameter estimation machine learning model 320. This parameter replacement allows the noise removal machine learning model 310 to quickly correct the diffusion generation model.

[0143] In the existing technology, modifying the diffusion generation model involves re-learning the noise removal machine learning model 310. In contrast, in the second embodiment, there is no need to re-learn the noise removal machine learning model 310, and the diffusion generation model can be modified simply by replacing the weights (parameters 311) of the noise removal machine learning model 310. Therefore, by applying the second embodiment, it is possible to modify the diffusion generation model more quickly and easily than in the existing technology.

[0144] FIG. 18 is a schematic diagram for explaining the learning method of the diffusion generative model.

[0145] In the diffusion generative model, input data 90, which is training data 910 that is image data and to which noise has been intentionally added, is input to a noise removal machine learning model 310. The diffusion generative model is trained to estimate the training data 910, which is image data before noise is added, using the noise removal machine learning model 310. More specifically, in the diffusion generative model, the noise removal machine learning model 310 is trained to bring the noise-removed output image data 80 closer to the training data 910.

[0146] (Details of Processing According to Second Embodiment) Next, the processing according to the second embodiment will be described in more detail. Fig. 19 is a flowchart illustrating an example of a procedure for training the parameter estimation machine learning model 320 according to the second embodiment. Each process in the flowchart illustrated in Fig. 19 may be executed by an operator, or some or all of the processes may be automatically executed by the server 1600, for example.

[0147] First, for example, the operator prepares a base diffusion generation model. In this example, the base noise removal machine learning model 310 is prepared (step S300).

[0148] Next, for example, the operator selects which parameters in the noise removal machine learning model 310 (diffusion generative model) are to be estimated by the parameter estimation machine learning model 320 (step S301). The parameter selection in step S301 can be performed using the LoRA method described with reference to Figs. 8, 9A, 9B, and 12.

[0149] Next, for example, the operator collects pairs of images and text data for explaining the images, and sets the collected data as training data (step S302). Here, the training data may be the training data used to train the original noise removal machine learning model 310 (diffusion generative model).

[0150] Next, the operator replaces the parameters 311 of the diffusion generation model (machine learning model for noise removal 310) with parameters estimated, for example, by the learning processing unit 1305 based on the output when text data 51 is input into the trained machine learning model for parameter estimation 320 (step S303).

[0151] Next, the operator, for example, uses the learning processing unit 1305 to train the machine learning model for parameter estimation 320 based on the diffusion generation model (machine learning model for noise removal 310) whose parameters have been replaced by the processing in step S303, the text data 50 input to the machine learning model for text information conversion 300, and the text data representing the opt-out target image (step S304).

[0152] The learning processing unit 1305 updates the parameters 311 of the diffusion generation model (noise removal machine learning model 310) using the parameters output by the thus-trained parameter estimation machine learning model 320. At this time, the learning processing unit 1305 does not update any parameters of the diffusion generation model parameters 311 other than the parameters selected in step S301.

[0153] Fig. 20 is a flowchart illustrating an example of a parameter update process according to the second embodiment. The flowchart in Fig. 20 shows the process of step S304 in the flowchart in Fig. 19 .

[0154] 20 , the learning processing unit 1305 executes a first process (step S310) and a second process (step S311) as learning processes for the parameter estimation machine learning model 320. The learning processing unit 1305 integrates the results of the first process and the second process to update the parameters 311 of the diffusion generative model (step S312).

[0155] 21A and 21B are schematic diagrams for explaining the parameter update process according to the second embodiment. The first process (step S310) and the second process (step S311) will be explained with reference to FIGS. 21A and 21B.

[0156] 21A shows an example of the first process in which the similarity between text data 50 (also shown as text data A3 in the figure) input to the machine learning model for text information conversion 300 and text data 51 (also shown as text data B4 in the figure) input to the machine learning model for parameter estimation 321 is higher than a predetermined threshold (step S310). The machine learning model for parameter estimation 321 estimates parameters 311 based on the input text data 51, and replaces the parameters selected in the machine learning model for noise removal 310 with the estimated parameters 311.

[0157] For example, the learning processing unit 1305 trains the machine learning model for parameter estimation 321 so that the noise-removed output image data 80 output from the machine learning model for noise removal 310 based on the input data 92 approaches image data A4 as learning data included in the input data 91.

[0158] Note that whether the text data 50 and the text data 51 are similar to each other may be determined, for example, by applying the method described in Non-Patent Document 3, "Alec Radford and 11 others, 'Learning transferable visual models from natural language supervision', [online], February 26, 2021, arXiv:2103.00020, [searched November 15, 2023], Internet, <https: / / arxiv.org / abs / 2103.00020>." The threshold for determination may be appropriately determined, for example, by an operator.

[0159] 21B shows an example of the second process in which the similarity between text data 50 (text data A3) input to the text information conversion machine learning model 300 and text data 51 (text data B4) input to the parameter estimation machine learning model 321 is less than a predetermined threshold (step S311). The parameter estimation machine learning model 321 estimates parameters 311 based on the input text data 51, and replaces the parameters selected in the noise removal machine learning model 310 with the estimated parameters 311.

[0160] For example, the learning processing unit 1305 trains the machine learning model for parameter estimation 321 so that the noise-removed output image data 80 output from the machine learning model for noise removal 310 based on the input data 92 moves away from the image data A4 as learning data included in the input data 91.

[0161] The learning processing unit 1305 uses a predetermined function according to the above-mentioned equation (1) as an objective function, and updates only the parameters of the learned machine learning model for parameter estimation 321, which has been learned by each of the above-mentioned first and second processes, so as to minimize the objective function.

[0162] As described above, in the second embodiment of the present disclosure, when invalidating specific learning data included in a learning dataset used to train a trained AI model in response to an opt-out request for the trained AI model, the learning data to be opted out can be invalidated simply by replacing parameters without re-training the AI ​​model. Therefore, it is possible to more quickly and easily execute processing in response to an opt-out request for a trained AI model.

[0163] (3-3-3. Example of System Configuration Applicable to Second Embodiment) Next, an example of a system configuration applicable to the second embodiment will be roughly described.

[0164] 15 distributes application software including a generative model, and may receive an opt-out request from the rights holder of training data (e.g., image data) included in a training dataset on which the generative model was trained. In this case, a conceivable configuration is that the server 1600 uses the parameter estimation machine learning model 320 to create modified parameters 311, and the server 1600 distributes the created parameters 311 as a modified program to the end user's user terminal 1400, etc.

[0165] Possible methods for replacing the parameters of the diffusion generation model (noise removal machine learning model 310) on the user terminal 1400 with the modified parameters 311 include performing the process on the server 1600 and performing the process on the user terminal 1400.

[0166] (Second System Configuration Example) In some cases, a generative model is provided in a format in which an environment in which the generative model operates is prepared on the server 1600 and access rights to the generative model are granted to users. In this case, the server 1600 may be configured to, upon receiving an opt-out request from the rights holder of the training data (e.g., image data) included in the training dataset on which the generative model was trained, create revised parameters 311 on the server 1600 using the parameter estimation machine learning model 320, and perform parameter replacement processing on the server 1600.

[0167] (Third System Configuration Example) There is a case where the parameters (weights) of a generative model are placed on a website different from the server 1600, and a user downloads the parameters to the user terminal 1400 to operate the generative model. In this case, the server 1600 may be configured to publish modified parameters on the website upon receiving an opt-out request from the rights holder of the training data (e.g., image data) included in the training dataset on which the generative model was trained. In this case, it is also considered that the parameter replacement process will be performed by each user.

[0168] (3-4. Modifications of the Second Embodiment) Next, modifications of the second embodiment will be described.

[0169] The process in step S204 in FIG. 19 can be replaced with "any learning method that adjusts the parameters of the original generative model so as not to generate a specific image."

[0170] In the above description, the parameters of the noise removal machine learning model 310 are replaced by the parameters 311 estimated by the parameter estimation machine learning model 320, but this is not limited to this example. For example, the parameters of the text information conversion machine learning model 300 may be replaced by the parameters estimated by the parameter estimation machine learning model 320.

[0171] The effects described in this specification are merely examples and are not limiting, and other effects may also be present.

[0172] Note that the present technology may also be configured as follows. (1) An information processing device including a machine learning model incorporating parameters for performing inference, wherein another parameter estimated by another machine learning model based on learning data to be invalidated for a learning dataset including learning data used to train the machine learning model is incorporated into the machine learning model. (2) The information processing device described in (1), wherein the parameter is replaced with the other parameter. (3) The information processing device described in (1), wherein the other parameter is added in parallel to the parameter in the machine learning model. (4) The information processing device described in (1), wherein the other parameter is added in series to the parameter in the machine learning model. (5) The information processing device described in any of (1) to (4), wherein the machine learning model is trained in advance to recognize a predetermined object included in an image using the training dataset, and wherein the other parameter learned by the other machine learning model different from the machine learning model is incorporated into the machine learning model using images that the machine learning model misrecognizes with a predetermined probability or higher as training data. (6) The information processing device according to (5), wherein the other parameters estimated by the other machine learning model trained using an image not including the predetermined object as the training data to be invalidated are incorporated into the machine learning model. (7) The information processing device according to (6), wherein the other parameters estimated by the other machine learning model trained so that a result of inference on the predetermined object by the machine learning model not incorporating the other parameters approaches a result of inference on the predetermined object by the machine learning model incorporating the other parameters.(8) The information processing device according to any one of (6) to (7), wherein the information processing device incorporates into the machine learning model the other parameters estimated by the other machine learning model trained to make the probability that a result of the machine learning model into which the other parameters are not incorporated, where the machine learning model has performed inference on an image that does not contain the predetermined object and is of a different class from the image that does not contain the predetermined object used as the learning data to be invalidated, approach a result of the machine learning model into which the other parameters are incorporated, where the result of the machine learning model has performed inference on an image that is of a different class from the image that does not contain the predetermined object used as the learning data to be invalidated, approach zero. (9) The information processing device according to any one of (6) to (8), wherein the information processing device incorporates into the machine learning model the other parameters estimated by the other machine learning model trained to make the probability that a different image of the same class, where the different image contains the predetermined object, approach zero, where the result of the machine learning model into which the other parameters are incorporated, where the result of the machine learning model into which the other parameters are incorporated, where the result of the machine learning model has performed inference on an image that does not contain the predetermined object and is of a different class from the image that does not contain the predetermined object used as the learning data to be invalidated, approach zero. (10) An information processing method comprising: a processor incorporating into a machine learning model another parameter different from a parameter to be incorporated into the machine learning model, the other parameter being estimated by another machine learning model based on learning data to be invalidated for a learning dataset including learning data used to train the machine learning model. (11) An information processing device comprising: a learning unit that trains a first machine learning model that estimates a first parameter in the machine learning model based on first learning data to be invalidated for the learning dataset, wherein the learning dataset includes second learning data used for learning a second machine learning model to perform inference, and the learning unit incorporates the first parameter into the second machine learning model. (12) The information processing device according to (11), wherein the learning unit replaces a second parameter in the second machine learning model with the first parameter.(13) The information processing device according to (11), wherein the learning unit adds the first parameter in parallel to the second parameter in the second machine learning model. (14) The information processing device according to (11), wherein the learning unit adds the first parameter in series to the second parameter in the second machine learning model. (15) The information processing device according to any of (11) to (14), wherein the second machine learning model is trained in advance to recognize a predetermined object included in an image using the training dataset, and the learning unit trains the first machine learning model using, as the first training data, images that the second machine learning model misrecognizes with a predetermined probability or higher. (16) The information processing device according to (15), wherein the learning unit trains the first machine learning model using, as the first training data, images that do not include the predetermined object. (17) The information processing device according to (16), wherein the learning unit trains the first machine learning model so that a result of inference made by the second machine learning model not incorporating the first parameter on the predetermined object approaches a result of inference made by the second machine learning model incorporating the first parameter on the predetermined object. (18) The information processing device according to (16) or (17), wherein the learning unit trains the first machine learning model so that a result of inference made by the second machine learning model not incorporating the first parameter on an image that does not include the predetermined object and is of a different class from images that do not include the predetermined object used as the first learning data approaches a result of inference made by the second machine learning model incorporating the first parameter on the image of the different class.(19) The information processing device according to any of (16) to (18), wherein the learning unit trains the first machine learning model into which the first parameters are incorporated so as to bring a probability that a different image of the same class as an image used as the first training data contains the predetermined object, which image is an image that does not contain the predetermined object and is obtained as a result of inference performed on the different image of the same class as an image that does not contain the predetermined object, closer to 0. (20) The information processing device according to (11), wherein the second machine learning model is trained in advance to generate a generated image corresponding to second text data based on an input image using the training dataset including the second training data, and the learning unit trains the first machine learning model using the first training data that is first text data corresponding to second training data to be invalidated or deleted from the second training data included in the training dataset. (21) The information processing device according to (20), wherein the learning unit trains the first machine learning model so that a similarity of the generated image to the input image decreases when a similarity between the first text data and the second text data is equal to or greater than a threshold. (22) The information processing device according to (20) or (21), wherein the learning unit trains the first machine learning model so that a similarity between the generated image and the input image increases when a similarity between the first text data and the second text data is less than a threshold. (23) An information processing method, wherein a processor includes a learning step of training a first machine learning model that estimates a first parameter in a machine learning model based on first training data to be invalidated for a training dataset, wherein the training dataset includes second training data used in training a second machine learning model to perform inference, and the learning step incorporates the first parameter into the second machine learning model.(24) An information processing device comprising: a learning unit that trains a first machine learning model that estimates a first parameter in a machine learning model based on first learning data, which is first text data corresponding to an opt-out target for a learning dataset; wherein the learning dataset includes second learning data corresponding to the second text data used in training a second machine learning model to generate a generated image in accordance with second text data; and the learning unit trains the first machine learning model using the first learning data corresponding to second learning data to be invalidated or deleted from the second learning data included in the learning dataset; and incorporates the first parameter into the second machine learning model. (25) An information processing method, comprising: a learning step in which a processor trains a first machine learning model that estimates a first parameter in a machine learning model based on first training data, which is first text data corresponding to an opt-out target for a training dataset; wherein the training dataset includes second training data corresponding to the second text data used in training a second machine learning model to generate a generated image in accordance with second text data; and the learning step trains the first machine learning model using the first training data corresponding to second training data to be invalidated or deleted from the second training data included in the training dataset; and incorporates the first parameter into the second machine learning model.

[0173] 10, 11, 90 Input data 20, 20c Machine learning model for face recognition 21, 311 Parameters 30, 31, 32, 33, 34, 35 Output data 40, 320, 321 Machine learning model for parameter estimation 50, 51 Text data 60 Input image data 70, 80 Output image data 210 Original parameters 220 Additional parameters 300 Machine learning model for text information conversion 310 Machine learning model for noise removal 910 Learning data 1000, 1004 Information processing system 1130 AI processing unit 1300, 1600 Server 1305 Learning processing unit 1400 User terminal 1500 Right holder terminal

Claims

1. An information processing device comprising a machine learning model incorporating parameters for performing inference, and incorporating into the machine learning model other parameters estimated by another machine learning model based on training data to be invalidated for a training dataset including training data used to train the machine learning model.

2. The information processing device according to claim 1, wherein the parameter is replaced with the other parameter.

3. The information processing device according to claim 1, wherein the other parameters are added in parallel to the parameter in the machine learning model.

4. The information processing device according to claim 1, wherein the other parameters are added in series to the parameter in the machine learning model.

5. The information processing device of claim 1, wherein the machine learning model is pre-trained to recognize a predetermined object contained in an image using the training dataset, and images that the machine learning model misrecognizes with a probability higher than a predetermined value are used as training data, and the other parameters learned by another machine learning model different from the machine learning model are incorporated into the machine learning model.

6. The information processing device according to claim 5, wherein the other parameters estimated by the other machine learning model trained using an image that does not include the specified object as training data to be invalidated are incorporated into the machine learning model.

7. The information processing device of claim 6, wherein the other parameters estimated by the other machine learning model that has been trained so that the result of inference made on the specified object by the machine learning model without the other parameters incorporated therein approaches the result of inference made on the specified object by the machine learning model with the other parameters incorporated therein.

8. The information processing device of claim 6, wherein the other parameters estimated by the other machine learning model that has been trained so that the result of inference performed by the machine learning model not incorporating the other parameters on an image that does not contain the specified object and is of a different class from the image that does not contain the specified object used as the learning data to be invalidated approaches the result of inference performed by the machine learning model incorporating the other parameters on an image of a different class.

9. The information processing device of claim 6, wherein the machine learning model into which the other parameters have been incorporated incorporates the other parameters estimated by the other machine learning model, which has been trained to bring the probability that a different image of the same class containing the specified object, which is obtained as a result of performing inference on an image that does not contain the specified object and is used as training data to be invalidated, closer to zero.

10. An information processing method comprising: a step in which a processor incorporates into the machine learning model other parameters different from the parameters incorporated into the machine learning model, the other parameters being estimated by another machine learning model based on training data to be invalidated for a training dataset including training data used to train the machine learning model.

11. An information processing device comprising: a learning unit that trains a first machine learning model that estimates a first parameter in a machine learning model based on first learning data that is to be invalidated from a learning dataset, wherein the learning dataset includes second learning data used for learning to enable a second machine learning model to make inference, and the learning unit incorporates the first parameter into the second machine learning model.

12. The information processing device according to claim 11, wherein the learning unit replaces second parameters in the second machine learning model with the first parameters.

13. The information processing device according to claim 11, wherein the learning unit adds the first parameter in parallel to a second parameter in the second machine learning model.

14. The information processing device according to claim 11, wherein the learning unit adds the first parameter in series to a second parameter in the second machine learning model.

15. The information processing device described in claim 11, wherein the second machine learning model is pre-trained to recognize a specified object contained in an image using the training dataset, and the training unit trains the first machine learning model using, as the first training data, images that the second machine learning model misrecognizes with a probability greater than or equal to a specified value.

16. The information processing device according to claim 15, wherein the learning unit uses an image that does not include the specified object as the first learning data to train the first machine learning model.

17. The information processing device described in claim 16, wherein the learning unit trains the first machine learning model so that the result of inference made on the specified object by the second machine learning model without incorporating the first parameter approaches the result of inference made on the specified object by the second machine learning model incorporating the first parameter.

18. The information processing device described in claim 16, wherein the learning unit trains the first machine learning model so that a result of inference performed by the second machine learning model not incorporating the first parameters on an image that does not contain the specified object and is of a different class from images that do not contain the specified object used as the first learning data approaches a result of inference performed by the second machine learning model incorporating the first parameters on an image of the different class.

19. The information processing device described in claim 16, wherein the learning unit trains the first machine learning model so that the probability that a different image of the same class as an image that does not contain the specified object and is obtained as a result of inference performed on an image that does not contain the specified object and is used as the first learning data, contains the specified object approaches zero.

20. The information processing device described in claim 11, wherein the second machine learning model is pre-trained to generate a generated image corresponding to second text data based on an input image using the training dataset including the second training data, and the training unit trains the first machine learning model using the first training data, which is first text data corresponding to second training data to be invalidated or deleted from the second training data included in the training dataset.

21. The information processing device described in claim 20, wherein the learning unit trains the first machine learning model so that the similarity of the generated image to the input image decreases when the similarity between the first text data and the second text data is equal to or greater than a threshold.

22. The information processing device described in claim 20, wherein the learning unit trains the first machine learning model so that the similarity between the generated image and the input image increases when the similarity between the first text data and the second text data is less than a threshold value.

23. An information processing method, comprising: a learning step in which a processor trains a first machine learning model that estimates a first parameter in the machine learning model based on first training data that is to be invalidated for a training dataset, wherein the training dataset includes second training data used for training a second machine learning model to perform inference, and the learning step incorporates the first parameter into the second machine learning model.

24. An information processing device comprising: a learning unit that trains a first machine learning model that estimates a first parameter in a machine learning model based on first training data, which is first text data corresponding to an opt-out target for a training dataset; wherein the training dataset includes second training data corresponding to the second text data used in training a second machine learning model to generate a generated image in accordance with the second text data; and the learning unit trains the first machine learning model using the first training data corresponding to second training data to be invalidated or deleted from the second training data included in the training dataset; and incorporates the first parameter into the second machine learning model.

25. An information processing method, comprising: a learning step in which a processor trains a first machine learning model that estimates a first parameter in a machine learning model based on first training data, which is first text data corresponding to an opt-out target in a training dataset; wherein the training dataset includes second training data corresponding to the second text data used in training a second machine learning model to generate a generated image in response to the second text data; and the learning step includes training the first machine learning model using the first training data corresponding to second training data to be invalidated or deleted from the second training data included in the training dataset; and incorporating the first parameter into the second machine learning model.

Images