Information processing system, information processing apparatus, method for training inference model, and program
Patent Information
- Application Number
- JP2022128578
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2022-08-12
- Publication Date
- 2025-08-14
Smart Images
Figure 00000000_0000_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to an information processing system, an information processing device, an inference model learning method, and a program for learning an inference model based on a neural network. [Background technology]
[0002] In deep learning, an inference model used for image recognition or the like may be pre-trained in an information processing device that distributes the model, and then distributed to other information processing devices, where additional training may be performed using training data in the information processing device of the distribution destination. Patent Document 1 discloses a technology for learning a watermark pattern in the information processing device of the distribution destination. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] Patent No. 6892844 Summary of the Invention [Problem to be solved by the invention]
[0004] However, in Patent Document 1, there is a possibility that learning may be performed unintended by the distributor of the model. In addition, when the learning data for additional learning is transmitted to the device that distributes the inference model and learning is performed at the distributor, it is difficult to guarantee the confidentiality of the training data. [Means for solving the problem]
[0005] In order to solve the above problem, an information processing system according to one embodiment of the present invention comprises a first information processing device and a second information processing device capable of communicating with the first information processing device via a network, the first information processing device including a first inference model acquisition unit that acquires a first inference model based on a first neural network composed of a first network including a first input layer, a first hidden layer group and a first output layer based on a common inference model, a first learning unit that learns the first inference model using teacher data, and a transmission unit that transmits output information based on forward propagation of the first learning unit to the second information processing device, the second information processing device including a second inference model acquisition unit that acquires a second inference model based on a second neural network composed of a second network including a second input layer, a second hidden layer group and a second output layer based on the common inference model, and a second learning unit that learns the second inference model based on the output information, has. Effect of the Invention
[0006] According to the present invention, it is possible to provide a highly accurate inference model while ensuring the confidentiality of training data. [Brief description of the drawings]
[0007] [Figure 1] FIG. 1 is a diagram showing the configuration of an information processing system according to a first embodiment. [Diagram 2] FIG. 2 is a diagram showing a hardware configuration of a first information processing apparatus according to the first embodiment. [Diagram 3] Schematic diagram showing the network configuration of the inference model related to the present invention. [Figure 4] Schematic diagram showing the learning process of the inference model related to the first embodiment. [Diagram 5] A flow chart showing the learning process of an inference model in the first embodiment. [Figure 6] A flow chart showing the inference processing of the inference model related to the first embodiment. [Figure 7] A flow diagram showing the learning process of an inference model related to variant example 1 of the first embodiment. [Figure 8] A schematic diagram showing the learning process of an inference model related to variant example 1 of the first embodiment. [Figure 9] FIG. 13 is a diagram showing the configuration of an information processing system according to a second embodiment. [Figure 10] A flow chart showing the learning process of an inference model in the second embodiment. [Figure 11] FIG. 13 is a diagram showing the configuration of an information processing system according to a third embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0008] The present invention can be preferably applied to medical data such as raw data (signal data) acquired by a modality and diagnostic data generated from raw data by image reconstruction. The modality includes, for example, an X-ray CT device, an MRI device, a SPECT device, a PET device, and an electrocardiograph. The data to be inferred and the teacher data may be not only medical data, but also information related to the privacy of a patient, such as age, sex, and disease information. The target data is not limited to medical data. For example, the target may be image data depicting a person, text data based on document data, and voice data based on voice, as long as it can be inferred by a neural network.
[0009] Hereinafter, preferred embodiments of the technology disclosed herein will be described with reference to the drawings. In each drawing, the same components are denoted by the same reference numerals, and the description of the same contents will be omitted as appropriate.
[0010] The information processing system of the present invention shows an example of a configuration for providing a highly accurate inference model while ensuring the confidentiality of training data.
[0011] The information processing system is composed of a first information processing device which is an information processing device on the user side of the inference model, and a second information processing device which is an information processing device on the model provider side and can communicate with the first information processing device via a network. The first information processing device and the second information processing device are each provided with a trained inference model which has been trained on the model provider side, or an inference model before training, as a common inference model.
[0012] The information processing system of the present invention performs a learning process of a first inference model, which is a common inference model provided in a first information processing device, and a second inference model, which is a common inference model provided in a second information processing device, based on the training data of the model provider while ensuring the confidentiality of the training data. The second information processing device learns the second inference model by acquiring output information based on the forward propagation of the first inference model to the training data of the model provider.
[0013] The inference model performs various tasks based on a neural network, such as segmenting disease regions from medical image data, detecting disease regions, determining the presence or absence of a disease, etc. The present invention is applicable to inference models that perform any of these tasks.
[0014] The training data refers to a set of data for learning an inference model based on a neural network, and includes learning data and correct answer data annotated to the learning data. The learning data is data used for learning the inference model, and is medical image data when performing a segmentation task, a detection task, or a presence / absence determination task. The correct answer data is data given as a correct answer when learning a task with the inference model, and holds data such as a segmentation mask for the segmentation task, a bounding box indicating the coordinates of the disease area for the detection task, and information on the presence or absence of a disease for the presence / absence determination task.
[0015] [First embodiment] The information processing system according to this embodiment will be described below with reference to FIG.
[0016] The information processing system 1 according to the first embodiment includes a first information processing device 2 which is the user side of the inference model, i.e., an information processing device to which the inference model is distributed, and a second information processing device 3 which is the provider side of the inference model, i.e., an information processing device from which the inference model is distributed, and which can communicate with the first information processing device 2 via a network 4. Here, the first information processing device 2 is provided with a trained inference model which has been trained in advance on the provider side of the model, or an inference model before the training process, as a common inference model with the second information processing device 3. In this embodiment, the provider side of the inference model provides the inference model before the training process to the user side.
[0017] In this embodiment, the information processing system 1 inputs learning data to a first inference model present in the first information processing device 2, and learns the first inference model in the first information processing device 2. Loss information calculated based on the output by forward propagation to the learning data and the correct answer data is transmitted to the second information processing device 3.
[0018] The second information processing device 3 uses the loss information to learn a second inference model, which is a common inference model with the first information processing device 2, by backpropagation. Here, a common inference model refers to an inference model that has a common network structure, the presence or absence of parameters, and the values of the parameters.
[0019] In this embodiment, the information processing system 1 is configured in this way, so that the second inference model present in the second information processing device on the inference model provider side can be learned without sending teacher data to the second information processing device 3. Furthermore, since the loss information and the parameter updates based on the loss information can be performed in a timely manner, the model provider side can check the progress of parameter updates. Therefore, this is more effective than the case where the model provider side performs the learning process once and then obtains the updated parameters.
[0020] The loss information may be calculated for a mini-batch. By acquiring the loss information for a mini-batch, the model learning can be made more efficient, and the loss information can ensure the irreversibility of the training data in the first information processing device 2.
[0021] The loss information is the difference between the correct data and the data output from the output layer after inputting medical image data into the inference model and forward propagating it when learning the inference model, and the inference model is learned by backpropagating the difference. Here, the first information processing device 2 is, for example, a workstation managed by a hospital, and the first inference model is learned on the workstation. The second information processing device 3 is, for example, a server managed by the model creator, and the second inference model is learned on the server.
[0022] The configuration of each information processing device will be described below.
[0023] The first information processing device 2, which constitutes the information processing system 1 and is the recipient of the model, is configured to include a first memory unit 5 that stores information on a common inference model and training data, a first acquisition unit 6 that acquires the first inference model, a first learning unit 7 that learns the first inference model using the training data, and a transmission unit 8 that transmits loss information to the second information processing device 3.
[0024] (First memory unit 5) The first storage unit 5 stores training data and correct answer data, which are teacher data managed by the user of the model, in the first information processing device 2, which is an information processing device on the user side of the model. It also stores the first inference model, which is a common inference model, distributed from the second information processing device 3.
[0025] The first storage unit 5 transmits the teacher data and the first inference model to the first acquisition unit 6 in response to an instruction from the first acquisition unit 6. The learning data and the correct answer data constituting the teacher data of the first information processing device 2 may be medical image data automatically transferred from a modality or an image server in a hospital. The first storage unit 5 may be replaced by an external storage device.
[0026] (First Acquisition Part 6) The first acquisition unit 6 acquires the teacher data and the first inference model from the first memory unit 5. Then, the first acquisition unit 6 transmits information on the learning data constituting the teacher data and the first inference model to the first learning unit 7.
[0027] (First Study Part 7) The first learning unit 7 inputs the learning data acquired from the first acquisition unit 6 to the input layer constituting the first inference model acquired from the first acquisition unit 6, forward propagates the data from the input layer to the intermediate layer, and transmits information based on the forward propagated output to the transmission unit 8. In this embodiment, the first learning unit 7 acquires loss information between the output of the output layer and the correct answer data, and learns the first inference model by backpropagating the loss information. The loss information is also transmitted to the transmission unit 8 as information based on the forward propagated output.
[0028] (Transmitter 8) The transmitting unit 8 transmits output information based on the forward propagated output to the second information processing device. In this embodiment, the output information is loss information calculated by calculating the difference between the correct answer data acquired from the first acquiring unit 6 and the output acquired from the first inference model, and transmits the loss information to the second acquiring unit 10 constituting the second information processing device 3 through the network 4.
[0029] Here, the second information processing device 3 is composed of a second memory unit 9 that stores information of the common inference model, a second acquisition unit 10 that acquires the second inference model and output information, and a second learning unit 11 that learns the second inference model by the backpropagation method based on the loss information, which is the output information.
[0030] (Second memory unit 9) The second storage unit 9 stores a second inference model common to the first inference model in the first information processing device 2. The second storage unit 9 may store teacher data for pre-learning performed before distribution of the model in the second information processing device 3, which is an information processing device on the model provider side, and for post-learning performed after distribution of the model. The second storage unit 9 may store inference target data for verification to verify whether the second inference model in the second learning unit 11 has been properly learned. The first inference model transmits information of the inference model to the first information processing device 2, which is an information processing device on the model user side, via the network 4 as appropriate. The second storage unit 9 also transmits information of the second inference model to the second acquisition unit based on an instruction from the second acquisition unit 10.
[0031] (Second Acquisition Unit 10) The second acquisition unit 10 acquires the loss information and information of the second inference model calculated by the calculation unit 8 of the first information processing device 2, and transmits them to the second learning unit 11.
[0032] (Second Study Section 11) The second learning unit 11 learns the second inference model by inputting the loss information acquired from the first information processing device 2 from the output layer of the second inference model acquired from the second acquisition unit 10 and performing backpropagation.
[0033] Here, the first information processing device 2 and the second information processing device 3 may be configured by a computer including a processor, a memory, a storage, and the like. In this case, a program stored in the storage is loaded into the memory, and the processor executes the program, whereby the functions and processing of the second information processing device 3, such as the second storage unit 9, the second acquisition unit 10, and the second learning unit 11, are realized from the first storage unit 5, the first acquisition unit 6, the first learning unit 7, and the transmission unit 8. However, this configuration is not limited, and for example, all or part of the configuration of the first information processing device 2 may be realized by a dedicated processor (such as ASIC) or FPGA. Alternatively, part of the arithmetic processing may be executed by a processor such as a GPU or DSP. In addition, the first information processing device 2 and the second information processing device 3 may be configured by a single piece of hardware, or may be configured by multiple pieces of hardware. For example, the functions and processing may be realized by multiple computers working together using cloud computing or distributed computing.
[0034] 2 shows an example of a specific configuration of the first information processing device 2. In this example, the first information processing device 2 has a CPU 20, a GPU 21, a RAM 22, a ROM 23, and a storage device 24, which are connected via a system bus 25. The second information processing device 3 may be configured in a similar manner. In addition, the first information processing device 2 is connected to a display device 26 and an input device 27 such as a mouse and a keyboard.
[0035] By configuring the first information processing device 2 as described above, a user of the first inference model, who is the administrator of the teacher data, does not need to send the teacher data to an external information processing device, and can learn the first inference model while ensuring the confidentiality of the teacher data.
[0036] In addition, for the provider of the inference model, the accuracy of the inference model can be improved by additionally learning the inference model or learning the inference model using the teacher data acquired by the user of the inference model. Furthermore, since the model provider can check the progress of parameter updates, it is also effective in terms of accuracy verification, etc., compared to acquiring updated parameters after performing a full set of learning processing on the model provider side.
[0037] Here, FIG. 3 is used to show the network configuration of the inference model in the information processing system 1 of the present invention.
[0038] The information processing system 1 of the present invention has a first inference model and a second inference model arranged in a first information processing device 2 and a second information processing device 3, respectively.
[0039] The first inference model and the second inference model are common inference models, and the input layer 201, the intermediate layer A202, the intermediate layer B203, the intermediate layer C204, and the output layer 205 constituting the intermediate layer group constituting the first inference model are common to the input layer 301, the intermediate layer A302, the intermediate layer B303, the intermediate layer C304, and the output layer 305 constituting the intermediate layer group constituting the second inference model, respectively, in terms of the network configuration, the presence or absence of learned parameters, and, if there are learned parameters, the parameters. In other embodiments, the network configuration of the inference models is also common.
[0040] FIG. 4 is a diagram showing a schematic diagram of the learning flow of the inference model in this embodiment.
[0041] The information processing system 1 is composed of a first information processing device 2 which is an information processing device on the user side of the inference model, and a second information processing device 3 which is an information processing device on the provider side of the inference model. Here, the first information processing device 2 and the second information processing device 3 each have a first inference model and a second inference model which are common inference models.
[0042] The first inference model is composed of an input layer 201, an intermediate layer A 202, an intermediate layer B 203, an intermediate layer C 204, and an output layer 205.
[0043] The second inference model is composed of an input layer 301, an intermediate layer A 302, an intermediate layer B 303, an intermediate layer C 304, and an output layer 305, and is an inference model common to the first inference model.
[0044] The first learning unit 7 applies the teacher data, which is obtained from the first acquisition unit 6 and is composed of the learning data and the correct answer data, to the first inference model to perform learning processing. Specifically, the first learning unit 7 in the first information processing device 2 inputs the learning data, which is among the teacher data, to the input layer 201 constituting the first inference model, and forward propagates it to the intermediate layer A 202, the intermediate layer B 203, the intermediate layer C 204, and the output layer 205. The first learning unit 7 calculates the difference between the output information from the output layer 205 and the correct answer data as loss information, and backpropagates it to the first inference model to train the first inference model.
[0045] In addition, the first learning unit 7 transmits the loss information acquired from the inference model to the transmission unit 8. The transmission unit 8 transmits the loss information, which is output information based on the forward propagation of the first learning unit 7, to the second information processing device 3.
[0046] The second learning unit 11 constituting the second information processing device 3 inputs loss information, which is output information based on forward propagation obtained from the first information processing device 2, to the output layer 305 of the second inference model, and learns the second inference model by an error backpropagation method in which the loss information is backpropagated in the order of the output layer 305, intermediate layer C304, intermediate layer B303, intermediate layer A302, and input layer 301.
[0047] The learning process of the inference model in the information processing system 1 in this embodiment will be described below with reference to FIG.
[0048] In step S31, the first acquisition unit 6 in the first information processing device 2 acquires teacher data consisting of learning data and correct answer data and information on the first inference model from the first memory unit 5 in the first information processing device 2, transmits them to the first learning unit 7, and then proceeds to the next step.
[0049] In step S32, the first learning unit 7 inputs learning data to the first inference model, forward propagates the data, obtains loss information between the output of the output layer and the correct answer data, and back-propagates the loss information to learn the first inference model. After the first learning unit 7 transmits the loss information to the transmitting unit 8, the process proceeds to the next step.
[0050] In step S33, the transmission unit 8 transmits the loss information, which is output information based on the forward propagation of the first learning unit 7, to the second acquisition unit 10 in the second information processing device 3 via the network 4, and proceeds to the next step.
[0051] In step S34, the second acquisition unit 10 in the second information processing device 3 transmits the acquired loss information to the second learning unit 11, and proceeds to the next step.
[0052] In step S35, the second learning unit 11 learns the second inference model by backpropagating the acquired loss information to the second inference model.
[0053] In step S36, it is determined whether learning of the first inference model has been completed. If there is training data for learning, the process proceeds to step S33; if there is no training data, the learning process ends.
[0054] According to the information processing system 1 of this embodiment, in the first information processing device 2, the learning data is input to the first inference model and forward propagated, and the loss information calculated based on the data output from the output layer and the correct answer data is transmitted to the second information processing device 3, and the second inference model is learned, so that both inference models can be learned while ensuring the confidentiality of the medical image data. This makes it possible to provide a highly accurate inference model while ensuring the confidentiality of the learning data.
[0055] Here, the inference process using the inference model by the information processing system 1 in this embodiment will be described with reference to Fig. 6. Here, the first information processing device 2 and the second information processing device 3 are information processing devices with different management entities, but the first inference model and the second inference model in each information processing device are inference models that can be inferred independently. Here, only the inference process of the first inference model will be described.
[0056] In step S40, the inference target data acquisition unit in the first information processing device 2 acquires the inference target data and the first inference model from the first memory unit 5, transmits each piece of information to the inference unit, and proceeds to the next step.
[0057] In step S41, the inference unit performs an inference process by applying a first inference model to the inference target data, and transmits the inference result to the output unit, and then proceeds to the next step.
[0058] In step S42, the output unit outputs the inference result to a display device, etc. With this configuration, a user of the inference model can obtain the inference result without transmitting the inference target data to an external information processing device.
[0059] The merit of the second information processing device 3 performing the inference process will be described below. The second information processing device 3 performs the inference process on inference target data for verification to verify whether the inference model has been properly learned, using the second inference model acquired through the above-mentioned learning process. With this configuration, the provider of the inference model can verify the performance of the inference model without transmitting the verification data to an external information processing device. As will be described in detail in the fourth embodiment, the more abundant the high-quality teacher data is, the better the performance of the inference model is expected to be. Therefore, by distributing a common inference model to users of the inference model consisting of multiple entities, the common inference model can be learned using teacher data held by the users of the multiple inference models, thereby constructing a highly accurate inference model. Since the second inference model is provided in the second information processing device 3, which is the information processing device on the inference model provider side, the second model, which is a common inference model, can be easily distributed to users of other inference models, which has the effect of making it easy to construct a learning environment for multiple users.
[0060] (Modification 1 of the first embodiment) In the above-described embodiment, a case has been described in which, when learning an inference model, the first learning unit 7 in the first information processing device 2 transmits loss information calculated based on the data output from the output layer and the correct answer data as output information based on forward propagation to the second information processing device 3 in the server by the transmitting unit 8, and the second inference model on the server side is learned. In this modified example, an example will be described in which output information from the intermediate layer constituting the first inference model is transmitted as output information based on forward propagation.
[0061] Specifically, the output from the intermediate layer of the inference model obtained by the first learning unit 7 and the correct answer data are transmitted to the second information processing device by the transmitting unit 8, and the second learning unit 11 in the second information processing device forward propagates the output information from the intermediate layer to the second inference model, calculates loss information between the correct answer data transmitted from the first information processing device 2, and back propagates the loss information to learn the second inference model on the server side.
[0062] FIG. 7 is a diagram showing a schematic flow of the learning process of the inference model in this modified example.
[0063] The information processing system 1 in this modified example includes a first information processing device 2 and a second information processing device 3, which are information processing devices on the user side of the model. The first information processing device 2 and the second information processing device 3 each have a first inference model and a second inference model, which are common inference models. The first inference model is composed of an input layer 201, an intermediate layer A 202, an intermediate layer B 203, an intermediate layer C 204, and an output layer 205. The second inference model is composed of an input layer 301, an intermediate layer A 302, an intermediate layer B 303, an intermediate layer C 304, and an output layer 305, and is the same as the first inference model.
[0064] The first learning unit 7 performs a learning process by applying the teacher data, which is composed of learning data and correct answer data and which is acquired from the first acquisition unit 6, to the first inference model. In this modification, the first learning unit 7 transmits the correct answer data to the transmission unit 8, and the transmission unit 8 transmits the correct answer data to the second information processing device 3.
[0065] The first learning unit 7 of the first information processing device 2 inputs the learning data into the first inference model, and forward propagates it up to the intermediate layer A202. Then, the output information of the intermediate layer A202 is duplicated in two, one of which is transmitted to the intermediate layer B202 and the other to the transmitting unit 8. From the intermediate layer B203, forward propagates it up to the intermediate layer C204 and the output layer 205, and the difference between the output information from the output layer 205 and the correct answer data is calculated as loss information, and the first inference model is backpropagated to learn it.
[0066] Furthermore, the transmission unit 8 transmits the output information from the hidden layer A 202 to the second information processing device 3.
[0067] The second learning unit 11 inputs output information from the intermediate layer A202, which is output information based on the forward propagation of the first learning unit 7 acquired from the first information processing device 2, from the intermediate layer B303 of the second inference model, and forward propagates it in the order of the intermediate layer C304 and the output layer 305. The second learning unit 11 further calculates the difference between the correct answer data acquired from the first information processing device 2 and the forward propagated output, and backpropagates it as loss information in the order of the output layer 305, the intermediate layer C304, the intermediate layer B303, the intermediate layer A302, and the input layer 301, thereby learning the second inference model.
[0068] The learning flow of the information processing system 1 in this modified example will be described below with reference to Fig. 8. Note that since S71 is the same as step S31, the description thereof will be omitted.
[0069] In step S72, the first learning unit 7 transmits the correct answer data to the second acquiring unit 11 via the transmitting unit 8, and proceeds to the next step.
[0070] In step S73, the first learning unit 7 inputs the learning data to the first inference model, forward propagates it up to a predetermined intermediate layer, and copies the output information from the predetermined intermediate layer. One of the output information is forward propagated as it is to the remaining part of the first inference model, and the other is transmitted via the transmitting unit 8 to the second acquiring unit 11 of the second information processing device 3 through the network 4, and the process proceeds to the next step.
[0071] In step S74, the first learning unit 7 acquires output information from the intermediate layer in step S73 for the remaining part of the first inference model (intermediate layers after the predetermined intermediate layer) and forward propagates it. Then, it calculates loss information, which is the difference between the forward propagated output and the correct answer data, and learns the first inference model by backpropagating the loss information, and proceeds to the next step.
[0072] In step S75, the second acquisition unit 10 acquires the output information and the correct answer data from the intermediate layer from the transmission unit 8, transmits the acquired information to the second learning unit 11, and proceeds to the next step.
[0073] In step S76, the second learning unit 11 inputs the acquired output information to the intermediate layer after the specified layer and propagates it forward, calculates the loss information with the acquired correct answer data, and propagates it backward, thereby learning a second inference model.
[0074] In step S77, it is determined whether learning of the first inference model has been completed. If there is training data for learning, the process proceeds to step S73. If there is no training data, the learning process ends.
[0075] With the configuration of this modified example, in addition to the effects of the above-mentioned embodiment, by sending the output of the intermediate layer, it is possible to increase the contribution of the learning processing speed by the second information processing device 3 when learning the second inference model.
[0076] [Second embodiment] The functional configuration of the information processing system 1 in this embodiment will be described with reference to FIG.
[0077] In the second embodiment, the second information processing device 3 further includes a judgment unit 910 that judges whether the learning of the inference model is being performed appropriately. The judgment unit 910 monitors the learning of the second inference model, and thereby it is possible to judge whether unintended learning is being performed in the inference model of the first information processing device, and to prevent unintended accuracy degradation. Unintended learning refers to learning an inference model using learning data that has been processed so that the inference model produces erroneous output for certain learning data, such as an adversarial attack. Unintended learning is detected by a known method.
[0078] Specifically, the judgment unit 910 compares the parameters of the inference model distributed to the user with the parameters of the inference model being learned by the user, and judges whether the change in the parameters is learning beyond a predetermined range. If the judgment unit 910 judges that the learning is beyond the predetermined range, it does not permit the update of the common model. Based on the judgment, at least one of the first learning unit 7 and the second learning unit 11 does not perform the learning process of the inference model, or does not use the second inference model after the learning process as the common model, but uses the second inference model before learning as the common model, thereby preventing a decrease in the accuracy of the inference model.
[0079] Here, the learning process of the information processing system 1 in this embodiment will be described with reference to FIG.
[0080] It should be noted that steps S1001 to S1005 are the same as steps S31 to S35, and therefore the description thereof will be omitted.
[0081] In step S1006, the judgment unit 910 determines whether the learning of the second inference model is being performed appropriately, thereby determining whether unintended learning is being performed in the inference model of the first information processing device, and if it is determined that unintended learning is not being performed, proceeds to the next step, and if it is determined that unintended learning is being performed, terminates.
[0082] In step S1007, it is determined whether learning of the first inference model has been completed. If learning has not been completed, the process proceeds to step S1003; if learning has been completed, the process ends.
[0083] (Modification 1 of the second embodiment) In the above-described embodiment, a case has been described in which unintended learning is detected in the learning of the inference model of the first information processing device by monitoring the learning of the inference model in the second information processing device. The implementation of the present invention is not limited to this, and when unintended learning is detected in the learning of the inference model of the first information processing device, the update of the learning of the inference model of the second information processing device may be stopped, and an alert may be issued to stop the update of the learning of the inference model of the first information processing device.
[0084] This enables the user of the inference model to recognize that unintended learning is occurring in the first information processing device 2, and to take action such as using the inference model before the unintended learning occurred, or not using the inference model.
[0085] [Third embodiment] In the third embodiment, we describe a case in which a common inference model is learned by multiple hospitals (users of multiple inference models), and the inference model on the server, which is an information processing device on the provider side of the inference model, is updated using loss information, which is output information based on forward propagation calculated by an information processing device on the user side of the inference model.
[0086] The information processing system according to this embodiment will be described below with reference to FIG.
[0087] The information processing system 1 according to this embodiment includes a first information processing device 2 and a third information processing device 110 which are the user side of the inference model, i.e., the information processing devices to which the inference model is distributed, and a second information processing device 3 which is the provider side of the inference model, i.e., the information processing device from which the inference model is distributed, and which can communicate via a network 4. Here, the first information processing device 2 and the third information processing device 110 are provided with a trained inference model which has been trained in advance on the model provider side, or an inference model before the training process, as a common inference model with the second information processing device 3.
[0088] In this embodiment, the information processing system 1 inputs learning data to a first inference model present in the first information processing device 2, and learns the first inference model in the first information processing device 2. Loss information calculated based on the output by forward propagation to the learning data and the correct answer data is transmitted to the second information processing device 3 and the third information processing device 110.
[0089] The second information processing device 3 and the third information processing device 110 learn a second inference model, which is a common inference model with the first information processing device 2, by backpropagation using the loss information.
[0090] In this embodiment, we have described a case where the information processing devices to which the inference model is distributed are the first information processing device 2 and the third information processing device 110, but this is not limited to this, and the number of information processing devices to which the inference model is distributed may be three or more.
[0091] In this embodiment, the information processing system 1 is configured in this way, so that the second inference model on the information processing device of the model provider can be learned without sending teacher data to the second information processing device 3. Furthermore, compared to a case where the model provider performs a full set of learning processes and then acquires updated parameters, the loss information and the update of parameters based on the loss information can be performed in a timely manner, so that the model provider can check the progress of parameter updates.
[0092] (Variation 1) In the above-described embodiment, the case where the inference model provided from the second information processing device to the first information processing device is an inference model before learning processing has been described. The implementation of the present invention is not limited to this, and an inference model pre-trained in the second information processing device may be provided to the user as the inference model. This makes it possible to pre-train the inference model to be provided with various images.
Claims
1. a first information processing device and a second information processing device capable of communicating with the first information processing device via a network; The first information processing device a first inference model acquisition unit that acquires a first inference model based on a first neural network including a first input layer, a first group of hidden layers, and a first output layer; a first learning unit that uses training data to learn the first inference model; a transmitter that transmits output information based on the forward propagation of the first learning unit to a second information processing device, The second information processing device a second inference model acquisition unit that acquires a second inference model based on a second neural network including a second input layer, a second group of hidden layers, and a second output layer, which is the same as the first inference model and is a common inference model; a second learning unit that learns the second inference model based on the output information; An information processing system comprising:
2. 2. The information processing system according to claim 1, wherein the output information is output information from a first intermediate layer group constituting the first neural network when the training data is forward propagated, among the teacher data including the training data.
3. The information processing system according to claim 1, characterized in that the output information is loss information between the output when the training data is forward propagated and the correct answer data, among the teacher data consisting of the training data and the correct answer data.
4. An information processing system as described in either one of claims 2 or 3, characterized in that it has an update unit that updates the common inference model using the second inference model learned by the second learning unit as a common inference model.
5. An information processing system as described in claim 4, characterized in that the second inference model is transmitted to another information processing device.
6. The first information processing device is managed by an entity to which the first inference model is distributed, 4. An information processing system according to claim 2 or 3, wherein the second information processing device is managed by the creator of the inference model.
7. The system is composed of a plurality of first information processing devices each having a different subject; 4. The information processing system according to claim 2, wherein the first learning unit performs a learning process.
8. An information processing system as described in claim 2 or 3, further comprising a judgment unit that compares the learning process by the second learning unit with the common inference model to determine whether the learning exceeds a predetermined range.
9. When the determination unit determines that the learning exceeds a predetermined range, The information processing system according to claim 8, wherein the common inference model is not updated.
10. 2. The information processing system according to claim 1, wherein the first inference model is an inference model distributed from the second information processing device.
11. 11. The information processing system according to claim 10, wherein the second information processing device pre-trains the second inference model and distributes it as the common inference model.
12. an acquisition step of acquiring loss information, which is a difference between correct answer data and an output by a common inference model, acquired from different information processing devices; a learning step of learning the common inference model based on the loss information; A method for learning an inference model, comprising:
13. A program for executing the learning method according to claim 12 on a computer.
14. A first inference model acquisition unit that acquires a first inference model based on a first neural network including a first input layer, a first group of hidden layers, and a first output layer; a first learning unit that uses training data to learn the first inference model; a transmitting unit that transmits output information based on the forward propagation of the first learning unit to a second information processing device that is another information processing device; An information processing device comprising:
15. an acquisition unit that acquires output information from the first inference model calculated by another information processing device; a second inference model acquisition unit that acquires a second inference model based on a second neural network including a second input layer, a second group of hidden layers, and a second output layer, the second inference model being the same as the first inference model and based on a common inference model; a second learning unit that learns the second inference model based on the output information; An information processing device comprising: