Information processing method, information processing device, and information processing program

Abstract

This technology provides the ability to generate more accurate models and improve the inference performance of those models. [Solution] The information processing device acquires reference data representing the initial distribution and training data representing the target distribution, inputs the reference data into a forward vector field model that continuously transforms the initial distribution into the target distribution, transforms the initial distribution of the reference data into the target distribution, inputs the training data into an inverse vector field model that continuously transforms the target distribution back into the initial distribution, transforms the target distribution of the training data into the initial distribution, and updates the first parameter of the forward vector field model and the second parameter of the inverse vector field model based on the forward vector field model into which the reference data was input and the inverse vector field model into which the training data was input.

Description

[Technical Field]

[0001] This disclosure relates to a technology for training a model and a technology for generating data using the model. [Background technology]

[0002] In recent years, to address the computational cost challenges of Continuous Normalizing Flows (CNF), Conditional Flow Matching (CFM) has been proposed, which enables efficient learning by having neural networks regress to a conditional vector field (minimizing the mean squared error) (see, for example, Non-Patent Documents 1 and 2). [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] Yaron Lipman, Ricky TQChen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le, "Flow Matching for Generative Modeling," ICLR 2023, February 2, 2023. [Non-Patent Document 2] Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio, "Improving and Generalizing Flow-Based Generative Models with Minibatch Optimal Transport," Transactions on Machine Learning Research, March 11, 2024. [Overview of the project] [Problems that the invention aims to solve]

[0004] However, in the above conventional technologies, various interpolation methods for interpolating the probability path between probability distributions have been proposed. However, depending on the selected interpolation method, the modeling may become inaccurate and the inference performance of the generated model may decrease, and further improvement has been required.

[0005] The present disclosure has been made to solve the above problems, and an object thereof is to provide a technology capable of generating a more accurate model and improving the inference performance of the model.

Means for Solving the Problems

[0006] An information processing method according to an aspect of the present disclosure is an information processing method executed by a computer, including acquiring reference data indicating an initial distribution and learning data indicating a target distribution, inputting the reference data into a forward vector field model that continuously transforms the initial distribution into the target distribution to transform the initial distribution of the reference data into the target distribution, inputting the learning data into a reverse vector field model that continuously reversely transforms the target distribution into the initial distribution to transform the target distribution of the learning data into the initial distribution, and updating a first parameter of the forward vector field model and a second parameter of the reverse vector field model based on the forward vector field model into which the reference data is input and the reverse vector field model into which the learning data is input.

Effects of the Invention

[0007] According to the present disclosure, a more accurate model can be generated and the inference performance of the model can be improved.

Brief Description of the Drawings

[0008] [Figure 1] It is a diagram showing an example of the configuration of an information processing apparatus according to Embodiment 1. [Figure 2]This diagram illustrates the difference between the learning process in conventional CFM and the learning process in this embodiment 1. [Figure 3] This figure shows an example of learning processing by the information processing device in this embodiment 1. [Figure 4] This figure shows an example of inference processing by the information processing device in this embodiment 1. [Figure 5] This figure shows an example of the configuration of the information processing device according to this second embodiment. [Figure 6] This figure shows an example of inference processing by the information processing device in this second embodiment. [Figure 7] This figure shows the results of experiments evaluating the performance of the conventional model and the model of the proposed method. [Modes for carrying out the invention]

[0009] (Knowledge that forms the basis of this disclosure) Traditionally, there is a deep generative model called the flow-based generative model. The flow-based generative model is a generative model that represents the function that generates data from latent variables as a composition of functions that can be inversely transformed. In the flow-based generative model, the generative model is trained to maximize the likelihood of the training data, and the distribution of the training data is modeled.

[0010] Continuous Normalizing Flows (CNF) introduces a continuous time point t[0,1] into the flow-based generative model, extending the flow to a continuous transformation. CNF represents the target distribution with an infinitely long transformation, enabling accurate modeling for discrete flows. However, CNF requires optimization over the entire time period, resulting in high computational costs and the potential for numerical instability and approximation errors.

[0011] Non-Patent Documents 1 and 2 propose Conditional Flow Matching (CFM) as a way to solve the computational cost problem of CNF. Non-Patent Documents 1 and 2 propose various interpolation methods for interpolating stochastic paths between probability distributions. However, depending on the interpolation method selected, the model may become inaccurate, and the inference performance of the generated model may decrease.

[0012] To address the above challenges, the following technologies are disclosed.

[0013] (1) An information processing method according to one aspect of the present disclosure is an information processing method performed by a computer, comprising: acquiring reference data representing an initial distribution and training data representing a target distribution; inputting the reference data into a forward vector field model that continuously transforms the initial distribution into the target distribution, and transforming the initial distribution of the reference data into the target distribution; inputting the training data into an inverse vector field model that continuously transforms the target distribution back into the initial distribution, and transforming the target distribution of the training data into the initial distribution; and updating a first parameter of the forward vector field model and a second parameter of the inverse vector field model based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input.

[0014] This configuration eliminates the need for interpolation of stochastic paths between probability distributions, which was required in conventional CFM. This allows for the generation of more accurate models and improves the inference performance of the models.

[0015] (2) In the information processing method described in (1) above, updating the first parameter and the second parameter may include updating the first parameter and the second parameter such that the cosine distance between the forward vector field model and the inverse vector field model becomes smaller.

[0016] With this configuration, the first and second parameters are updated so that the cosine distance between the forward and backward vector field models is minimized, making it easy to train both the forward and backward vector field models.

[0017] (3) In the information processing method described in (1) above, the reference data may be noise data randomly sampled based on the initial distribution, and the training data may be image data that is the target of image generation.

[0018] With this configuration, the reference data is noise data randomly sampled based on the initial distribution, and the training data is image data that is the target of image generation. Therefore, it is possible to generate a model that generates the desired image data from the noise data.

[0019] (4) In the information processing method described in (1) above, the reference data may be noise data randomly sampled based on the initial distribution, and the training data may be image data that is the target of image recognition.

[0020] With this configuration, the reference data is noise data randomly sampled based on the initial distribution, and the training data is image data that is the target of image recognition, so it is possible to generate a model that recognizes image data.

[0021] (5) In the information processing method described in (1) above, the reference data may be noise data randomly sampled based on the initial distribution, and the training data may be time-series sensing data that is the target of anomaly detection.

[0022] With this configuration, the reference data is noise data randomly sampled based on the initial distribution, and the training data is time-series sensing data that is the target of anomaly detection. Therefore, it is possible to generate a model that can detect anomalies in time-series sensing data.

[0023] Furthermore, this disclosure can be implemented not only as an information processing method that performs the characteristic processing described above, but also as an information processing device having a characteristic configuration corresponding to the characteristic processing performed by the information processing method. It can also be implemented as a computer program that causes a computer to execute the characteristic processing included in such an information processing method. Therefore, the same effects as the above-described information processing method can be achieved in the following other embodiments.

[0024] (6) An information processing apparatus according to another aspect of the present disclosure is an information processing apparatus comprising a processor, the processor acquires reference data representing an initial distribution and training data representing a target distribution, inputs the reference data to a forward vector field model that continuously transforms the initial distribution to the target distribution, transforms the initial distribution of the reference data to the target distribution, inputs the training data to an inverse vector field model that continuously transforms the target distribution back to the initial distribution, transforms the target distribution of the training data to the initial distribution, and updates a first parameter of the forward vector field model and a second parameter of the inverse vector field model based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input.

[0025] (7) An information processing program according to another aspect of the present disclosure acquires reference data representing an initial distribution and training data representing a target distribution, inputs the reference data into a forward vector field model that continuously transforms the initial distribution into the target distribution, transforms the initial distribution of the reference data into the target distribution, inputs the training data into an inverse vector field model that continuously transforms the target distribution back into the initial distribution, transforms the target distribution of the training data into the initial distribution, and causes the computer to function to update a first parameter of the forward vector field model and a second parameter of the inverse vector field model based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input.

[0026] Furthermore, non-temporary computer-readable recording media relating to other aspects of this disclosure record the information processing programs described in (7) above.

[0027] (8) Information processing methods relating to other aspects of the present disclosure are information processing methods performed by a computer, which include: acquiring input data indicating an initial distribution; inputting the input data into a trained forward vector field model that continuously transforms the initial distribution into a target distribution, and transforming the initial distribution of the input data into the target distribution; and acquiring output data output from the forward vector field model as inference data, wherein in training the forward vector field model, reference data indicating the initial distribution is input into the forward vector field model, the initial distribution of the reference data is transformed into the target distribution; training data indicating the target distribution is input into an inverse vector field model that continuously transforms the target distribution back into the initial distribution, the target distribution of the training data is transformed into the initial distribution; and the parameters of the forward vector field model are updated based on the forward vector field model into which the reference data was input and the inverse vector field model into which the training data was input.

[0028] This configuration eliminates the need for interpolation of stochastic paths between probability distributions, which was required in conventional CFM. This allows for the generation of more accurate models and improves the inference performance of the models.

[0029] (9) In the information processing method described in (8) above, the input data and the reference data may be noise data randomly sampled based on the initial distribution, and the learning data and the output data may be image data that is the target of image generation.

[0030] With this configuration, the input data and reference data are noise data randomly sampled based on the initial distribution, and the training data and output data are image data that are the target of image generation. Therefore, it is possible to generate the desired image data from the noise data.

[0031] Furthermore, this disclosure can be implemented not only as an information processing method that performs the characteristic processing described above, but also as an information processing device having a characteristic configuration corresponding to the characteristic processing performed by the information processing method. It can also be implemented as a computer program that causes a computer to execute the characteristic processing included in such an information processing method. Therefore, the same effects as the above-described information processing method can be achieved in the following other embodiments.

[0032] (10) An information processing device according to another aspect of the present disclosure is an information processing device comprising a processor, wherein the processor acquires input data representing an initial distribution, inputs the input data to a trained forward vector field model that continuously transforms the initial distribution to a target distribution, transforms the initial distribution of the input data to the target distribution, acquires output data output from the forward vector field model as inference data, and in training the forward vector field model, reference data representing the initial distribution is input to the forward vector field model, the initial distribution of the reference data is transformed to the target distribution, training data representing the target distribution is input to an inverse vector field model that continuously transforms the target distribution back to the initial distribution, the target distribution of the training data is transformed to the initial distribution, and the parameters of the forward vector field model are updated based on the forward vector field model into which the reference data was input and the inverse vector field model into which the training data was input.

[0033] (11) Information processing programs according to other embodiments of the present disclosure include a computer that acquires input data representing an initial distribution, inputs the input data into a trained forward vector field model that continuously transforms the initial distribution into a target distribution, transforms the initial distribution of the input data into the target distribution, and acquires output data output from the forward vector field model as inference data, wherein in the training of the forward vector field model, reference data representing the initial distribution is input into the forward vector field model, the initial distribution of the reference data is transformed into the target distribution, training data representing the target distribution is input into an inverse vector field model that continuously transforms the target distribution back into the initial distribution, the target distribution of the training data is transformed into the initial distribution, and the parameters of the forward vector field model are updated based on the forward vector field model into which the reference data was input and the inverse vector field model into which the training data was input.

[0034] Furthermore, non-temporary computer-readable recording media relating to other aspects of this disclosure record the information processing programs described in (11) above.

[0035] (12) Information processing methods relating to other aspects of the present disclosure are information processing methods performed by a computer, which include: acquiring evaluation data indicating a target distribution; inputting the evaluation data into a trained inverse vector field model that continuously inversely transforms the target distribution into an initial distribution, converting the target distribution of the evaluation data into the initial distribution; and acquiring output data output from the inverse vector field model as inference data, wherein in the training of the inverse vector field model, reference data indicating the initial distribution is input into a forward vector field model that continuously transforms the initial distribution into the target distribution, the initial distribution of the reference data is converted into the target distribution, training data indicating the target distribution is input into the inverse vector field model, the target distribution of the training data is converted into the initial distribution, and the parameters of the inverse vector field model are updated based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input.

[0036] This configuration eliminates the need for interpolation of stochastic paths between probability distributions, which was required in conventional CFM. This allows for the generation of more accurate models and improves the inference performance of the models.

[0037] (13) In the information processing method described in (12) above, the reference data is noise data randomly sampled based on the initial distribution, and the evaluation data and the training data may be image data that is the target of recognition for image recognition.

[0038] In this configuration, the reference data is noise data randomly sampled based on the initial distribution, and the evaluation data and training data are image data that are the target of image recognition, so the image data can be recognized.

[0039] (14) In the information processing method described in (13) above, the image recognition may further include anomaly detection for detecting anomalies in objects contained in an image, calculating the likelihood between the acquired inference data and a reference distribution indicating a normal state or an abnormal state that is stored in memory in advance, and detecting anomalies in objects contained in the image of the evaluation data based on the calculated likelihood.

[0040] This configuration allows for the detection of anomalies in objects contained within the images of the evaluation data.

[0041] (15) In the information processing method described in (13) above, the image recognition may further include class classification of objects contained in an image into one of a plurality of classes, calculating the likelihood of the acquired inference data and each of a plurality of reference distributions that represent each of the plurality of classes and are stored in memory in advance, and classifying the objects contained in the image of the evaluation data into one of the plurality of classes based on the calculated plurality of likelihoods.

[0042] This configuration allows objects contained in the evaluation data images to be classified into one of several classes.

[0043] (16) In the information processing method described in (12) above, the reference data is noise data randomly sampled based on the initial distribution, the evaluation data and the learning data are time-series sensing data that are the target of anomaly detection, and the method further includes calculating the likelihood between the acquired inference data and a reference distribution that indicates a normal state or an abnormal state and is stored in memory in advance, and detecting anomalies in the sensing data which is the evaluation data based on the calculated likelihood.

[0044] This configuration allows for the detection of anomalies in time-series sensing data, which is used as evaluation data.

[0045] Furthermore, this disclosure can be implemented not only as an information processing method that performs the characteristic processing described above, but also as an information processing device having a characteristic configuration corresponding to the characteristic processing performed by the information processing method. It can also be implemented as a computer program that causes a computer to execute the characteristic processing included in such an information processing method. Therefore, the same effects as the above-described information processing method can be achieved in the following other embodiments.

[0046] (17) An information processing device according to another aspect of the present disclosure is an information processing device comprising a processor, wherein the processor acquires evaluation data representing a target distribution, inputs the evaluation data to a trained inverse vector field model that continuously inversely transforms the target distribution to an initial distribution, converts the target distribution of the evaluation data to the initial distribution, acquires output data output from the inverse vector field model as inference data, and in training the inverse vector field model, reference data representing the initial distribution is input to a forward vector field model that continuously transforms the initial distribution to the target distribution, the initial distribution of the reference data is converted to the target distribution, training data representing the target distribution is input to the inverse vector field model, the target distribution of the training data is converted to the initial distribution, and the parameters of the inverse vector field model are updated based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input.

[0047] (18) Information processing programs according to other embodiments of the present disclosure include: acquiring evaluation data representing a target distribution; inputting the evaluation data into a trained inverse vector field model that continuously inversely transforms the target distribution into an initial distribution; converting the target distribution of the evaluation data into the initial distribution; and causing the computer to function to acquire output data output from the inverse vector field model as inference data; wherein, in training the inverse vector field model, reference data representing the initial distribution is input into a forward vector field model that continuously transforms the initial distribution into the target distribution; the initial distribution of the reference data is converted into the target distribution; training data representing the target distribution is input into the inverse vector field model; the target distribution of the training data is converted into the initial distribution; and the parameters of the inverse vector field model are updated based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input.

[0048] Furthermore, non-temporary computer-readable recording media relating to other aspects of this disclosure record the information processing programs described in (18) above.

[0049] Embodiments of this disclosure will be described below with reference to the attached drawings. Note that the embodiments described below are all specific examples of this disclosure. The numerical values, shapes, components, steps, and order of steps shown in the following embodiments are examples only and are not intended to limit this disclosure. Furthermore, components in the following embodiments that are not described in the independent claim representing the highest-level concept will be described as optional components. Also, in all embodiments, the contents of each can be combined.

[0050] (Embodiment 1) Figure 1 shows an example of the configuration of the information processing device 1 according to this embodiment 1.

[0051] The information processing device 1 comprises a noise data generation unit 11, a learning data storage unit 12, a data acquisition unit 13, a forward conversion unit 14, a reverse conversion unit 15, a learning unit 16, an inference unit 17, and an inference data storage unit 18.

[0052] The information processing device 1 includes, for example, a computer system comprising a control program, a processing circuit such as a processor or logic circuit that executes the control program, and a recording device such as an internal memory or an accessible external memory that stores the control program. The information processing device 1 may be implemented, for example, by a hardware implementation using a processing circuit, by the execution of a software program held in memory by the processing circuit or distributed from an external server, or by a combination of these hardware and software implementations. Furthermore, the information processing device 1 may be a server, a terminal, or a system comprising both a server and a terminal.

[0053] The learning processing unit also includes a noise data generation unit 11, a learning data storage unit 12, a data acquisition unit 13, a forward transformation unit 14, a reverse transformation unit 15, and a learning unit 16. In the learning phase, the learning processing unit learns a forward vector field model and a reverse vector field model. The inference processing unit includes a noise data generation unit 11, a data acquisition unit 13, a forward transformation unit 14, an inference unit 17, and an inference data storage unit 18. In the inference phase, the inference processing unit performs inference on the input data using the forward vector field model.

[0054] During the learning phase, the noise data generation unit 11 generates noise data as reference data, randomly sampled based on an initial distribution. The initial distribution is a simple probability distribution. For example, the initial distribution is a Gaussian distribution.

[0055] The learning data storage unit 12 stores learning data that represents the target distribution. The target distribution is a probability distribution that is more complex than the initial distribution. In this embodiment 1, the learning data is image data that is the target of image generation. For example, if the target of generation is a human face image, the learning data is human face image data.

[0056] During the learning phase, the data acquisition unit 13 acquires reference data representing the initial distribution and learning data representing the target distribution. At this time, the data acquisition unit 13 acquires reference data generated by the noise data generation unit 11. The reference data is noise data randomly sampled based on the initial distribution. Also during the learning phase, the data acquisition unit 13 acquires learning data stored in the learning data storage unit 12. Also during the learning phase, the data acquisition unit 13 samples time.

[0057] The forward transformation unit 14 includes a forward vector field model that continuously transforms the initial distribution to the target distribution. In the learning phase, the forward transformation unit 14 inputs reference data into the forward vector field model that continuously transforms the initial distribution to the target distribution, and transforms the initial distribution of the reference data to the target distribution. The forward vector field model is a neural network model. The type of neural network is not limited. The neural network may be any neural network, such as a deep neural network, a convolutional neural network, or a recurrent neural network.

[0058] The inverse transformation unit 15 includes an inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution. In the learning phase, the inverse transformation unit 15 inputs training data into the inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and transforms the target distribution of the training data back to the initial distribution. The inverse vector field model is a neural network model. The type of neural network is not limited. The neural network may be any neural network, such as a deep neural network, a convolutional neural network, or a recurrent neural network.

[0059] The learning unit 16 updates the first parameter of the forward vector field model and the second parameter of the inverse vector field model based on the forward vector field model with reference data input and the inverse vector field model with training data input. The learning unit 16 updates the first and second parameters so that the cosine distance between the forward and inverse vector field models is minimized. More specifically, the learning unit 16 updates the first and second parameters so that the cosine distance between the forward and inverse vector field models is minimized.

[0060] In the inference phase, the noise data generation unit 11 generates noise data as input data, randomly sampled based on an initial distribution. The initial distribution is a simple probability distribution. For example, the initial distribution is a Gaussian distribution.

[0061] In the inference phase, the data acquisition unit 13 acquires input data that represents the initial distribution. At this time, the data acquisition unit 13 acquires input data generated by the noise data generation unit 11. The input data is noise data randomly sampled based on the initial distribution.

[0062] In the inference phase, the forward transformation unit 14 includes a trained forward vector field model that continuously transforms the initial distribution to the target distribution. In the inference phase, the forward transformation unit 14 inputs the input data into the trained forward vector field model that continuously transforms the initial distribution to the target distribution, and transforms the initial distribution of the input data to the target distribution.

[0063] In training a forward vector field model, reference data representing the initial distribution is input to the forward vector field model, and the initial distribution of the reference data is transformed into the target distribution. Similarly, training data representing the target distribution is input to an inverse vector field model that continuously transforms the target distribution back to the initial distribution, and the target distribution of the training data is transformed into the initial distribution. Then, the parameters of the forward vector field model are updated based on the forward vector field model with the reference data input and the inverse vector field model with the training data input.

[0064] The inference unit 17 acquires the output data from the forward vector field model as inference data. The output data is the image data that is to be generated for image generation.

[0065] The inference data storage unit 18 stores the inference data acquired by the inference unit 17.

[0066] Here, the learning process for the forward vector field model and the inverse vector field model in this embodiment 1 will be explained in more detail.

[0067] Figure 2 is a diagram illustrating the difference between the learning process in a conventional CFM and the learning process in this embodiment 1. In Figure 2, the upper diagram shows the learning process in a conventional CFM, and the lower diagram shows the learning process in this embodiment 1.

[0068] Traditional CFM uses neural networks θ (t,x t ) by the affine transformation φ t Gaussian interpolated vector field u having (x) t(x|z) is regressed. That is, v θ (t, x t ) and u t (x|z), the parameter θ that can be learned is updated so that the mean square error (L CFM ) is minimized.

[0069] On the other hand, the learning of Embodiment 1 has two neural networks with free-form transformations having only the bijective objective x t = φ λ -1 (φ θ (x t )). In the learning of Embodiment 1, in order to eliminate the path interpolation of the CFM, a reverse vector field model v λ (t, y t ) having a learnable parameter λ is further introduced.

[0070] The learning of Embodiment 1 only requires the bijectivity of the free-form transformations φ θ (t, x t ) and the reverse vector field model v λ (t, y t ) respectively generated by φ t and φ t -1 . And the learning of Embodiment 1 by the bijective φ t can be expressed as an ordinary differential equation (ODE) represented by the following formula (1).

[0071]

Equation

[0072] Assuming the bijectivity of φ t (y) = φ t -1 (x), the above formula (1) can be rewritten as the following formula (2) using the inverse function theorem of a single variable.

[0073]

Equation

[0074] Equation (2) above can be expressed as equation (3) below, for the forward vector field model v θ (t,x t ) and the inverse vector field model v λ (t,y t The cosine distance between ) and dist cos The loss function L that minimizes DFM This can be achieved.

[0075]

number

[0076] In equations (1) to (3), x and y represent data points, t represents time, q represents probability density, φ represents flow, u and v represent vector field models, θ and λ represent learnable parameters, and L DFM This represents the loss function.

[0077] The learning unit 16 inputs a forward vector field model v with reference data. θ (t,x t ) and the training data are input to the inverse vector field model v λ (t,y t The cosine distance between () and dist cos The first parameter θ of the forward vector field model and the second parameter λ of the inverse vector field model are updated to minimize the given value.

[0078] In the learning process of this embodiment 1, interpolation of stochastic paths is unnecessary, and only bijectivity is assumed, resulting in fewer constraints and more accurate stochastic modeling compared to conventional CFM.

[0079] Next, the learning process performed by the information processing device 1 in this embodiment 1 will be described.

[0080] Figure 3 shows an example of the learning process performed by the information processing device 1 in this embodiment 1.

[0081] First, in step S1, the noise data generation unit 11 generates noise data randomly sampled based on the initial distribution as reference data.

[0082] Next, in step S2, the data acquisition unit 13 acquires reference data representing the initial distribution generated by the noise data generation unit 11 and training data representing the target distribution stored in the training data storage unit 12.

[0083] Next, in step S3, the data acquisition unit 13 samples a time from the range of 0 to 1.

[0084] Next, in step S4, the forward transformation unit 14 transforms the initial distribution of the reference data into the target distribution based on the reference data, time, and forward vector field model. The forward transformation unit 14 inputs the reference data into the forward vector field model that continuously transforms the initial distribution into the target distribution, and transforms the initial distribution of the reference data into the target distribution. The forward transformation unit 14 then converts the data point x of the reference data at the sampled time t. t Forward vector field model v representing the vector θ (t,x t ) obtain.

[0085] Next, in step S5, the inverse transformation unit 15 transforms the target distribution of the training data into the initial distribution based on the training data, time, and the inverse vector field model. The inverse transformation unit 15 inputs the training data into the inverse vector field model, which continuously transforms the target distribution back to the initial distribution, and transforms the target distribution of the training data into the initial distribution. The inverse transformation unit 15 transforms the data point y of the training data at the sampled time t. t v is an inverse vector field model representing a vector. λ (t,y t ) obtain.

[0086] Next, in step S6, the learning unit 16 updates the first parameter of the forward vector field model and the second parameter of the inverse vector field model based on the forward vector field model and the inverse vector field model. The learning unit 16 updates the forward vector field model v θ (t,x t ) and the inverse vector field model v λ (t,y t The cosine distance between ( ) and ( ) is calculated, and the forward vector field model v is set such that the calculated cosine distance is minimized. θ (t,x t The first parameter θ of the inverse vector field model v λ (t,y t The second parameter λ of ) is updated.

[0087] Next, in step S7, the learning unit 16 determines whether the first and second parameters have been updated a predetermined number of times. The predetermined number of times is predetermined.

[0088] If it is determined that the first and second parameters have not been updated a predetermined number of times (NO in step S7), the process returns to step S1.

[0089] On the other hand, if it is determined that the first and second parameters have been updated a predetermined number of times (YES in step S7), the learning process ends.

[0090] Thus, since interpolation of stochastic paths between probability distributions, which was necessary in conventional CFM, is no longer required, it is possible to generate a more accurate model and improve the inference performance of the model.

[0091] In step S7, the learning unit 16 may determine whether the first and second parameters have converged. If it is determined that the first and second parameters have not converged, the process returns to step S1. If it is determined that the first and second parameters have converged, the learning process may be terminated.

[0092] Next, the inference process performed by the information processing device 1 in this embodiment 1 will be described.

[0093] Figure 4 shows an example of inference processing by the information processing device 1 in this embodiment 1.

[0094] First, in step S11, the noise data generation unit 11 generates noise data as input data, which is randomly sampled based on the initial distribution.

[0095] Next, in step S12, the data acquisition unit 13 acquires input data that represents the initial distribution.

[0096] Next, in step S13, the forward transformation unit 14 transforms the initial distribution of the input data into a target distribution based on the input data and the trained forward vector field model. The forward transformation unit 14 inputs the input data into the trained forward vector field model that continuously transforms the initial distribution into the target distribution, and transforms the initial distribution of the input data into the target distribution.

[0097] Next, in step S14, the inference unit 17 acquires output data representing the target distribution output from the forward vector field model as inference data. The inference data is, for example, image data that is the target of image generation. The inference unit 17 stores the acquired inference data in the inference data storage unit 18. The inference data stored in the inference data storage unit 18 may be output to a terminal connected to the information processing device 1 via a network, or to a display device connected to the information processing device 1. The terminal or display device may display the inference data.

[0098] In this embodiment 1, the training data and output data are image data that are the target of image generation, but this disclosure is not limited to this, and any sensing data may be used. The training data and output data may be, for example, time-series sensing data. The training data may also be time-series sensing data that is the target of anomaly detection. As training data and output data, sensing data such as audio data output from a microphone, point cloud data output from a radar such as LiDAR (Light Detection and Ranging), pressure data output from a pressure sensor, temperature data or humidity data output from a temperature sensor or humidity sensor, or scent data output from a scent sensor may be used.

[0099] (Embodiment 2) In Embodiment 1 described above, the training data is image data that is the target of image generation, whereas in Embodiment 2, the training data is image data that is the target of image recognition. Furthermore, in Embodiment 1, input data representing the initial distribution is acquired, and the input data is input to a trained forward vector field model that continuously transforms the initial distribution to the target distribution, the initial distribution of the input data is transformed to the target distribution, and the output data output from the forward vector field model is acquired as inference data. In contrast, in Embodiment 2, evaluation data representing the target distribution is acquired, and the evaluation data is input to a trained inverse vector field model that continuously transforms the target distribution inversely to the initial distribution, the target distribution of the evaluation data is transformed to the initial distribution, and the output data output from the inverse vector field model is acquired as inference data.

[0100] Figure 5 shows an example of the configuration of the information processing device 1A according to this second embodiment.

[0101] The information processing device 1A includes a noise data generation unit 11, a learning data storage unit 22, a data acquisition unit 23, a forward conversion unit 14, a reverse conversion unit 15, a learning unit 16, an inference unit 27, an inference data storage unit 28, an evaluation data storage unit 29, a likelihood calculation unit 30, a determination unit 31, and a determination result storage unit 32. In this second embodiment, the same reference numerals are used for components that are the same as in the first embodiment, and their descriptions are omitted.

[0102] The information processing device 1A comprises, for example, a computer system including a control program, a processing circuit such as a processor or logic circuit that executes the control program, and a recording device such as an internal memory or an accessible external memory that stores the control program. The information processing device 1A may be implemented, for example, by a hardware implementation using a processing circuit, by the execution of a software program held in memory by the processing circuit or distributed from an external server, or by a combination of these hardware and software implementations. Furthermore, the information processing device 1A may be a server, a terminal, or a system comprising both a server and a terminal.

[0103] The learning processing unit also includes a noise data generation unit 11, a learning data storage unit 22, a data acquisition unit 23, a forward transformation unit 14, a reverse transformation unit 15, and a learning unit 16. In the learning phase, the learning processing unit learns a forward vector field model and a reverse vector field model. The inference processing unit includes a noise data generation unit 11, a data acquisition unit 23, a reverse transformation unit 15, an inference unit 27, an inference data storage unit 28, an evaluation data storage unit 29, a likelihood calculation unit 30, a determination unit 31, and a determination result storage unit 32. In the inference phase, the inference processing unit performs inference on the input data using the reverse vector field model.

[0104] The learning data storage unit 22 stores learning data that represents the target distribution. The target distribution is a probability distribution that is more complex than the initial distribution. In this embodiment 2, the learning data is image data that is the target of image recognition. Image recognition includes anomaly detection, which detects abnormalities in objects contained in the image. For example, anomaly detection may detect abnormalities in items manufactured in a factory.

[0105] During the learning phase, the data acquisition unit 23 acquires reference data representing the initial distribution and learning data representing the target distribution. At this time, the data acquisition unit 23 acquires reference data generated by the noise data generation unit 11. The reference data is noise data randomly sampled based on the initial distribution. Also during the learning phase, the data acquisition unit 23 acquires learning data stored in the learning data storage unit 22. Also during the learning phase, the data acquisition unit 23 samples time.

[0106] The learning method in this second embodiment is the same as the learning method in the first embodiment.

[0107] The evaluation data storage unit 29 stores evaluation data that indicates the target distribution. The evaluation data is image data that is the target of image recognition. If the image recognition is anomaly detection that detects abnormalities in objects contained in an image, the image data is image data of the object to be detected.

[0108] During the inference phase, the data acquisition unit 23 acquires evaluation data that represents the target distribution. At this time, the data acquisition unit 23 reads out the evaluation data stored in the evaluation data storage unit 29.

[0109] In the inference phase, the inverse transformation unit 15 includes a trained inverse vector field model that continuously transforms the target distribution back to the initial distribution. In the inference phase, the inverse transformation unit 15 inputs evaluation data into the trained inverse vector field model that continuously transforms the target distribution back to the initial distribution, and transforms the target distribution of the evaluation data back to the initial distribution.

[0110] In training an inverse vector field model, reference data representing the initial distribution is input to a forward vector field model that continuously transforms the initial distribution to the target distribution, and the initial distribution of the reference data is transformed to the target distribution. Similarly, training data representing the target distribution is input to the inverse vector field model, and the target distribution of the training data is transformed to the initial distribution. Then, the parameters of the inverse vector field model are updated based on the forward vector field model with the reference data input and the inverse vector field model with the training data input.

[0111] The inference unit 27 acquires the output data from the inverse vector field model as inference data. The output data represents the initial distribution, which is a simple probability distribution transformed from the evaluation data. The initial distribution is, for example, a Gaussian distribution.

[0112] The inference data storage unit 28 stores the inference data acquired by the inference unit 27.

[0113] The likelihood calculation unit 30 calculates the likelihood between the inference data acquired by the inference unit 27 and a reference distribution that represents a normal or abnormal state and is stored in memory beforehand. The reference distribution is, for example, a Gaussian distribution that represents a normal or abnormal state. For example, the likelihood calculation unit 30 calculates the likelihood between the Gaussian distribution, which is the inference data, and a Gaussian distribution that represents a normal state (a state without abnormalities) and is stored in memory beforehand.

[0114] The determination unit 31 detects abnormalities in objects included in the image of the evaluation data based on the likelihood calculated by the likelihood calculation unit 30. Specifically, the determination unit 31 compares the likelihood calculated by the likelihood calculation unit 30 with a threshold. If the likelihood calculated by the likelihood calculation unit 30 is greater than or equal to the threshold, the determination unit 31 determines that the object included in the image of the evaluation data is normal. If the likelihood calculated by the likelihood calculation unit 30 is lower than the threshold, the determination unit 31 determines that the object included in the image of the evaluation data is abnormal. If the likelihood between the inferred Gaussian distribution and the Gaussian distribution representing a normal state is greater than or equal to the threshold, the determination unit 31 can determine that the object included in the image of the evaluation data is normal. On the other hand, if the likelihood between the inferred Gaussian distribution and the Gaussian distribution representing a normal state is lower than the threshold, the determination unit 31 can determine that the object included in the image of the evaluation data is abnormal.

[0115] The judgment result storage unit 32 stores the judgment result made by the judgment unit 31. That is, the judgment result storage unit 32 stores a judgment result indicating whether or not an abnormality was detected in an object included in the image of the evaluation data.

[0116] Next, the inference process performed by the information processing device 1A in this second embodiment will be described.

[0117] Figure 6 shows an example of inference processing by the information processing device 1A in this second embodiment.

[0118] First, in step S21, the data acquisition unit 23 acquires evaluation data that shows the target distribution.

[0119] Next, in step S22, the inverse transformation unit 15 transforms the target distribution of the evaluation data into an initial distribution based on the evaluation data and the trained inverse vector field model. The inverse transformation unit 15 inputs the evaluation data into a trained inverse vector field model that continuously transforms the target distribution into an initial distribution, and transforms the target distribution of the evaluation data into an initial distribution.

[0120] Next, in step S23, the inference unit 27 acquires output data representing the initial distribution output from the inverse vector field model as inference data. The inference data is, for example, a Gaussian distribution converted from the image data that is the target of recognition in image recognition. The inference unit 27 stores the acquired inference data in the inference data storage unit 28.

[0121] Next, in step S24, the likelihood calculation unit 30 calculates the likelihood between the inference data acquired by the inference unit 27 and a reference distribution representing a normal state that is stored in memory beforehand. For example, the likelihood calculation unit 30 calculates the likelihood between the Gaussian distribution, which is the inference data, and a Gaussian distribution representing a normal state that is stored in memory beforehand.

[0122] Next, in step S25, the determination unit 31 detects abnormalities in objects included in the image of the evaluation data based on the likelihood calculated by the likelihood calculation unit 30. If the likelihood calculated by the likelihood calculation unit 30 is greater than or equal to a threshold, the determination unit 31 determines that the objects included in the image of the evaluation data are normal. If the likelihood calculated by the likelihood calculation unit 30 is lower than a threshold, the determination unit 31 determines that the objects included in the image of the evaluation data are abnormal. The determination unit 31 stores the determination result in the determination result storage unit 32. The determination result stored in the determination result storage unit 32 may be output to a terminal connected to the information processing device 1 via a network, or to a display device connected to the information processing device 1. The terminal or display device may display the determination result of abnormality detection.

[0123] Thus, since interpolation of stochastic paths between probability distributions, which was necessary in conventional CFM, is no longer required, it is possible to generate a more accurate model and improve the inference performance of the model.

[0124] In this embodiment 2, the training data and evaluation data are image data that are the target of image recognition, but this disclosure is not limited to this, and any sensing data may be used. The training data and evaluation data may be, for example, time-series sensing data. Alternatively, the training data and evaluation data may be time-series sensing data that is the target of anomaly detection. As training data and evaluation data, sensing data such as audio data output from a microphone, point cloud data output from a radar such as LiDAR (Light Detection and Ranging), pressure data output from a pressure sensor, temperature data or humidity data output from a temperature sensor or humidity sensor, or scent data output from a scent sensor may be used.

[0125] The determination unit 31 may detect anomalies in the sensing data, which is evaluation data, based on the likelihood calculated by the likelihood calculation unit 30.

[0126] Next, we will describe the experimental results evaluating the performance of the model generated by the learning process in this second embodiment.

[0127] Figure 7 shows the results of experiments evaluating the performance of the conventional model and the model of the proposed method.

[0128] The experiment used the real-world SMAP (Soil Moisture Active Passive) dataset for unsupervised anomaly detection. The SMAP dataset includes soil samples and telemetry information with 135K data points without anomalies used for training and 428K data points used for testing.

[0129] Furthermore, the models used for comparison were a model generated by conventional CNF ("Base CNF" in Figure 7), a model generated by flow matching (FM) from Non-Patent Literature 1 ("FM" in Figure 7), and a model generated by I-CFM from Non-Patent Literature 2 ("I-CFM" in Figure 7).

[0130] Furthermore, the following metrics were used to evaluate the model: precision, recall, AUC (Area Under the Curve), and F1 score.

[0131] In Figure 7, the numbers shown in bold represent the highest values ​​among the four models, while the numbers shown underlined represent the second highest values ​​among the four models.

[0132] Experiments showed that the performance of the proposed model was generally superior to that of the other three models. In particular, the proposed model increased precision by 6.5% (from 88.2% to 94.7%) and F1 score by 3.1% (from 93.3% to 96.4%).

[0133] In this second embodiment, image recognition may be a class classification that classifies objects contained in an image into one of a plurality of classes. In this case, the likelihood calculation unit 30 may calculate the likelihood of the inference data acquired by the inference unit 27 and each of the plurality of reference distributions that represent each of the plurality of classes and are stored in memory in advance. The memory may store in advance a mixture of Gaussian distributions obtained by superimposing a plurality of Gaussian distributions that represent each of the plurality of classes. The likelihood calculation unit 30 may calculate the likelihood of the Gaussian distribution that is the inference data and each of the Gaussian distributions of the mixture of Gaussian distributions. The determination unit 31 may then classify the objects contained in the image of the evaluation data into one of the plurality of classes based on the plurality of likelihoods calculated by the likelihood calculation unit 30. The determination unit 31 may classify the objects contained in the image of the evaluation data into the class corresponding to the reference distribution with the highest likelihood among the plurality of likelihoods calculated by the likelihood calculation unit 30.

[0134] Furthermore, some or all of the functions of the apparatus according to the embodiments of this disclosure may be realized by a processor such as a CPU executing a program.

[0135] Furthermore, all figures used above are illustrative examples provided to illustrate this disclosure, and this disclosure is not limited to these illustrative figures.

[0136] Furthermore, the order in which the steps shown in the flowchart above are performed is illustrative for the purpose of specifically illustrating this disclosure, and other orders may be used to the extent that similar effects can be achieved. Also, some of the above steps may be performed simultaneously (in parallel) with other steps. [Industrial applicability]

[0137] The technology described herein is useful as a technology for training models and a technology for generating data using models, as it can generate more accurate models and improve the inference performance of the models. [Explanation of symbols]

[0138] 1,1A Information Processing Device 11. Noise data generation unit 12,22 Learning data storage unit 13,23 Data Acquisition Section 14 Forward Direction Conversion Unit 15 Reverse Direction Conversion Section 16. Learning Department 17,27 Reasoning part 18,28 Inference Data Storage Unit 29 Evaluation data storage unit 30 Likelihood calculation unit 31 Judgment section 32 Judgment result storage unit

Claims

1. A method of information processing performed by a computer, Obtain reference data showing the initial distribution and training data showing the target distribution, The process involves inputting the reference data into a forward vector field model that continuously transforms the initial distribution into the target distribution, and transforming the initial distribution of the reference data into the target distribution. The process involves inputting the training data into an inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and transforming the target distribution of the training data back to the initial distribution. Based on the forward vector field model into which the reference data is input and the inverse vector field model into which the learning data is input, the first parameter of the forward vector field model and the second parameter of the inverse vector field model are updated. Information processing methods including

2. Updating the first and second parameters includes updating the first and second parameters such that the cosine distance between the forward vector field model and the inverse vector field model decreases. The information processing method according to claim 1.

3. The aforementioned reference data is noise data randomly sampled based on the aforementioned initial distribution. The aforementioned training data is image data that is the target of image generation. The information processing method according to claim 1.

4. The aforementioned reference data is noise data randomly sampled based on the aforementioned initial distribution. The aforementioned training data is image data that is the target of image recognition. The information processing method according to claim 1.

5. The aforementioned reference data is noise data randomly sampled based on the aforementioned initial distribution. The aforementioned training data is time-series sensing data that is the target of anomaly detection. The information processing method according to claim 1.

6. An information processing device equipped with a processor, The aforementioned processor, We obtain reference data showing the initial distribution and training data showing the target distribution. The reference data is input to a forward vector field model that continuously transforms the initial distribution into the target distribution, and the initial distribution of the reference data is transformed into the target distribution. The training data is input to an inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and the target distribution of the training data is transformed back to the initial distribution. Based on the forward vector field model into which the reference data is input and the inverse vector field model into which the learning data is input, the first parameter of the forward vector field model and the second parameter of the inverse vector field model are updated. Information processing device.

7. We obtain reference data showing the initial distribution and training data showing the target distribution. The reference data is input to a forward vector field model that continuously transforms the initial distribution into the target distribution, and the initial distribution of the reference data is transformed into the target distribution. The training data is input to an inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and the target distribution of the training data is transformed back to the initial distribution. The computer is made to update the first parameter of the forward vector field model and the second parameter of the inverse vector field model based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the learning data has been input. Information processing program.

8. A method of information processing performed by a computer, Obtain input data that shows the initial distribution, The input data is input to a trained forward vector field model that continuously transforms the initial distribution into a target distribution, and the initial distribution of the input data is transformed into the target distribution. The output data from the aforementioned forward vector field model is acquired as inference data, Includes, In the training of the forward vector field model, reference data representing the initial distribution is input to the forward vector field model, the initial distribution of the reference data is transformed into the target distribution, training data representing the target distribution is input to an inverse vector field model that continuously transforms the target distribution back to the initial distribution, the target distribution of the training data is transformed into the initial distribution, and the parameters of the forward vector field model are updated based on the forward vector field model with the reference data input and the inverse vector field model with the training data input. Information processing methods.

9. The input data and the reference data are noise data randomly sampled based on the initial distribution. The aforementioned training data and output data are image data that are the target of image generation. The information processing method according to claim 8.

10. An information processing device equipped with a processor, The aforementioned processor, Obtain input data that shows the initial distribution, The input data is input to a trained forward vector field model that continuously transforms the initial distribution to the target distribution, and the initial distribution of the input data is transformed to the target distribution. The output data from the aforementioned forward vector field model is acquired as inference data. In the training of the forward vector field model, reference data representing the initial distribution is input to the forward vector field model, the initial distribution of the reference data is transformed into the target distribution, training data representing the target distribution is input to an inverse vector field model that continuously transforms the target distribution back to the initial distribution, the target distribution of the training data is transformed into the initial distribution, and the parameters of the forward vector field model are updated based on the forward vector field model with the reference data input and the inverse vector field model with the training data input. Information processing device.

11. Obtain input data that shows the initial distribution, The input data is input to a trained forward vector field model that continuously transforms the initial distribution to the target distribution, and the initial distribution of the input data is transformed to the target distribution. The computer is configured to acquire the output data from the aforementioned forward vector field model as inference data. In the training of the forward vector field model, reference data representing the initial distribution is input to the forward vector field model, the initial distribution of the reference data is transformed into the target distribution, training data representing the target distribution is input to an inverse vector field model that continuously transforms the target distribution back to the initial distribution, the target distribution of the training data is transformed into the initial distribution, and the parameters of the forward vector field model are updated based on the forward vector field model with the reference data input and the inverse vector field model with the training data input. Information processing program.

12. A method of information processing performed by a computer, To obtain evaluation data that shows the target distribution, The evaluation data is input into a trained inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and the target distribution of the evaluation data is transformed back to the initial distribution. The output data from the aforementioned inverse vector field model is acquired as inference data, Includes, In the training of the inverse vector field model, reference data representing the initial distribution is input to a forward vector field model that continuously transforms the initial distribution to the target distribution, the initial distribution of the reference data is transformed to the target distribution, training data representing the target distribution is input to the inverse vector field model, the target distribution of the training data is transformed to the initial distribution, and the parameters of the inverse vector field model are updated based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input. Information processing methods.

13. The aforementioned reference data is noise data randomly sampled based on the aforementioned initial distribution. The aforementioned evaluation data and training data are image data that are the target of image recognition. The information processing method according to claim 12.

14. The aforementioned image recognition includes anomaly detection, which detects abnormalities in objects contained in the image. The likelihood of the acquired inference data being compared with a reference distribution indicating a normal or abnormal state that is pre-stored in memory is calculated. Based on the calculated likelihood, the abnormality of the object contained in the image of the evaluation data is detected, Further including, The information processing method according to claim 13.

15. The aforementioned image recognition includes class classification, which classifies objects contained in an image into one of several classes. The process involves calculating the likelihood of the acquired inference data and each of the multiple reference distributions representing each of the multiple classes that are pre-stored in memory, Based on the calculated likelihoods, the objects contained in the images of the evaluation data are classified into one of several classes. Further including, The information processing method according to claim 13.

16. The aforementioned reference data is noise data randomly sampled based on the aforementioned initial distribution. The evaluation data and the training data are time-series sensing data that are the target of anomaly detection. The likelihood of the acquired inference data being compared with a reference distribution indicating a normal or abnormal state that is pre-stored in memory is calculated. Based on the calculated likelihood, anomalies are detected in the sensing data, which is the evaluation data. Further including, The information processing method according to claim 12.

17. An information processing device equipped with a processor, The aforementioned processor, We obtained evaluation data showing the target distribution, The evaluation data is input to a trained inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and the target distribution of the evaluation data is transformed back to the initial distribution. The output data from the aforementioned inverse vector field model is acquired as inference data. In the training of the inverse vector field model, reference data representing the initial distribution is input to a forward vector field model that continuously transforms the initial distribution to the target distribution, the initial distribution of the reference data is transformed to the target distribution, training data representing the target distribution is input to the inverse vector field model, the target distribution of the training data is transformed to the initial distribution, and the parameters of the inverse vector field model are updated based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input. Information processing device.

18. We obtained evaluation data showing the target distribution, The evaluation data is input to a trained inverse vector field model that continuously inversely transforms the target distribution back to the initial distribution, and the target distribution of the evaluation data is transformed back to the initial distribution. The computer is configured to acquire the output data from the aforementioned inverse vector field model as inference data. In the training of the inverse vector field model, reference data representing the initial distribution is input to a forward vector field model that continuously transforms the initial distribution to the target distribution, the initial distribution of the reference data is transformed to the target distribution, training data representing the target distribution is input to the inverse vector field model, the target distribution of the training data is transformed to the initial distribution, and the parameters of the inverse vector field model are updated based on the forward vector field model into which the reference data has been input and the inverse vector field model into which the training data has been input. Information processing program.

Images