Prediction program, training program, prediction method, training method, and information processing apparatus

The autoencoder training method improves noise removal performance in denatured data by using time-varying data and advanced learning techniques, effectively addressing the limitations of conventional deep learning approaches.

JP2025166634APending Publication Date: 2025-11-06FUJITSU LTD +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
JP2024070804
Authority / Receiving Office
JP · JP
Patent Type
Applications
Current Assignee / Owner
Filing Date
2024-04-24
Publication Date
2025-11-06

AI Technical Summary

Technical Problem

Conventional noise removal techniques using deep learning suffer from poor performance and lack statistical guarantees, particularly in denatured data with noise.

Method used

A computer-based method trains an autoencoder to predict clean data by using data items with time-varying noise, adjusting its parameters to minimize loss functions and employing techniques like Monte Carlo methods and transfer learning to improve noise removal performance.

Benefits of technology

The method enhances noise removal performance by accurately predicting clean data from denatured and noisy data, addressing the limitations of conventional techniques.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 2025166634000001_ABST
    Figure 2025166634000001_ABST
Patent Text Reader

Abstract

To improve noise removal performance.SOLUTION: An autoencoder is trained so that, with a plurality of pieces of data including a change with the lapse of time and including noise as training data, its output when data of reference timing and information on the lapse of time are input thereto is closer to data of timing corresponding to the information on the lapse of time. An information processing apparatus predicts the output from the autoencoder when the data of the reference timing and information indicating the reference timing as the information on the lapse of time are input thereto, as noiseless data corresponding to the inputted data of the reference timing.SELECTED DRAWING: Figure 2
Need to check novelty before this filing date? Find Prior Art

Description

[Technical Field]

[0001] The present invention relates to a prediction program and the like. [Background technology]

[0002] When target data is acquired (photographed, recorded, etc.) over a period of time under certain circumstances, the clean data may be denatured over time, and the denatured data may contain additional noise. In the following description, data that contains changes over time and noise will be referred to as "denatured data with noise."

[0003] There is a conventional technique that uses deep learning to remove noise from data with denatured noise. For example, in the conventional technique, image pairs are specified from video data, and a learning model is trained using the specified image pairs. In such a conventional technique, the denatured noise-containing data is input to the trained learning model, and clean data is estimated. [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] J. Xu, E. Adalsteinsson, Deformed2Self: Self-supervised denoising for dynamic medical imaging, in: Medical Image Computing and Computer Assisted Intervention - MICCAI 2021, Springer International Publishing, Cham,2021, pp. 25-35. Summary of the Invention [Problem to be solved by the invention]

[0005] The above-mentioned conventional techniques have the problem of poor noise removal performance, and the conventional techniques do not provide any statistical guarantee regarding the noise removal performance.

[0006] In one aspect, the present invention aims to provide a prediction program, a training program, a prediction method, a training method, and an information processing device that can improve noise removal performance. [Means for solving the problem]

[0007] In the first proposal, a computer executes the following process: Using a plurality of data items that contain changes over time and that contain noise as training data, the computer predicts the output of an autoencoder that is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time, as noise-free data corresponding to the input data at the reference timing. [Effects of the Invention]

[0008] The noise removal performance can be improved. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 1 is a diagram for explaining training data. [Figure 2] FIG. 2 is a diagram for explaining the process of training the autoencoder according to the first embodiment. [Figure 3] FIG. 3 is a functional block diagram illustrating a configuration of the information processing device according to the first embodiment. [Figure 4] FIG. 4 is a flowchart showing a processing procedure during training according to the first embodiment. [Figure 5] FIG. 5 is a flowchart illustrating a processing procedure at the time of prediction according to the first embodiment. [Figure 6]FIG. 6 is a diagram illustrating a process of training an autoencoder according to the second embodiment. [Figure 7] FIG. 7 is a functional block diagram illustrating a configuration of an information processing device according to the second embodiment. [Figure 8] FIG. 8 is a flowchart showing a processing procedure during training according to the second embodiment. [Figure 9] FIG. 9 is a flowchart illustrating a processing procedure at the time of prediction according to the second embodiment. [Figure 10] FIG. 10 is a diagram for explaining the process of creating a plurality of tasks. [Figure 11] FIG. 11 is a diagram illustrating a process of training an autoencoder according to the third embodiment. [Figure 12] FIG. 12 is a functional block diagram illustrating a configuration of an information processing device according to the third embodiment. [Figure 13] FIG. 13 is a flowchart illustrating a processing procedure during training according to the third embodiment. [Figure 14] FIG. 14 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing apparatus of this embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the prediction program, training program, prediction method, training method, and information processing device disclosed in the present application will be described in detail with reference to the accompanying drawings. However, the present invention is not limited to these embodiments. [Example]

[0011] The information processing device according to the first embodiment is referred to as "information processing device 100." The information processing device 100 trains an autoencoder and predicts clean data for data with denatured noise using the trained autoencoder. The process of training the autoencoder and the process of predicting clean data executed by the information processing device 100 will be described below in order.

[0012] An example of processing for training an autoencoder by the information processing device 100 will be described. The information processing device 100 trains the autoencoder using training data shown in Fig. 1. Fig. 1 is a diagram for explaining the training data.

[0013] As shown in Fig. 1, the training data 141 includes modified noise-added data for each time change. For example, the modified noise-added data for time change 0 is "y 0,j The data with denatured noise for time change τ1 is "y τ1,j ". Time change τ N The denatured and noisy data is "y τN,j " "The "j" assigned to each denatured noise-added data represents the data series number. M is the maximum value of the data series number.

[0014] x0 to x shown in Figure 1 τn is "clean data for each theoretical time change." Clean data is subject to denaturation and damage due to time change. Clean data "x0" with a time change of 0 is clean data that the information processing device 100 is to predict.

[0015] The clean data "x0" is given noise and the modified noise-added data "y 0,j」 " corresponds to "x τ1 The data with noise added to "y" is the modified noise-added data "y 0,τ1 " corresponds to "x τN The data with noise added to "y" is the modified noise-added data "y 0,τN " corresponds to

[0016] The assumptions regarding the denatured and noisy data are shown in equations (1), (2), and (3).

[0017]

number

number

number

[0018] The information processing device 100 does not generate training data 141 by adding noise to clean data of each theoretical time change, but directly acquires training data 141 from an external device. For example, the external device is an MRI (Magnetic Resonance Imaging) device. The training data 141 is MRI image data of each time change. In the first embodiment, the training data 141 is described as image data, but may be audio data or the like.

[0019] The information processing device 100 trains an autoencoder using the training data 141 described in Fig. 1. Fig. 2 is a diagram for explaining the process of training the autoencoder according to the first embodiment. For example, the autoencoder 30 includes an encoding unit 30a and a decoding unit 30b.

[0020] The information processing device 100 inputs the denatured noise-added data “y 0,j ” and time-varying information “τ i ” is input, the output of the autoencoder 30 is “f(y 0,j ,τ i The information processing device 100 calculates f(y 0,j ,τ i ) but y τi,j The information processing device 100 updates the parameters of the autoencoder 30 so that the value approaches j=1 to M and i=1 to N. The information processing device 100 trains the autoencoder 30 by repeatedly executing the above process for j=1 to M and i=1 to N.

[0021] The expected loss for the autoencoder 30 is given by equation (4). Approximating equation (4) using the Monte Carlo method allows us to define the empirical loss shown in equation (5).

[0022]

number

[0023]

number

[0024] Training the autoencoder 30 by the information processing device 100 means updating the parameters of the autoencoder 30 so as to minimize the value of equation (5).

[0025] An example of the process of training the autoencoder 30 executed by the information processing device 100 has been described above. As described above, the information processing device 100 uses the denatured noise-added data “y 0,j " and time change τ i The output when inputting f(y 0,j ,τ i ) but y τi,j The parameters of the autoencoder 30 are updated so that the value approaches the value of x0. This makes it possible to generate an autoencoder 30 that can predict the clean data x0 when the time change τ=0.

[0026] Next, an example of a process for predicting clean data executed by the information processing device 100 will be described. The clean data to be predicted is "x0" explained in FIG. 1. The information processing device 100 inputs the modified noise-added data "y 0,j " and time change "0 (τ=0)" to obtain f(y 0,j The information processing device 100 repeats the above process for j=1 to M, and calculates M f(y 0,j ,0) is predicted as the clean data x0.

[0027] For example, the information processing device 100 predicts clean data based on equation (6).

[0028]

number

[0029] The above describes an example of the process of predicting clean data executed by the information processing device 100. As described above, the information processing device 100 inputs the modified noise-added data “y 0,j " and time change "0 (τ=0)" to obtain f(y 0,j , 0) is calculated. This allows us to predict the clean data x0.

[0030] Next, a configuration example of the information processing device 100 will be described. Fig. 3 is a functional block diagram showing the configuration of the information processing device according to the present embodiment 1. As shown in Fig. 3, the information processing device 100 includes a communication unit 110, an input unit 120, a display unit 130, a storage unit 140, and a control unit 150.

[0031] The communication unit 110 executes data communication with an external device or the like via a network. The communication unit 110 is a network interface card (NIC) or the like. For example, the communication unit 110 may acquire training data 141 or the like from the external device or the like.

[0032] The input unit 120 is an input device that inputs various types of information to the control unit 150 of the information processing device 100. For example, the input unit 120 corresponds to a keyboard, a mouse, a touch panel, or the like.

[0033] The display unit 130 is a display device that displays information output from the control unit 150.

[0034] The storage unit 140 includes the autoencoder 30 and training data 141. The storage unit 140 is a memory or the like.

[0035] The autoencoder 30 is the autoencoder 30 described in Fig. 2. The autoencoder 30 is a neural network (NN) or the like.

[0036] The training data 141 is the training data 141 described in Fig. 1. As described in Fig. 1, the training data 141 includes denatured noise-added data for each time change.

[0037] Next, we move on to explaining the control unit 150. The control unit 150 has an acquisition unit 151, a training processing unit 152, and a prediction processing unit 153. The control unit 150 is a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), or the like.

[0038] The acquiring unit 151 acquires the training data 141 from an external device or the like. The acquiring unit 151 stores the training data 141 in the storage unit 140. Note that the training data 141 may be stored in the storage unit 140 in advance.

[0039] The training processing unit 152 trains the autoencoder 30 using the training data 141. For example, the training processing unit 152 trains the denatured noise-added data “y 0,j " and time change τ i The output when inputting f(y 0,j ,τ i ) but y τi,j For example, when training the autoencoder 30, the training processing unit 152 uses the backpropagation method.

[0040] Other operations performed by the training processing unit 152 are the same as those described in FIG. 2 for training the autoencoder 30.

[0041] The prediction processing unit 153 predicts the clean data x0 using the trained autoencoder 30. For example, the prediction processing unit 153 predicts the denatured noise-added data "y 0,j " and time change "0 (τ=0)" to obtain f(y 0,j , 0) <corresponding to clean data>. The prediction processing unit 153 may output the predicted clean data to the display unit 130 for display, or may transmit it to a pre-specified external device.

[0042] Other processes executed by the prediction processing unit 153 are similar to the process of predicting clean data described above.

[0043] Next, an example of the processing procedure of the information processing apparatus 100 according to the first embodiment will be described. FIG. 4 is a flowchart showing the processing procedure during training according to the first embodiment. As shown in FIG. 4, the training processing unit 152 of the information processing apparatus 100 sets j = 1 (step S101). The training processing unit 152 sets i = 0 (step S102).

[0044] The training processing unit 152 inputs the modified noisy data y 0,j and the time change τ i to the autoencoder 30 (step S103). The training processing unit 152 updates the parameters of the autoencoder 30 so that f(y 0,j , 0) output from the autoencoder 30 approaches y τi,j (step S104).

[0045] The training processing unit 152 updates i by i = i + 1 (step S105). When the training processing unit 152 satisfies the condition i < N (step S106, Yes), it proceeds to step S103. On the other hand, when the training processing unit 152 does not satisfy the condition i < N (step S106, No), it proceeds to step S107.

[0046] The training processing unit 152 updates j by j = j + 1 (step S107). When the training processing unit 152 satisfies the condition j < M (step S108, Yes), it proceeds to step S102. On the other hand, when the training processing unit 152 does not satisfy the condition j < M (step S108, No), it outputs the trained autoencoder 30 (step S109).

[0047] In the processing shown in FIG. 4, for one pair of "y 0,j " and τ i , f(y 0,j , 0) approaches y τi,jWhile updating the parameters of the autoencoder 30 so as to approach it, it is not limited to this. For example, the training processing unit 152 may update the parameters of the autoencoder 30 by applying the mini-batch learning method. That is, the training processing unit 152 is "y 0,j ", and for m pairs with τ i , the parameters of the autoencoder 30 may be updated so that each of m f(y 0,j , 0) and y τi,j approaches.

[0048] FIG. 5 is a flowchart showing the processing procedure at the time of prediction according to the first embodiment. Note that the autoencoder 30 described in FIG. 5 is a trained autoencoder 30. As shown in FIG. 5, the prediction processing unit 153 of the information processing apparatus 100 sets j = 1 (step S201).

[0049] The prediction processing unit 153 inputs the modified noisy data "y 0,j " and the time change "0 (τ = 0)" into the autoencoder 30 to calculate f(y 0,j , 0) (step S202).

[0050] The prediction processing unit 153 updates j by j = j + 1 (step S203). When the prediction processing unit 153 satisfies the condition of j < M (step S204, Yes), it proceeds to step S202. On the other hand, when the prediction processing unit 153 does not satisfy the condition of j < M (step S204, No), it proceeds to step S205.

[0051] The prediction processing unit 153 predicts the clean data based on Equation (6) (step S205). The prediction processing unit 153 outputs the predicted clean data (step S206).

[0052] Next, the effects of the information processing apparatus 100 according to the first embodiment will be described. When the information processing apparatus 100 inputs the modified noisy data "y 0,j " included in the training data 141 with the time change τ = 0 and the time change τ i , the output f(y0,j ,τ i ) but y τi,j The parameters of the autoencoder 30 are updated so that the value approaches the value of x0. This makes it possible to generate an autoencoder 30 that can predict the clean data x0 when the time change τ=0.

[0053] The information processing device 100 also inputs the modified noise-added data “y 0,j " and time change "0 (τ=0)" to obtain f(y 0,j , 0) is calculated. This allows us to predict the clean data x0.

[0054] However, the method of the information processing device 100 described above has an issue of difficulty in generalizing denatured noise data when the degree of denaturation over time is strong. For this reason, the autoencoder 30 may be trained, for example, by the following procedure.

[0055] The noise ε that satisfies the condition shown in equation (7) is defined. The operation of adding noise ε to the modified noise-added data "y0" is defined in equation (8). Similarly, the operation of adding noise ε to the modified noise-added data "y τ The operation of adding noise ε' to τ ≠ 0 is defined as m'.

[0056]

number

[0057]

number

[0058] Here, the main objective function L for training the autoencoder 30 (autoencoder f) is D (f) is defined by equation (9).

[0059]

number

[0060] Next, normalization L is performed to bring the mean of the predicted data closer to the mean of the observed data. A (f) is defined by equation (10).

[0061]

number

[0062] Next, the main objective function L shown in Equation (9) D (f) is the empirical loss L approximated by the Monte Carlo method D (f) (hat) is defined by equation (11). Also, the normalized L A (f) is the empirical loss L approximated by the Monte Carlo method A (f) (hat) is defined by equation (12).

[0063]

number

number

[0064] The information processing device 100 calculates the experience loss L D (f) (hat) and experience loss L A (f) (hat) and the sum of the objective function L T Define the objective function L T The experience loss L included in A (f) (hat) may be weighted according to M and N.

[0065] The information processing device 100 calculates the objective function L T For example, the empirical loss L A (f) (hat) is the autoencoder 30, k (y 0,j ) and τ i This means that the autoencoder 30 is trained so that the average value of the input data approaches the average value of the training data 141 to which noise has been added.

[0066] The process in which the information processing device 100 predicts clean data using the trained autoencoder 30 is similar to the process described above. [Example]

[0067] Next, an information processing device according to the second embodiment will be described. The information processing device according to the second embodiment will be referred to as an "information processing device 200." The method of the information processing device 100 according to the first embodiment described above has a problem that when there is a large amount of data with modified noise in the training data 141, the time required for training becomes long. For example, there is a large amount of data with modified noise when N is equal to or greater than a predetermined number.

[0068] For example, the information processing device 200 classifies the training data 141 into a plurality of clusters so as to maximize the amount of mutual information. The information processing device 200 prepares the same number of decoding units included in the autoencoder as the number of cluster labels, and creates a decoding unit specialized for each cluster.

[0069] An example of a process in which the information processing device 200 trains an autoencoder will be described. FIG. 6 is a diagram for explaining a process in which the information processing device 200 trains an autoencoder according to the second embodiment. For example, the information processing device 200 classifies the training data 141 into clusters C1, C2, and C3 by performing clustering on the training data 141. The information processing device 200 also prepares an autoencoder 40. The autoencoder 40 includes an encoding unit 40a and decoding units 40b-1, 40b-2, and 40b-3.

[0070] In FIG. 6, a case where the training data 141 is classified into clusters C1, C2, and C3 will be described as an example, but the training data 141 may be classified into other clusters.

[0071] When the information processing device 200 trains the autoencoder 40 using the training data of the cluster C1, it updates the parameters of the pair of the encoding unit 40a and the decoding unit 40b-1.

[0072] For example, the information processing device 200 may input the denatured noise-added data “y 0,j ” and time-varying information “τ i " is input, the "f(y 0,j ,τ i The information processing device 200 calculates f(y 0,j ,τ i ) but y τi,j The parameters of the autoencoder 40 (parameters of the encoding unit 40a and the decoding unit 40b-1) are updated so that the value approaches

[0073] When the information processing device 200 trains the autoencoder 40 using the training data of the cluster C2, it updates the parameters of the pair of the encoding unit 40a and the decoding unit 40b-2.

[0074] For example, the information processing device 200 may input the denatured noise-added data “y 0,j ” and time-varying information “τ i " is input, the "f(y 0,j ,τ i The information processing device 200 calculates f(y 0,j ,τ i ) but y τi,j The parameters of the autoencoder 40 (parameters of the encoding unit 40a and the decoding unit 40b-2) are updated so that the value approaches

[0075] When training the autoencoder 40 using the training data of the cluster C3, the information processing device 200 performs parameter update for the pair of the encoding unit 40a and the decoding unit 40b-3.

[0076] For example, the information processing device 200 may input the denatured noise-added data “y 0,j ” and time-varying information “τ i" is input, the "f(y 0,j ,τ i The information processing device 200 calculates f(y 0,j ,τ i ) but y τi,j The parameters of the autoencoder 40 (parameters of the encoding unit 40a and the decoding unit 40b-3) are updated so that the value approaches

[0077] The above describes an example of the process of training the autoencoder 40 executed by the information processing device 200. Although the number of decoding units increases and the amount of calculation per epoch increases, since a decoding unit specialized for each cluster is created, the training can be completed in a shorter time than the time required for training by the information processing device 100 of the first embodiment.

[0078] Next, an example of a process for predicting clean data executed by the information processing device 200 will be described. The clean data to be predicted is "x0" explained in FIG. 1. The information processing device 200 predicts the denatured noise-added data "y 0,j The cluster to which " belongs is identified from clusters C1, C2, and C3.

[0079] The information processing device 200 processes the denatured noise-added data “y 0,j If the cluster to which "y" belongs is cluster C1, the denatured noise-added data "y 0,j " and the time change "0 (τ=0)" are input to the encoding unit 40a of the autoencoder 40, and the output of the decoding unit 40b-1 is predicted as clean data.

[0080] The information processing device 200 processes the denatured noise-added data “y 0,j If the cluster to which "y" belongs is cluster C2, the denatured noise-added data "y 0,j " and the time change "0 (τ=0)" are input to the encoding unit 40a of the autoencoder 40, and the output of the decoding unit 40b-2 is predicted as clean data.

[0081] The information processing device 200 processes the denatured noise-added data “y0,j If the cluster to which "y" belongs is cluster C3, the denatured noise-added data "y 0,j " and the time change "0 (τ=0)" are input to the encoding unit 40a of the autoencoder 40, and the output of the decoding unit 40b-3 is predicted as clean data.

[0082] When there are multiple pieces of modified noise-added data to be input, the information processing device 200 may repeatedly perform the above process on each piece of modified noise-added data, and average the output results of each decoding unit 40b-1 to 40b-3 to predict clean data.

[0083] The above describes the process of predicting clean data executed by the information processing device 200. As described above, even if there are multiple decoding units, clean data can be predicted appropriately.

[0084] Next, a description will be given of an example of the configuration of the information processing device 200. Fig. 7 is a functional block diagram showing the configuration of the information processing device according to Example 2. As shown in Fig. 7, the information processing device 200 includes a communication unit 210, an input unit 220, a display unit 230, a storage unit 240, and a control unit 250.

[0085] The explanations regarding the communication unit 210, the input unit 220, and the display unit 230 are the same as the explanations regarding the communication unit 110, the input unit 120, and the display unit 130 explained in FIG.

[0086] The storage unit 240 includes the autoencoder 40 and training data 141. The storage unit 240 is a memory or the like.

[0087] The autoencoder 40 is the autoencoder 40 described in Fig. 6. The autoencoder 40 is a neural network or the like.

[0088] The training data 141 is the training data 141 described in Fig. 1. As described in Fig. 1, the training data 141 includes denatured noise-added data for each time change.

[0089] Next, we move on to explaining the control unit 250. The control unit 250 has an acquisition unit 251, a clustering unit 252, a training processing unit 253, and a prediction processing unit 254. The control unit 250 is a CPU, a GPU, or the like.

[0090] The acquiring unit 251 acquires the training data 141 from an external device or the like. The acquiring unit 251 stores the training data 141 in the storage unit 240. Note that the training data 141 may be stored in the storage unit 240 in advance.

[0091] The clustering unit 252 classifies the training data 141 into a plurality of clusters so as to maximize the amount of mutual information. Based on the classification results, the clustering unit 252 assigns a cluster label to each piece of denatured noise-added data included in the training data 141. The clustering unit 252 may perform clustering in advance, or may learn the relationship between the denatured noise-added data and the cluster labels during training, and assign a cluster label to each piece of denatured noise-added data based on the learning results.

[0092] The training processing unit 253 trains the autoencoder 40 using the training data 141. For example, based on the cluster labels of the denatured noise-added data of the training data 141, the training processing unit 253 selects a decoding unit from the decoding units 40b-1 to 40b-3 according to the cluster labels, and performs training.

[0093] Other operations performed by the training processing unit 253 are the same as those described in FIG. 6 for training the autoencoder 40.

[0094] The prediction processing unit 254 predicts the clean data x0 using the trained autoencoder 40. For example, the prediction processing unit 254 selects a decoding unit according to the cluster label of the denatured noise-added data and performs prediction of the clean data.

[0095] Other operations performed by the prediction processing unit 254 are similar to those of the above-described process for predicting clean data.

[0096] Next, an example of the processing procedure of the information processing apparatus 200 according to the second embodiment will be described. FIG. 8 is a flowchart showing the processing procedure during training according to the second embodiment. As shown in FIG. 8, the clustering unit 252 of the information processing apparatus 200 executes clustering on the training data 141 and assigns a cluster label to each modified noisy data (step S301).

[0097] The training processing unit 253 of the information processing apparatus 200 sets j = 1 (step S302). The training processing unit 253 sets i = 0 (step S303).

[0098] The training processing unit 253 selects a decoder corresponding to the cluster label of the modified noisy data y 0,j and inputs the modified noisy data y 0,j and the time change τ i into the autoencoder 40 (step S304). The training processing unit 253 updates the parameters of the autoencoder 40 so that f(y 0,j , 0) output from the autoencoder 40 approaches y τi,j (step S305).

[0099] The training processing unit 253 updates i by i = i + 1 (step S306). When the condition i < N is satisfied (step S307, Yes), the training processing unit 253 proceeds to step S304. On the other hand, when the condition i < N is not satisfied (step S307, No), the training processing unit 152 proceeds to step S308.

[0100] The training processing unit 253 updates j by j = j + 1 (step S308). When the condition j < M is satisfied (step S309, Yes), the training processing unit 253 proceeds to step S303. On the other hand, when the condition j < M is not satisfied (step S309, No), the training processing unit 253 outputs the trained autoencoder 40 (step S310).

[0101] Note that in the processing shown in FIG. 8, "y 0,j " and τi For one pair with, f(y 0,j , 0) was updated the parameters of the autoencoder 40 so as to approach y τi,j , but it is not limited to this. For example, the training processing unit 253 may update the parameters of the autoencoder 40 by applying the mini-batch learning method. That is, the training processing unit 253 is "y 0,j ", for m pairs with τ i , the parameters of the autoencoder 30 may be updated so that each of the m f(y 0,j , 0) and y τi,j approaches each other.

[0102] FIG. 9 is a flowchart showing the processing procedure at the time of prediction according to the second embodiment. Note that the autoencoder 40 described in FIG. 9 is a trained autoencoder 40. As shown in FIG. 9, the prediction processing unit 254 of the information processing apparatus 200 sets j = 1 (step S401).

[0103] The prediction processing unit 254 selects a decoding unit according to the cluster label of the modified noisy data y 0,j , and by inputting the modified noisy data "y 0,j " and the time change "0 (τ = 0)" into the autoencoder 40, f(y 0,j , 0) is calculated (step S402).

[0104] The prediction processing unit 254 updates j by j = j + 1 (step S403). When the prediction processing unit 254 satisfies the condition of j < M (step S404, Yes), it moves to step S402. On the other hand, when the prediction processing unit 254 does not satisfy the condition of j < M (step S404, No), it moves to step S405.

[0105] The prediction processing unit 254 predicts the clean data based on Equation (6) (step S405). The prediction processing unit 254 outputs the predicted clean data (step S406).

[0106] Next, the effects of the information processing device 200 according to the second embodiment will be described. The information processing device 200 classifies the training data 141 into a plurality of clusters, prepares the number of decoding units included in the autoencoder equal to the number of cluster labels, and creates a decoding unit specialized for each cluster. According to the information processing device 200, although the number of decoding units increases and the amount of calculation per epoch increases, since a decoding unit specialized for each cluster is created, it is possible to complete training in a shorter time than the time required for training by the information processing device 100 according to the first embodiment. [Example]

[0107] Next, an information processing device according to the third embodiment will be described. The information processing device according to the third embodiment will be referred to as "information processing device 300." Data handled as training data (and evaluation data) by the information processing device 300 will be referred to as "cryo-EM micrograph data." In the third embodiment, a set of training data and evaluation data will be referred to as a "task" as appropriate. The evaluation data is data used when evaluating the training results of a target (autoencoder), and is, for example, cryo-EM (Cryo-electron microscopy) micrograph data of time change (time) 0 (τ=0).

[0108] For example, cryo-EM micrograph data is widely used for imaging biomolecules, but measurements are performed under strict conditions with low doses to avoid radiation damage to the biomolecules being sampled. As a result, the signal-to-noise ratio of each acquired "cryo-EM micrograph data" is poor, which can significantly reduce the accuracy of parameter estimation in subsequent analysis, such as 3D reconstruction.

[0109] For example, the information processing device 300 uses an external device such as an electron microscope (cryo-EM microscope) to perform repeated low-dose measurements of the same field of view of the same sample for a longer period of time than usual. The signal-to-noise ratio of the cumulative image of the time-series data obtained by such measurements is higher than usual, but the data (long-exposure cryo-EM micrograph data) is affected by radiation damage towards the latter half of the measurement.

[0110] By applying the method described in Example 1 (D2N2 method) to long-exposure cryoEM micrograph data, clean data with minimal radiation damage can be predicted. Using long-exposure cryoEM micrograph data and the D2N2 method, high-quality cryoEM micrograph data with minimal radiation damage and a high signal-to-noise ratio, which is difficult to obtain using conventional methods, can be obtained. Using this high-quality cryoEM micrograph data can significantly improve the accuracy of parameter estimation in subsequent analysis, thereby improving the performance of biomolecular imaging.

[0111] However, if we try to apply cryoEM micrograph data directly to the method of Example 1 (D2N2 method), the following problem arises. Because the particles in cryoEM micrograph data deteriorate rapidly over time, it is difficult to ensure a sufficient length of time-series data recording the time course of the same particle. This corresponds to the fact that N of the training data in Example 1 is small. For example, if there is insufficient task (training data), it may not be possible to generate an autoencoder with sufficient noise removal capabilities.

[0112] In relation to the above-mentioned problem, Example 3 focuses on the following point. For example, because the field of view of the camera capturing cryo-EM micrograph data is wide, it is possible to simultaneously capture multiple particles in the same environment, and there are multiple other particles whose time changes and noise processes are strongly correlated. Furthermore, it is assumed that there are M other tasks with properties similar to the noise removal task for the target particle.

[0113] Based on the above-mentioned focus, the information processing device 300 trains an autoencoder by M-Task-N-shot transfer learning or meta-learning on M pieces of time-series data of length N acquired under a similar environment.

[0114] An example of a process in which the information processing device 300 creates multiple tasks will be described. FIG. 10 is a diagram for explaining the process of creating multiple tasks. In the following description, cryo-EM micrograph data will be abbreviated as "cryo data." The information processing device 300 acquires time-series cryo-data (training data 341) from an external device such as an electron microscope.

[0115] In the example shown in FIG. 10, the time series cryo data (training data 341) includes cryo data 1-0 with time change 0, cryo data 1-1 with time change τ1, cryo data 1-2 with time change τ2, and cryo data 1-3 with time change τ N The black circles in cryo data 1-0 to 1-N indicate particles.

[0116] The information processing device 300 divides the cryo data 1-0 to 1-N into a plurality of regions. Fig. 10 shows an example in which the cryo data 1-0 to 1-N are divided into 3 rows and 3 columns, but this is not limiting, and the regions may be divided for each particle. In this Example 3, an image of the region in the Xth row and Yth column among the divided regions of the cryo data is represented as a partial image (X, Y).

[0117] The information processing device 300 generates the partial image (1,1) of the cryo data 1-0, the partial image (1,1) of the cryo data 1-1, the partial image (1,1) of the cryo data 1-2, ..., the partial image (1,1) of the cryo data 1-N as one task. The information processing device 300 may set the partial image (1,1) of the cryo data 1-0 as evaluation data for the task.

[0118] The information processing device 300 generates multiple tasks by grouping together partial images at the same position using other partial images of the cryo data 1-0 to 1-N. For example, partial image (1,2) of the cryo data 1-0 to 1-N is grouped together as one task. The same applies to partial images (1,3) to (3,3) of the cryo data 1-0 to 1-N. As a result, nine tasks are created from the time-series cryo data (training data 341).

[0119] The information processing device 300 also creates multiple tasks for cryo data of other time series (times) acquired from an external device by executing the process described in Fig. 10. Note that the time series cryo data acquired from the external device is provided with information about the imaging environment. The information processing device 300 performs a process of classifying tasks generated from cryo data of the same imaging environment into the same group.

[0120] The information processing device 300 trains an autoencoder using a plurality of tasks created by the process described with reference to Fig. 10. An example of the process in which the information processing device 300 trains an autoencoder will be described below.

[0121] 11 is a diagram illustrating a process of training an autoencoder according to the third embodiment. For example, the information processing device 100 described in the first embodiment sets random initial parameters as the initial parameters of the autoencoder 30. In contrast, the information processing device 300 performs pre-learning using multiple tasks classified into the same group, and uses the parameters obtained by the pre-learning as the initial parameters. The multiple tasks classified into the same group are each created from cryodata captured under the same environment.

[0122] 11, the other tasks that belong to the same group as task Ts1 are tasks Ts1-1, Ts1-2, and Ts1-3. The information processing device 300 pre-trains the autoencoder 30 using tasks Ts1, and Ts1-1 to Ts1-3.

[0123] When the pre-learning is completed, the information processing device 300 uses the parameters of the autoencoder 30 at the time of the completion of the pre-learning as initial parameters for training using the task Ts1. The information processing device 300 trains the autoencoder 30 using the task Ts1.

[0124] The details of the process in which the information processing device 300 pre-learns (trains) the autoencoder 30 using the tasks Ts1, Ts1-1 to Ts1-3 are the same as the details of the training described in the first embodiment.

[0125] 11 corresponds to transfer learning (for example, joint training). The information processing device 300 may train the autoencoder 30 by meta-learning, which will be described below.

[0126] For example, the information processing device 300 may classify multiple tasks classified into the same group into evaluation data (partial images where τ=0) and other training data, and perform training using common initial parameters that minimize the loss of each evaluation data due to learned parameters for each task learned with the loss of each training data. When calculating such initial parameters, the information processing device 300 uses an algorithm such as MAML (Model-Agnostic Meta-Learning [Finn et al., 2017]).

[0127] The above describes an example of the process of training the autoencoder 30 executed by the information processing device 300. As described above, the information processing device 300 performs pre-learning using multiple tasks created from cryodata captured under the same environment, and then trains the autoencoder 30 using the parameters obtained as a result of the pre-learning as initial parameters. This makes it possible to create an autoencoder 30 that can predict clean data with high accuracy, even when it is difficult to ensure a sufficient sequence length for time-series data recording the time course of the same particle.

[0128] Next, an example of a process for predicting clean data (target molecules) executed by the information processing device 300 will be described. The information processing device 300 calculates a partial image containing a target particle, which has a time change of 0, by dividing the partial image into "y 0,j The information processing device 300 uses "y 0,j" and the time change "0 (τ=0)" are input to the trained autoencoder 30, and f(y 0,j The information processing device 300 repeats the above process for j=1 to M, and calculates M f(y 0,j ,0) is predicted as the clean data x0.

[0129] An example of the process of predicting clean data executed by the information processing device 300 has been described above.

[0130] Next, a description will be given of an example of the configuration of the information processing device 300. Fig. 12 is a functional block diagram showing the configuration of the information processing device according to the third embodiment. As shown in Fig. 12, the information processing device 300 includes a communication unit 310, an input unit 320, a display unit 330, a storage unit 340, and a control unit 350.

[0131] The communication unit 310 executes data communication with an external device (electron microscope) or the like via a network. The communication unit 310 is an NIC or the like. For example, the communication unit 310 may acquire training data 341 or the like from the external device or the like.

[0132] The input unit 320 is an input device that inputs various types of information to the control unit 350 of the information processing device 300. For example, the input unit 320 corresponds to a keyboard, a mouse, a touch panel, or the like.

[0133] The display unit 330 is a display device that displays information output from the control unit 350.

[0134] The storage unit 340 includes the autoencoder 30, training data 341, and a task table 342. The storage unit 340 is a memory or the like.

[0135] The autoencoder 30 is the autoencoder 30 described in Fig. 2. The autoencoder 30 is a neural network or the like.

[0136] The training data 341 is the training data 341 described in FIG. 10. The training data 341 is time-series cryodata. The storage unit 340 may store multiple training data captured under the same environment. Each training data is provided with information about the imaging environment when the time-series cryodata was captured.

[0137] The task table 342 is a table that holds a plurality of tasks created from the training data 341 described in Fig. 10. For example, in the task table 342, tasks created from cryodata captured under the same environment are classified into the same group.

[0138] Next, we move on to explaining the control unit 350. The control unit 350 has an acquisition unit 351, a task creation unit 352, a training processing unit 353, and a prediction processing unit 354. The control unit 350 is a CPU, a GPU, or the like.

[0139] The acquiring unit 351 acquires training data 341 from an external device or the like. The acquiring unit 351 stores the training data 341 in the storage unit 340. Note that the training data 341 may be stored in the storage unit 340 in advance.

[0140] The task creating unit 352 creates a plurality of tasks based on the training data 341. The process by which the task creating unit 352 creates a plurality of tasks is the same as the process described with reference to FIG.

[0141] The task creation unit 352 stores the created tasks in the task table 342. The task creation unit 352 stores the tasks created from cryodata captured under the same environment in the task table 342 in a state where the tasks are classified into the same group.

[0142] The training processing unit 353 performs pre-learning of the autoencoder 30 using multiple tasks classified into the same group stored in the task table 342, and uses the parameters obtained by the pre-learning as initial values ​​of the autoencoder 30. After the pre-learning, the training processing unit 353 trains the autoencoder 30 using a task of an image (partial image) containing a target particle.

[0143] Other operations performed by the training processing unit 353 are the same as those described in FIG. 11 for training the autoencoder 30.

[0144] The prediction processing unit 354 calculates a partial image "y 0,j " and the time change "0 (τ=0)" are input to the trained autoencoder 30, and f(y 0,j The prediction processing unit 354 repeats the above process for j=1 to M, and calculates M f(y 0,j ,0) is predicted as the clean data x0.

[0145] Next, an example of a processing procedure of the information processing device 300 according to the present embodiment 3 will be described. Fig. 13 is a flowchart showing a processing procedure during training according to the present embodiment 3. As shown in Fig. 13, the task creation unit 352 of the information processing device 300 acquires training data 341 (step S501).

[0146] The task creating unit 352 creates a plurality of tasks based on the training data 341 (step S502). The task creating unit 352 classifies the plurality of tasks created from the training data 341 captured under the same environment into the same group (step S503).

[0147] The training processing unit 353 performs pre-learning of the autoencoder 30 using a plurality of tasks classified into the same group, and identifies initial parameters (step S504).

[0148] The training processing unit 353 uses the initial parameters obtained by pre-learning to train the autoencoder 30 (step S505). The training processing unit 353 outputs the trained autoencoder 30 (step S506).

[0149] The processing procedure for prediction according to the third embodiment is the same as that described in FIG. 5 of the first embodiment except for the target image, and therefore will not be illustrated.

[0150] Next, the effects of the information processing device 300 according to the third embodiment will be described. The information processing device 300 performs pre-learning using multiple tasks created from cryodata captured under the same environment, and trains the autoencoder 30 using parameters obtained as a result of the pre-learning as initial values. This makes it possible to create an autoencoder 30 that can predict clean data with high accuracy, even when it is difficult to ensure a sufficient sequence length for time-series data recording the time course of the same particle.

[0151] In addition, when the information processing device 300 performs pre-training of the autoencoder 30 for a target particle, it may calculate the similarity between the task of the target particle and other tasks, and select tasks with high similarity as a priority to perform pre-training of the autoencoder 30.

[0152] 10, the information processing device 300 creates multiple tasks, but the present invention is not limited to this. For example, the information processing device 300 may prepare multiple series of whole cryo images (training data 341) and use each of the multiple series of training data as a task.

[0153] Next, an example of the hardware configuration of a computer that realizes the same functions as the above-described information processing device 100 (200, 300) will be described. Fig. 14 is a diagram showing an example of the hardware configuration of a computer that realizes the same functions as the information processing device of this embodiment.

[0154] 14, the computer 400 has a CPU 401 that executes various types of arithmetic processing, an input device 402 that accepts data input from a user, and a display 403. The computer 400 also has a communication device 404 that exchanges data with external devices via a wired or wireless network, and an interface device 405. The computer 400 also has a RAM 406 that temporarily stores various types of information, and a hard disk drive 407. The devices 401 to 407 are connected to a bus 408.

[0155] The hard disk drive 407 stores an acquisition program 407a, a clustering program 407b, a task creation program 407c, a training processing program 407d, and a prediction processing program 407e. The CPU 401 reads out each of the programs 407a to 407e and loads them into the RAM 406.

[0156] The acquisition program 407a functions as an acquisition process 406a. The clustering program 407b functions as a clustering process 406b. The task creation program 407c functions as a task creation process 406c. The training process program 407d functions as a training process 406d. The prediction process program 407e functions as a prediction process 406e.

[0157] The processing of the acquisition process 406a corresponds to the processing of the acquisition units 151, 252, and 352. The processing of the clustering process 406b corresponds to the processing of the clustering unit 252. The processing of the task creation process 406c corresponds to the processing of the task creation unit 353. The processing of the training processing process 406d corresponds to the processing of the training processing units 152, 253, and 353. The processing of the prediction processing process 406e corresponds to the processing of the prediction processing units 153, 254, and 354.

[0158] It should be noted that each of the programs 407a to 407d does not necessarily have to be stored in the hard disk drive 407 from the beginning. For example, each of the programs may be stored in a "portable physical medium" such as a flexible disk (FD), CD-ROM, DVD, magneto-optical disk, or IC card that is inserted into the computer 400. Then, the computer 400 may read and execute each of the programs 407a to 407e.

[0159] The following supplementary notes are further disclosed regarding the embodiments including the above examples.

[0160] (Supplementary Note 1) A plurality of data items including changes over time and including noise are used as training data, and an autoencoder is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time. When the data at the reference timing and information indicating the reference timing as information on the passage of time are input to the autoencoder, the output is predicted as noise-free data corresponding to the input data at the reference timing. A prediction program that causes a computer to execute a process.

[0161] (Appendix 2) The prediction program according to Appendix 1, characterized in that the prediction process inputs, for each index, a pair of data on the reference timing and information indicating the reference timing as information on the passage of time into the autoencoder, and predicts the average value of the output of the autoencoder as the noise-free data.

[0162] (Appendix 3) A plurality of data including changes over time and noise is acquired as training data, training the autoencoder so that an output when data on a reference timing included in the training data and information on the passage of time are inputted approaches data on a timing corresponding to the information on the passage of time; A training program that causes a computer to execute a process.

[0163] (Appendix 4) The training program described in Appendix 3, characterized in that the computer is further caused to execute a process of adding noise to multiple data included in the training data, calculating a first average value of the multiple data with the added noise, and training the autoencoder so that a second average value of the output when multiple pairs of data with the added noise at the reference timing and information on the passage of time are input to the autoencoder approaches the first average value.

[0164] (Supplementary Note 5) The autoencoder has a plurality of decoders; The training program described in Appendix 3 is characterized in that the computer is further made to execute a process of classifying the training data into a plurality of groups, and training the autoencoder so that when the data of the reference timing belonging to a certain group and information on the passage of time are input, the output from a decoder corresponding to the certain group among the plurality of decoders approaches the data of the timing corresponding to the information on the passage of time.

[0165] (Appendix 6) The training program described in Appendix 3, characterized in that it further causes a computer to execute a process of creating a plurality of tasks based on the training data, pre-training the autoencoder using a first task and a plurality of tasks similar to the first task among the plurality of tasks, thereby identifying initial parameters of the autoencoder, and training the autoencoder using the initial parameters and the first task.

[0166] (Supplementary Note 7) A plurality of data items including changes over time and including noise are used as training data, and an autoencoder is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time. When the data at the reference timing and information indicating the reference timing as information on the passage of time are input to the autoencoder, the output is predicted as noise-free data corresponding to the input data at the reference timing. A prediction method characterized in that the processing is performed by a computer.

[0167] (Appendix 8) The prediction method described in Appendix 7, characterized in that the prediction process inputs, for each index, a pair of data on the reference timing and information indicating the reference timing as information on the passage of time into the autoencoder, and predicts the average value of the output of the autoencoder as the noise-free data.

[0168] (Appendix 9) A plurality of data including changes over time and noise is acquired as training data, training the autoencoder so that an output when data on a reference timing included in the training data and information on the passage of time are inputted approaches data on a timing corresponding to the information on the passage of time; A training method characterized in that the processing is performed by a computer.

[0169] (Appendix 10) The training method described in Appendix 9, characterized in that the computer is further made to execute a process of adding noise to multiple data included in the training data, calculating a first average value of the multiple data with added noise, and training the autoencoder so that a second average value of the output when multiple pairs of data with added noise at the reference timing and information on the passage of time are input to the autoencoder approaches the first average value.

[0170] (Supplementary Note 11) The autoencoder has a plurality of decoders; The training method described in Appendix 9, characterized in that the computer further executes a process of classifying the training data into a plurality of groups, and training the autoencoder so that the output when the data of the reference timing belonging to a certain group and information on the passage of time are inputted, from one of the plurality of decoders corresponding to the certain group, approaches the data of the timing corresponding to the information on the passage of time.

[0171] (Appendix 12) The training method described in Appendix 9, characterized in that the computer further executes a process of creating a plurality of tasks based on the training data, pre-training the autoencoder using a first task and a plurality of tasks similar to the first task among the plurality of tasks, thereby identifying initial parameters of the autoencoder, and training the autoencoder using the initial parameters and the first task.

[0172] (Supplementary Note 13) A plurality of data including changes over time and including noise are used as training data, and an autoencoder is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time. When the data at the reference timing and information indicating the reference timing as information on the passage of time are input to the autoencoder, the output is predicted as noise-free data corresponding to the input data at the reference timing. An information processing device having a control unit that executes processing.

[0173] (Appendix 14) The information processing device described in Appendix 13, characterized in that the prediction process inputs, for each index, a pair of data on the reference timing and information indicating the reference timing as information on the passage of time into the autoencoder, and predicts the average value of the output of the autoencoder as the noise-free data.

[0174] (Appendix 15) A plurality of data including changes over time and noise is acquired as training data, training the autoencoder so that an output when data on a reference timing included in the training data and information on the passage of time are inputted approaches data on a timing corresponding to the information on the passage of time; An information processing device having a control unit that executes processing.

[0175] (Appendix 16) The information processing device described in Appendix 15, characterized in that the control unit further adds noise to multiple data included in the training data, calculates a first average value of the multiple data with the added noise, and further executes a process of training the autoencoder so that a second average value of the output when multiple pairs of data with the added noise at the reference timing and information on the passage of time are input to the autoencoder approaches the first average value.

[0176] (Supplementary Note 17) The autoencoder has a plurality of decoders; The information processing device described in Appendix 15, characterized in that the control unit further executes a process of classifying the training data into a plurality of groups, and training the autoencoder so that the output from a decoder corresponding to the group among the plurality of decoders approaches the data at the timing corresponding to the information on the passage of time, when the data at the reference timing belonging to a certain group and information on the passage of time are input.

[0177] (Appendix 18) The information processing device described in Appendix 15, characterized in that the control unit creates a plurality of tasks based on the training data, and pre-trains the autoencoder using a first task and a plurality of tasks similar to the first task among the plurality of tasks, thereby identifying initial parameters of the autoencoder, and further executes a process of training the autoencoder using the initial parameters and the first task. [Explanation of symbols]

[0178] 30,40 autoencoder 100, 200, 300 Information processing equipment 110,210,310 Communications Department 120,220,320 Input section 130,230,330 Display section 140,240,340 storage section 141,341 training data 150,250,350 Control unit 151,251,351 Acquisition Department 152,253,353 Training Processing Unit 153,254,354 Prediction processing section

Claims

1. A plurality of data items including changes over time and including noise are used as training data, and an autoencoder is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time. When the data at the reference timing and information indicating the reference timing as information on the passage of time are input to the autoencoder, the output is predicted as noise-free data corresponding to the input data at the reference timing. A prediction program that causes a computer to execute a process.

2. The prediction program according to claim 1, characterized in that the prediction process inputs, for each index, a pair of data on the reference timing and information indicating the reference timing as information on the passage of time into the autoencoder, and predicts the average value of the output of the autoencoder as the noise-free data.

3. A plurality of pieces of data that contain time-dependent changes and noise are acquired as training data; training the autoencoder so that an output when data on a reference timing included in the training data and information on the passage of time are inputted approaches data on a timing corresponding to the information on the passage of time; A training program that causes a computer to execute a process.

4. The training program of claim 3 further comprises causing a computer to execute a process of adding noise to a plurality of data included in the training data, calculating a first average value of the plurality of data to which noise has been added, and training the autoencoder so that a second average value of the output when a plurality of sets of data to which noise has been added at the reference timing and information on the passage of time are input to the autoencoder approaches the first average value.

5. the autoencoder comprises a plurality of decoders; The training program of claim 3 further comprises causing a computer to execute a process of classifying the training data into a plurality of groups, training the autoencoder so that when the data of the reference timing belonging to a certain group and information on the passage of time are input, the output from a decoder corresponding to the certain group among the plurality of decoders approaches the data of the timing corresponding to the information on the passage of time.

6. The training program of claim 3 further causes a computer to execute a process of creating a plurality of tasks based on the training data, pre-training the autoencoder using a first task and a plurality of tasks similar to the first task among the plurality of tasks, thereby identifying initial parameters of the autoencoder, and training the autoencoder using the initial parameters and the first task.

7. A plurality of data items including changes over time and including noise are used as training data, and an autoencoder is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time. When the data at the reference timing and information indicating the reference timing as information on the passage of time are input to the autoencoder, the output is predicted as noise-free data corresponding to the input data at the reference timing. A prediction method characterized in that the processing is performed by a computer.

8. A plurality of pieces of data that contain time-dependent changes and noise are acquired as training data; training the autoencoder so that an output when data on a reference timing included in the training data and information on the passage of time are inputted approaches data on a timing corresponding to the information on the passage of time; A training method characterized in that the processing is performed by a computer.

9. A plurality of data items including changes over time and including noise are used as training data, and an autoencoder is trained so that when data at a reference timing and information on the passage of time are input, the output approaches data at a timing corresponding to the information on the passage of time. When the data at the reference timing and information indicating the reference timing as information on the passage of time are input to the autoencoder, the output is predicted as noise-free data corresponding to the input data at the reference timing. An information processing device having a control unit that executes processing.

10. A plurality of pieces of data that contain time-dependent changes and noise are acquired as training data; training the autoencoder so that an output when data on a reference timing included in the training data and information on the passage of time are inputted approaches data on a timing corresponding to the information on the passage of time; An information processing device having a control unit that executes processing.