Information processing device, information processing method, and information processing system

Abstract

Detect trends in data with missing information. [Solution] The information processing device comprises a processing circuit. The processing circuit acquires second data by interpolating the missing values ​​of data with a first interpolation value when the first data is missing, and trains an autoencoder model that outputs the first data when the second data is input. It acquires a plurality of fourth data by interpolating the missing values ​​of data with a plurality of second interpolation values ​​when the third data is missing, calculates the error between the third data and the output data obtained by inputting the plurality of fourth data to the autoencoder model, updates the first interpolation value with an interpolation value extracted from the plurality of second interpolation values ​​based on the error, and repeatedly executes the process from acquiring the second data using the updated first interpolation value to optimize the first interpolation value and the autoencoder model.

Description

Technical Field

[0001] Embodiments of the present invention relate to an information processing apparatus, an information processing method, and an information processing system.

Background Art

[0002] In manufacturing, selling, etc. products, it is important to highly and stably improve the yield in the manufacturing process. Appropriately obtaining information such as locations and areas where defects have occurred throughout the product often leads to stabilizing the yield. For example, when manufacturing a large number of parts together, by determining under what circumstances the parts are appropriately manufactured when viewed as a whole, the final yield can be maintained at a high level.

[0003] In such a case where there are a large number of parts as a whole, in the inspection process, the overall tendency may be obtained by a sensor or the like in the form of an image or the like, but due to the performance of the sensor or the like, there may be a defect in the image or the like. Since the interpolation of this missing value may affect the overall tendency, it is desirable to perform it appropriately.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Therefore, one of the non-limiting problems that the embodiment aims to solve is to appropriately obtain the overall trend even when there are defects. The problems that the embodiment aims to solve may also be, in some more limited examples, problems corresponding to the effects described in the following description. That is, any problem corresponding to at least one of the effects described in the description of the embodiment may be a problem that the embodiment aims to solve. [Means for solving the problem]

[0006] According to one embodiment, the information processing device includes a processing circuit. The processing circuit acquires second data by interpolating the missing values ​​of data with a first interpolation value when the first data is missing, and trains an autoencoder model that outputs the first data when the second data is input. It acquires a plurality of fourth data by interpolating the missing values ​​of data with a plurality of second interpolation values ​​when the third data is missing, calculates the error between the third data and the output data obtained by inputting the plurality of fourth data to the autoencoder model, updates the first interpolation value with an interpolation value extracted from the plurality of second interpolation values ​​based on the error, and repeatedly executes the processing from the acquisition of the second data using the updated first interpolation value to optimize the first interpolation value and the autoencoder model. [Brief explanation of the drawing]

[0007] [Figure 1] A diagram showing an example of the processing flow according to one embodiment. [Figure 2] A figure showing an example of missing data generation according to one embodiment. [Figure 3] A figure showing an example of interpolation data generation according to one embodiment. [Figure 4] A conceptual diagram showing an example of an autoencoder model according to one embodiment. [Figure 5] A conceptual diagram showing an example of training a model for estimating a state according to one embodiment. [Figure 6] A conceptual diagram showing an example of training a model for estimating a state according to one embodiment. [Figure 7] A conceptual diagram showing an example of training a model for estimating a state according to one embodiment. [Figure 8] A conceptual diagram showing an example of training a model for estimating a state according to one embodiment. [Figure 9] A conceptual diagram showing an example of training a model for estimating a state according to one embodiment. [Modes for carrying out the invention]

[0008] Embodiments will be described below with reference to the drawings. The information processing apparatus, information processing method, and information processing system in this disclosure realize information processing that appropriately estimates a certain trend from data containing missing values, even when such data contains missing values, in data where a certain trend can be observed when viewed as a whole.

[0009] The data may be, for example, image data, audio data, other binary data, or text data. While a semiconductor device may be used as an example, it should be noted that the forms disclosed herein are applicable to a variety of other fields.

[0010] One of the information processing devices in this disclosure operates as an estimation device that estimates the overall trend related to data from data that contains missing values. Another of the information processing devices in this disclosure operates as a learning device that optimizes the model used for estimation in the above information processing device.

[0011] Furthermore, one of the information processing devices in this disclosure may be an information processing device that implements the operation of both the learning device and the estimation device described above. In addition, the above information processing can be implemented not by a single information processing device, but by an information processing system in which multiple information processing devices cooperate and operate at least at some point in time.

[0012] (First Embodiment)

[0013] First, an embodiment of an information processing apparatus as a training apparatus will be described. The information processing apparatus includes, for example, a memory circuit and a processing circuit. The processing circuit is formed as a circuit for training. The memory circuit stores necessary data.

[0014] The processing circuit may be in a form that uses a dedicated digital or analog circuit such as an ASIC (Application Specific Integrated Circuitry) or a DSP (Digital Signal Processor), or may be in a form that uses a general-purpose processor such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), or may be in a form that uses a programmable circuit such as an FPGA (Field Programmable Gate Array).

[0015] The memory circuit may be provided inside the information processing apparatus, or may be provided outside and be able to acquire data through an appropriate interface. The memory circuit may store data required in training, data during training, and data after training completion respectively. When the information processing of the information processing apparatus is specifically realized by using a processing circuit, which is a hardware resource, through software, the memory circuit may store a program, an executable file, or data equivalent thereto for controlling this software-based information processing.

[0016] As described above, these may be composed of several memory circuits and processing circuits connected to each other as an information processing system. In the following, the processing executed by the information processing apparatus can be realized by appropriately using this processing circuit and the memory circuit.

[0017] FIG. 1 is a diagram showing the flow of processing according to an embodiment. In this embodiment, the information processing apparatus generates a model that appropriately generates data without missing data from data with missing data.

[0018] The information processing apparatus first prepares a training data group and a verification data group (S100). These data are stored in an external storage circuit of the information processing apparatus, such as a storage or the like, and it is sufficient that they are prepared to be available for the information processing apparatus at the timing when the information processing apparatus uses them. The training data is data used for learning the autoencoder model AE. The verification data is data used for verifying the autoencoder model AE and is also data for optimizing the interpolation value.

[0019] The information processing apparatus extracts first data for executing training of the autoencoder model AE from the training data (S102). Subsequently, the information processing apparatus generates missing data by missing the first data (S104).

[0020] For example, when the input data (training data and verification data) is image data, the information processing apparatus can generate data by missing a part of the image. The information processing apparatus can similarly generate missing data in voice data, other binary data, and text data.

[0021] The information processing apparatus can generate missing data according to the characteristics of the sensor that acquires the input data. For example, when the input data is image data, the information processing apparatus can generate missing data by missing the image data based on the missing that can occur in the sensor that acquires the image.

[0022] FIG. 2 is a diagram showing an example of generating missing data for image data according to an embodiment. The information processing apparatus generates the missing data shown on the right side by missing the values of some or a plurality of regions for some image (for example, the first data) shown on the left side. The portions filled in black in the right figure indicate the missing portions.

[0023] Returning to Figure 1, the information processing device interpolates the missing parts of the generated missing data using the first interpolation value to generate the second data (S106). In the first iteration, a predetermined initial value may be set for the first interpolation value. By interpolating the missing data in this way, the information processing device generates data that can be input to the autoencoder model AE.

[0024] Figure 3 shows an example of interpolation data generation for image data according to one embodiment. The information processing device interpolates the missing parts of the missing data shown on the left using a first interpolation value. The first interpolation value is set, for example, between the minimum and maximum values ​​of brightness. The figure on the right shows the interpolated data (second data), and the shaded areas indicate the areas interpolated by the first interpolation value.

[0025] Returning to Figure 1, the information processing device inputs the second data to the autoencoder model AE and obtains output data for the second data (S108). The autoencoder model AE is a model that includes an encoder ENC and a decoder DEC, where the encoder ENC converts the input image into a latent variable Z (latent vector), and the decoder DEC converts this latent variable Z into output data.

[0026] The information processing device compares the output data obtained by inputting the second data into the autoencoder model AE with the first data, and trains the autoencoder model AE so that when the second data is input, it outputs the first data (S112). This training may be performed based on a general autoencoder method.

[0027] The information processing device continues training the autoencoder model AE until optimization is complete. During training, the information processing device may extract multiple first data points from the training data as needed and perform optimization. Alternatively, during training, the information processing device may generate multiple missing data points from the same first data point and generate second data points using first interpolated values ​​for these missing data points.

[0028] The learning termination conditions are set to ensure that the autoencoder model AE is properly trained. These termination conditions can include, for example, completing a predetermined number of iterations, performing calculations for a predetermined duration, or the evaluation function falling below a predetermined threshold.

[0029] Prior to this learning, the information processing device may, in an alternative configuration, first train the autoencoder model AE so that when the first data is input, the first data is output, without using missing data, and then perform learning using the second data using these trained parameters. In other words, the information processing device can use an autoencoder model AE that has been pre-trained so that when the first data is input, the first data is output. In this case, the processing in S112 can also be performed using techniques such as network distillation and dropout of the autoencoder model AE.

[0030] After the training using the training data is completed in the autoencoder model AE, the information processing device moves on to processing using the validation data. The information processing device extracts the third data from the validation data (S112). Subsequently, the information processing device generates missing data for the fourth data, similar to the first data (S114).

[0031] Next, the information processing device generates multiple fourth data using multiple second interpolation values ​​for the missing data in the acquired third data (S116). The second interpolation values ​​may or may not include the first interpolation values, but it is preferable that they include the first interpolation values ​​so that other second interpolation values ​​can be compared with the first interpolation values.

[0032] The information processing device can, for example, randomly extract multiple values ​​from the acceptable range of interpolated values ​​to obtain multiple second interpolated values. Alternatively, the information processing device can, for example, extract multiple values ​​at equal intervals within the range of interpolated values ​​to obtain multiple second interpolated values.

[0033] The information processing device inputs multiple fourth data points into the autoencoder model AE and acquires output data (S118). The information processing device then updates the first interpolation value by comparing the multiple output data points with the third data points (S120).

[0034] For example, the information processing device may acquire the error (residual) between multiple output data and a third data, extract a second interpolated value corresponding to the output data with the smallest error, and update the first interpolated value with the extracted second interpolated value.

[0035] After the first interpolation value is updated, the information processing device can repeatedly execute the processes from S102 to S120 until the termination condition is reached (S122). That is, the information processing device learns the autoencoder model AE using the new first interpolation value, and updates the first interpolation value using the second interpolation value for the learned autoencoder model AE.

[0036] In this way, the information processing device alternately repeats updating the autoencoder model AE and updating the first interpolation value that fills in missing values. Ultimately, the information processing device can repeat these processes until it satisfies conditions suitable for completing learning, such as until the error (which may be the minimum value or the average value, etc.) between the output data input to the autoencoder model AE with the third data falls below a predetermined threshold, or until the first interpolation value has been updated a predetermined number of times.

[0037] Furthermore, the information processing device may use other selection methods for the second interpolation value in the processing of S116, rather than just randomly and uniformly. For example, the selection may be made according to a normal distribution centered on the current first interpolation value.

[0038] This selection can also be changed depending on the number of times S120 has been executed. For example, the information processing device may control parameters such as the standard deviation of a normal distribution according to the number of times S120 has been executed.

[0039] As another example, the information processing device may change the method for selecting the second interpolation value to be selected in the next iteration based on the error between the output data and the third data. The information processing device can select the next second interpolation value according to the error in a way that avoids falling into a local minimum, for example, as seen in simulated annealing.

[0040] In the processing of S120, the information processing device may extract the second interpolation value that minimizes the error between the output data and the third data, and use this as the first interpolation value for the next iteration. However, the information processing device may also calculate the first interpolation value for the next iteration from the second interpolation value (and the first interpolation value) based on the error between the output data and the third data.

[0041] The information processing device may update the first interpolated value based on various statistics, such as the average value obtained by weighting each second interpolated value based on the error between the output data of the fourth data input to the autoencoder model AE and the third data.

[0042] As described above, according to this embodiment, by repeatedly updating the autoencoder model and updating the first interpolated value alternately, it is possible to optimize the model for estimating the original data from data with missing values. By using the original data thus reconstructed, it becomes possible to understand the appropriate overall data trend from the data with missing values. The overall data trend may be determined by rule-based determination, or by using a trained model that outputs the overall data trend.

[0043] (Second Embodiment)

[0044] In the first embodiment described above, an example of restoring data with missing information using an autoencoder model was explained. In this embodiment, a method for improving the accuracy of the restoration by this autoencoder model will be described.

[0045] Figure 4 is a conceptual diagram showing an example of an autoencoder model AE according to one embodiment. The autoencoder model AE includes, for example, an encoder ENC and a decoder DEC, in addition to a Gaussian layer GL.

[0046] The Gaussian layer GL is placed before the input layer of the encoder ENC and superimposes Gaussian noise on the data input to the encoder ENC. When the information processing device inputs the second and fourth data shown in Figure 1 to the autoencoder model AE, it first adds noise to the data using this Gaussian layer GL, and then inputs this noisy data to the encoder ENC.

[0047] A Gaussian layer (GL) can superimpose noise onto data not by continuously adding a predetermined noise, but by generating noise according to a distribution that follows, for example, the noise superposition parameters.

[0048] The information processing device can perform the optimization of the autoencoder model AE and the first interpolation value using the autoencoder model AE in Figure 4 instead of the autoencoder model AE in Figure 1.

[0049] As described above, this embodiment can include a Gaussian layer GL that superimposes noise onto the data input to the encoder ENC of the autoencoder model AE. By including this Gaussian layer GL, it becomes possible to train a more robust model that takes into account noise in the input data acquired by the sensor and variations due to individual differences in the sensor.

[0050] (Third embodiment)

[0051] In the embodiments described above, a model for reconstructing intact data from missing data was explained. By using this model, intact data can be obtained, and the overall trends of the data can be obtained using various methods, such as commonly known rule-based methods and methods using models trained by machine learning.

[0052] In this embodiment, we will describe an unspecified example of the overall trend of data using the autoencoder model AE described in each of the embodiments described above. In the following description, we will use an embodiment that includes the Gaussian layer GL described in the second embodiment, but of course, the same processing can be achieved even without the Gaussian layer GL shown in Figure 1, and the same effects as in the first embodiment can be obtained.

[0053] Figure 5 is a conceptual diagram illustrating the configuration of a model that acquires state Y (state-indicating data) from data with missing data X (characteristic data) according to one embodiment, and an example of training this model. The autoencoder model AE may be formed according to the embodiments described above.

[0054] The information processing device, using the same processing as in the previously described embodiment, causes data X to be missing, interpolates it, inputs the interpolated data into the autoencoder model AE, and obtains estimated reconstructed data f(X) of data X (S200). The interpolated value can be the first interpolated value optimized according to the previously described embodiment.

[0055] The information processing device optimizes the first model so that when it receives the acquired estimated reconstruction data f(X) and the data X as input, it outputs a state Y (S202). The state Y for data X can be generated from training data such as experimental values ​​or theoretical values.

[0056] The first model may be, for example, a linear regression model. In this case, the information processing device can optimize the model by setting the state Y as the dependent variable for the data X as the explanatory variable. The linear regression method may be any method.

[0057] Furthermore, as an example that is not limited to this, the first model may be any other model such as MLP (Multi-Layer Perceptron), CNN (Convolutional Neural Network), or GNN (Graph Neural Network), and in these cases, the information processing device can perform model training using any method suitable for each model.

[0058] The first model, thus trained, is optimized to obtain state Y1 (first state) from the results of inputting data X containing missing values ​​into the autoencoder model AE.

[0059] As described above, according to this embodiment, by using optimized interpolation values, an optimized autoencoder model AE, and an optimized first model, it becomes possible to estimate the overall trend of the data from data with missing values.

[0060] Furthermore, the information processing device can perform training of the autoencoder model AE and optimization of the first interpolation value in parallel with training the first model.

[0061] (Fourth Embodiment)

[0062] In the third embodiment, the state was predicted using the reconstructed data output from the autoencoder model AE. However, in this embodiment, we will describe a model that predicts the state using the latent variable Z output from the encoder ENC of the autoencoder model AE.

[0063] Figure 6 is a conceptual diagram showing the configuration of a model that acquires state Y from data with missing data X according to one embodiment, and an example of training this model. The autoencoder model AE may be formed according to the embodiments described above.

[0064] The information processing device decrypts the data X in the same manner as in the embodiments described above, interpolates it with an optimized first interpolation value, and inputs it to the optimized autoencoder model AE. The information processing device obtains the latent variable Z output from the encoder ENC of the autoencoder model AE (S300).

[0065] The information processing device trains a second model to obtain the state Y from the latent variable Z (S302). The second model may be formed, for example, by a decision tree. If it is formed by a decision tree, the information processing device can train the second model to estimate the state Y from the latent variable Z using a machine learning technique such as XGBoost.

[0066] The latent variable Z, which is dimensionally compressed data by the encoder ENC, is reconstructed into data without missing values ​​by the decoder DEC. This data without missing values ​​is data that has a form from which the state can be extracted. That is, the latent variable Z is a low-dimensional vector that has features of the state Y for the input data X.

[0067] The second model is a model formed to extract the features of state Y from this latent variable Z and output state Y. The information processing device is optimized to extract state Y from the latent variable Z which has the features of state Y in this way.

[0068] For example, the information processing device takes an optimized interpolated value from missing data X and a latent variable Z obtained by solving the autoencoder model AE as input, and performs machine learning (e.g., training with XGBoost) to output a state Y as training data, thereby optimizing the second model.

[0069] The optimized second model is formed as a model that takes a latent variable Z as input and outputs state Y2 (second state).

[0070] As described above, according to this embodiment, by using optimized interpolation values, an optimized autoencoder model AE, and an optimized second model, it becomes possible to estimate the overall trend of the data from data with missing values.

[0071] Furthermore, similar to the third embodiment, the information processing device can perform training of the autoencoder model AE and optimization of the first interpolation value as multitask learning in parallel with training of the second model.

[0072] (Fifth embodiment) In the third embodiment, the estimation of the state from data output from the autoencoder model AE by the first model was described, and in the fourth embodiment, the estimation of the state from latent variables extracted from the autoencoder model AE by the second model was described. These estimations may be performed separately as described above, but it is also possible to learn to estimate the state Y using the results of both.

[0073] Figure 7 is a conceptual diagram showing the configuration of a model that acquires state Y from data with missing data X according to one embodiment, and an example of training this model. The autoencoder model AE, the first model, and the second model have the same configuration as the embodiments described above.

[0074] The information processing device performs ensemble learning using the optimization results of the first and second models to train these models (S400). As an example, the information processing device can perform ensemble learning after optimizing the first interpolation value and the autoencoder model AE. Also, as an example, the information processing device can achieve training of the first interpolation value, the autoencoder model AE, the first model and the second model through ensemble learning.

[0075] As mentioned earlier, this training can also be conducted as multitasking learning.

[0076] More specifically, the information processing device trains the first and second models according to the processing of the third and fourth embodiments, for example, using the first interpolated value and autoencoder model AE learned from the training data. The information processing device can acquire data (X, Y), i.e., input data X and state data Y, from the same or different training data for these models, and perform learning based on this data (X, Y).

[0077] By performing ensemble learning using the state Y1 output from the first model and the state Y2 output from the second model, the information processing device can achieve model training that suppresses overfitting in both the first and second models.

[0078] As described above, this embodiment makes it possible to further improve the accuracy of the optimized interpolated values, the optimized autoencoder model AE, the optimized first model, and the optimized second model. As a result, it becomes possible to estimate the overall trend of the data from data with missing values ​​with higher accuracy.

[0079] (Sixth Embodiment)

[0080] In addition to using the autoencoder model AE described above, it is also possible to form a model that directly obtains the state Y from missing data. This model can be formed, for example, by a model such as XGBoost that estimates the state from missing data.

[0081] Figure 8 conceptually illustrates a third model that directly estimates state Y from missing data according to one embodiment. The information processing device acquires missing data from data X and trains the third model to output state Y when this missing data is input (S500). This training is implemented by performing learning that supports missing data, such as XGBoost.

[0082] The third model differs from the second model in that it directly applies a model such as XGBoost to missing data. It is also possible to train using this third model as a weak learner. Therefore, in this embodiment, a further training method using this third model as a weak learner will be described.

[0083] Figure 9 is a conceptual diagram showing the configuration of a model that acquires state Y from missing data X according to one embodiment, and an example of training this model. The autoencoder model AE, the first model and the second model have the same configuration as the embodiments described above. The third model has the same configuration as the model shown in Figure 8.

[0084] The information processing device first generates missing data from data X by the same process as in each of the embodiments described above. The information processing device generates interpolated data for the missing data using a first interpolation value.

[0085] The information processing device inputs the stored data into the autoencoder model AE and obtains the output data f(X) and the latent variable Z. The information processing device inputs the output data f(X) into the first model to obtain state Y1. Simultaneously, the information processing device inputs the latent variable Z into the second model to obtain state Y2. Furthermore, the information processing device inputs the missing data (data X) into the third model to obtain state Y3 (third state).

[0086] The information processing device performs ensemble learning using these states Y1, Y2, and Y3 to improve the accuracy of each model.

[0087] As an example that is not limited to this, the information processing device may learn the first model, the second model, and the third model through ensemble learning.

[0088] As an example that is not limited to this, the information processing device may learn the first interpolated value, the autoencoder model AE, the first model, the second model, and the third model by ensemble learning.

[0089] By performing ensemble learning that references output values ​​from multiple paths that produce the same state, the accuracy of each model can be improved. Furthermore, this learning can be multi-task learning as needed.

[0090] As described above, this embodiment makes it possible to optimize not only the autoencoder model AE, but also the model that acquires the overall trend of the data. In other words, the information processing device can generate a highly accurate model by ensemble learning using at least two of the first, second, and third states acquired by separate means.

[0091] An information processing device can use an optimized model to form a model for restoring missing data and / or a model for estimating a state that indicates an overall trend in the data. The scope of this disclosure, of course, also includes forms in which the information processing device uses this optimized model.

[0092] Furthermore, an information processing system can be formed that uses multiple information processing devices to recover data with missing information and / or estimate the overall state from the data with missing information.

[0093] As an example without limitation, the optimizations and optimized models in this disclosure can be applied to the characteristics and wafer conditions of semiconductor devices. For example, data X (characteristic) may be an electrical characteristic, which may be, as an example without limitation, a leakage current, a breakdown voltage, or an on-resistance.

[0094] For example, state Y may represent the electronic state of a device predicted from its electrical properties, or the state of a semiconductor device. As an example, state Y can be used to estimate impurity concentration, device processing shape (trench width, depth, taper angle, etc.), and also to estimate photolithography accuracy and processing accuracy that affect the shape. Furthermore, state Y can be used as an electrical property to estimate other electrical properties.

[0095] Data with missing values ​​may include, for example, data exhibiting characteristics below a certain threshold. For instance, these may include missing data such as dust particles or other particles in an image, or voltage resistance values. Missing data can also include defects. A defect may include, for example, a case where, if a certain characteristic is judged to be defective, subsequent characteristics are not measured. In this way, it is possible to handle missing values ​​of any characteristic, such as missing data in images, current values, and voltage values.

[0096] While several embodiments of the present invention have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These novel embodiments can be carried out in a variety of other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents.

Claims

1. Equipped with a processing circuit, The aforementioned processing circuit is The missing values ​​in the first data set are interpolated using the first interpolated value to obtain the second data set. An autoencoder model is trained that outputs the first data when the second data is input. The missing values ​​in the data with missing third data are interpolated using multiple second interpolation values ​​to obtain multiple fourth data. The error between the third data and the output data obtained by inputting the plurality of fourth data into the autoencoder model is calculated. Based on the aforementioned error, the first interpolation value is updated using the interpolation value extracted from the plurality of second interpolation values. Using the updated first interpolation value, the process from acquiring the second data is repeatedly executed to optimize the first interpolation value and the autoencoder model. Information processing device.

2. The aforementioned autoencoder model includes a layer that superimposes Gaussian noise onto the input data. The information processing apparatus according to claim 1.

3. The aforementioned processing circuit is A first model is generated that performs linear regression on state data for input data containing missing values ​​and output data obtained by interpolating the input data with the first interpolation value and inputting it into the autoencoder model. The information processing apparatus according to claim 1.

4. The processing circuit further, Input data containing missing values ​​is interpolated with the first interpolation value and input to the autoencoder model to obtain latent variables. A second model is trained to obtain the state data from the latent variable, based on the latent variable and the state data for the input data. The information processing apparatus according to claim 3.

5. The aforementioned processing circuit is Input data having missing values ​​is interpolated with the first interpolation value and input to the autoencoder model and the first model to obtain the first state. The input data is interpolated with the first interpolation value and input to the autoencoder model and the second model to obtain the second state. Based on the first and second states, perform ensemble learning. The information processing apparatus according to claim 4.

6. The processing circuit further, A third model that outputs state data when input data containing missing values ​​is input is trained based on the state data obtained by inputting the input data and the state data for the input data. The information processing apparatus according to claim 4.

7. The aforementioned processing circuit is Input data having missing values ​​is interpolated with the first interpolation value and input to the autoencoder model and the first model to obtain the first state. The input data is interpolated with the first interpolation value and input to the autoencoder model and the second model to obtain the second state. The input data is then input to the third model to obtain the third state. Based on the first state, the second state, and the third state, ensemble learning is performed. The information processing apparatus according to claim 6.

8. Using a model trained by an information processing device according to any one of claims 1 to 7, the state of input data having missing values ​​is obtained from the input data. Information processing device.

9. The processing circuit, The missing values ​​in the first data set are interpolated using the first interpolated value to obtain the second data set. An autoencoder model is trained that outputs the first data when the second data is input. The missing values ​​in the data with missing third data are interpolated using multiple second interpolation values ​​to obtain multiple fourth data. The error between the third data and the output data obtained by inputting the plurality of fourth data into the autoencoder model is calculated. Based on the aforementioned error, the first interpolation value is updated using the interpolation value extracted from the plurality of second interpolation values. Using the updated first interpolation value, the process from acquiring the second data is repeatedly executed to optimize the first interpolation value and the autoencoder model. Information processing methods.

10. One or more memory circuits, One or more processing circuits, Equipped with, Using at least one of the one or more processing circuits described above, The missing values ​​in the first data set are interpolated using the first interpolated value to obtain the second data set. An autoencoder model is trained that outputs the first data when the second data is input. The missing values ​​in the data with missing third data are interpolated using multiple second interpolation values ​​to obtain multiple fourth data. The error between the third data and the output data obtained by inputting the plurality of fourth data into the autoencoder model is calculated. Based on the aforementioned error, the first interpolation value is updated using the interpolation value extracted from the plurality of second interpolation values. Using the updated first interpolation value, the process from acquiring the second data is repeatedly executed to optimize the first interpolation value and the autoencoder model. Information processing system.

Images