WO2025224959A1 - Information processing device, training method, and program - Google Patents

Abstract

An information processing device (1) comprises: a feature amount extraction unit (11) that extracts feature amounts from data; a clustering unit (12) that determines a cluster to which each feature amount belongs by clustering of the feature amounts in a feature amount space; a pseudo label assignment unit (13) that sequentially sets a range in each cluster, generates a pseudo label on the basis of cluster information corresponding to feature amounts included in the range, and assigns the generated pseudo label to at least one feature amount included in the range; an inference unit (14) that calculates an inference value on the basis of the feature amounts; and a parameter update unit (15) that updates parameters of the feature amount extraction unit (11) and the inference unit (14) on the basis of the pseudo label and the inference value.

Description

Information processing device, learning method and program

This disclosure relates to an information processing device, a learning method, and a program.

A technique is known in which features extracted from data are clustered, and the clusters obtained by the clustering are used as pseudo labels to train a classifier. For example, Non-Patent Document 1 describes a technique in which features extracted from image data are clustered, and the clusters obtained by the clustering are assigned as pseudo labels to each image.

Caron, Mathilde, et al. “Deep clustering for unsupervised learning of visual features.” Proceedings of the European conference on computer vision (ECCV).　2018.

The conventional technology described in Non-Patent Document 1 had the problem that if labels were not assigned to each feature point within a cluster, it was not possible to perform learning by assigning pseudo-labels to these points. For example, the conventional technology assumes that a pseudo-label is assigned to each image corresponding to a cluster, and therefore cannot be applied to image segmentation, which requires that a pseudo-label be assigned to each pixel within an image.

The present disclosure aims to solve the above problem and provide an information processing device capable of learning by assigning pseudo labels to each feature point of a cluster.

The information processing device according to the present disclosure includes a feature extraction unit that extracts features from data, a clustering unit that determines the cluster to which a feature belongs by clustering the features in feature space, a pseudo label assignment unit that sequentially sets ranges within the cluster, generates pseudo labels based on cluster information corresponding to the features included within the range, and assigns the generated pseudo label to at least one feature included within the range, an inference unit that calculates an inferred value based on the feature, and a parameter update unit that updates parameters of the feature extraction unit and the inference unit based on the pseudo label and the inferred value.

The information processing device disclosed herein enables learning in which pseudo labels are assigned to each feature point of a cluster.

1 is a block diagram showing an example of the configuration of an information processing device according to a first embodiment; 4 is a flowchart illustrating an operation of the information processing device according to the first embodiment. 3 is a flowchart showing a learning method according to the first embodiment. FIG. 1 is a diagram illustrating an overview of a process for extracting features from data. FIG. 1 is a diagram illustrating an overview of clustering of feature amounts. FIG. 10 is a diagram illustrating an overview of a process for assigning a pseudo label to each point in a cluster. FIG. 1 is a diagram illustrating an outline of inference using feature amounts. FIG. 10 is a diagram illustrating an overview of parameter updates in a feature extraction unit and an inference unit. 1 is a block diagram showing a hardware configuration for realizing the functions of an information processing device according to a first embodiment; FIG. 10 is a block diagram showing a configuration example of an information processing device according to a second embodiment. 10 is a flowchart showing a learning method according to the second embodiment. 12A, 12B, 12C, and 12D are diagrams showing examples of data having spatial continuity.

Embodiment 1.
In supervised learning, a correct label (hereinafter simply referred to as "label") is assigned to a dataset, whereas in unsupervised learning, unlabeled data is used. Here, a label is identification information used to identify the cluster into which the data constituting the dataset is classified. A cluster is a set of points with similar features. Features are expressed in the form of numbers or categories, etc., of each piece of data constituting the dataset. Points with features that share characteristics different from those of other clusters are classified within the same cluster.

The information processing device according to the first embodiment is a device that includes an inference unit, and that performs inference processing by learning the inference unit by updating the parameters of an inference model through unsupervised learning. Here, the unsupervised learning performed by the information processing device according to the first embodiment calculates pseudo labels corresponding to features extracted from unlabeled data, and assigns the calculated pseudo labels to each feature point within a cluster. Note that a pseudo label is a label that is artificially assigned to unlabeled data.

(Outline of information processing device)
1 is a block diagram showing an example configuration of an information processing device 1 according to embodiment 1. In FIG. 1, the information processing device 1 extracts features from data acquired as a dataset, clusters the features in a feature space, sequentially sets ranges within the clusters, generates pseudo labels based on cluster information corresponding to the features included in the ranges, and assigns the pseudo labels to at least one feature included in the ranges. For example, each point in the feature space represents a point at each time in time-series data of features, or a feature of each pixel of an image, etc.

Cluster information is information that indicates the relationship between points that indicate feature quantities and the clusters into which these points are classified, and is generated in the process of clustering features extracted from data. Cluster information includes information that indicates which cluster a point that indicates a feature quantity has been classified into, information that indicates the reference point of the cluster, or information that indicates the characteristics or attributes of the cluster. Information that indicates which cluster a point that indicates a feature quantity has been classified into includes, for example, information that indicates a cluster number used to identify the cluster. Examples of reference points for a cluster include the center point, center of gravity, average value, or median of the cluster. Information that indicates the characteristics or attributes of a cluster is information in which the characteristics or attributes have been quantified, and may be, for example, a score used to evaluate the characteristics of the cluster.

When the cluster information is a score for evaluating the characteristics of a cluster, the score is calculated based on the feature amount. For example, the information processing device 1 assigns the average value of the scores corresponding to the feature amounts included within a range set within the cluster as a pseudo label to the target points included within the range. The information processing device 1 then sets the above range sequentially and assigns a pseudo label to each point so that all points included in the cluster become target points. This allows the information processing device 1 to learn to assign pseudo labels to all feature amount points included in the feature amount space.

As shown in FIG. 1, the information processing device 1 also includes a feature extraction unit 11, a clustering unit 12, a pseudo label assignment unit 13, an inference unit 14, and a parameter update unit 15. For example, the information processing device 1 is implemented by a computer. A memory included in the computer stores information processing application programs for implementing the functions of the feature extraction unit 11, the clustering unit 12, the pseudo label assignment unit 13, the inference unit 14, and the parameter update unit 15. A processor included in the computer executes the information processing application read from the memory, thereby implementing the functions of the feature extraction unit 11, the clustering unit 12, the pseudo label assignment unit 13, the inference unit 14, and the parameter update unit 15.

(Feature extraction unit)
The feature extraction unit 11 extracts features from data. For example, when extracting features from data in a three-dimensional space, the feature extraction unit 11 selects features to be extracted depending on the purpose for which the data will be used. Position information, normal information, color information, or a combination of these, of an object existing in the three-dimensional space is selected as the feature to be extracted. Then, the feature extraction unit 11 extracts the selected feature.

For example, when position information is selected as the feature, the feature extracting unit 11 extracts the position coordinates of the object as the feature.
Furthermore, when shape information is selected as the feature to be extracted, the feature extraction unit 11 calculates a feature (volume, surface area, length, etc.) representing the shape of the object. When color information is selected as the feature to be extracted, the feature extraction unit 11 extracts the hue, saturation, brightness, etc. of the object.
The feature extraction unit 11 may also scale the features. For example, if the scales of the position information are different, the feature extraction unit 11 may perform normalization to unify the ranges of each dimension. In the case of color information, the feature extraction unit 11 may scale the ranges of color values to unify them.

The learning data from which features are extracted includes sensor data, web data, image data, video data, text data, and the like.
The sensor data is data detected by a physical sensor. For example, the sensor data may include data such as temperature, humidity, or atmospheric pressure detected by a weather sensor, environmental monitoring data detected by a sensor network, and patient biometric information detected by a medical sensor. The feature extraction unit 11 can extract features from the sensor data acquired from the sensor by the information processing device 1.

Web data is data available on the Internet, such as information from web pages or user behavior data. The information processing device 1 may acquire web data by web scraping using a communication unit not shown in FIG. 1. The information processing device 1 may also acquire data from social media using the communication unit. The feature extraction unit 11 is capable of extracting features from the web data acquired by the information processing device 1 using the communication unit.

Image data and video data are information collected using cameras or sensors, and include, for example, video data captured by surveillance cameras, medical image data, camera image data for automobile detection or segmentation, and the like.
The feature extraction unit 11 is capable of extracting features from image data or video data acquired by the information processing device 1 from a camera or a sensor.

Text data is text information collected from books, news articles, web pages, social media posts, etc. It is used for natural language processing (NLP) tasks or text mining. The feature extraction unit 11 is able to extract features from text data acquired by the information processing device 1 using the communication unit.

(Feature extraction using convolution operations)
The feature extraction unit 11 may also extract features by a convolution operation. For example, a convolutional neural network (CNN) is used for the convolution operation. A matrix called a convolution kernel is used in the convolution operation. The convolution kernel is a weight matrix used when performing a convolution operation on input data. The convolution kernel functions as a filter for detecting specific patterns or features. The feature extraction unit 11 slides the convolution kernel along each position of the data, sequentially calculates the product of each element of the convolution kernel and the corresponding element of the data, and extracts the value at a specific position of the convolution kernel as a feature. This makes it possible to extract features that indicate local characteristics of the data.

The feature extraction unit 11 also extracts features from data based on the values of parameters set therein. For example, when the feature extraction unit 11 extracts features through a convolution operation using a CNN model, parameters for setting the size and stride of the convolution kernel affect the feature extraction. The larger the convolution kernel size, the wider the range of features extracted, and the larger the stride, the fewer the number of features extracted.
Furthermore, when the feature extraction unit 11 extracts features by moving a moving window over sequence data such as time-series data or text data, the values of parameters for setting the window width and step size affect the feature extraction: the larger the window width, the wider the range of features extracted, and the larger the step size, the fewer the number of features.

When the data from which features are to be extracted by convolution operations is one-dimensional data, two-dimensional data, three-dimensional data, or graph data, the feature extraction unit 11 performs one-dimensional convolution, two-dimensional convolution, three-dimensional convolution, or graph convolution to extract features.

(1) One-dimensional convolution For example, one-dimensional data includes time series data, audio data, sensor data, etc. When extracting features from one-dimensional data by convolution, the length of the data to be convolved becomes the kernel size, and the convolution kernel is defined as a one-dimensional array.
The feature extraction unit 11 slides the convolution kernel along each position of the data and sequentially calculates the product of each element of the convolution kernel and the corresponding element of the data. By extracting the value at a specific position of the convolution kernel as a feature, a feature indicating a local feature that emphasizes the relationship between neighboring elements in the data is extracted.
The position of the feature from which the feature is extracted may be the beginning, the center, or the end of the convolution kernel.

(2) Two-dimensional convolution For example, two-dimensional data includes image data. When extracting features from two-dimensional data by convolution calculation, the convolution kernel is defined as a two-dimensional matrix. The feature extraction unit 11 slides the convolution kernel along each position of the data and sequentially calculates the product of each element of the convolution kernel and the corresponding element of the data. By extracting values at specific positions of the convolution kernel as features, features indicating local features that emphasize the relationship between neighboring elements in the data are extracted.
The position of the feature from which the feature is extracted may be the beginning, the center, or the end of the convolution kernel.

(3) 3D Convolution For example, 3D data includes video data. When extracting features from 3D data using a convolution operation, the convolution kernel is defined as a 3D matrix having three dimensions: width, height, and channel. The feature extraction unit 11 slides the convolution kernel along each position of the data and sequentially calculates the product of each element of the convolution kernel and the corresponding element of the data. By extracting values at specific positions of the convolution kernel as features, features indicating local features that emphasize the relationship between neighboring elements in the data are extracted.
The position of the feature from which the feature is extracted may be the beginning, the center, or the end of the convolution kernel.

(4) Graph Convolution The feature extraction unit 11 may perform graph convolution, which is a convolution operation on data having a graph structure. Examples of data having a graph structure include social networks, power grids, and molecular structures. In graph convolution, a graph structure is provided as input data. A graph is composed of a set of nodes (vertices) and edges (lines), where the nodes represent elements of the data and the edges represent relationships between the nodes. The feature extraction unit 11 performs a convolution operation while moving a convolution kernel on the graph, thereby extracting specific patterns in the graph as features.

Alternatively, the feature extraction unit 11 may perform two-dimensional convolution on the three-dimensional data in the spatial direction, and then perform one-dimensional convolution on the three-dimensional data in the time direction.
The feature extraction unit 11 may convert the image data into a feature vector of a fixed length, and then perform one-dimensional convolution in the time direction.
The feature extraction unit 11 may perform derated convolution, which uses a convolution kernel to convolve elements at regular intervals in the data.

Although feature extraction using a CNN has been described, the feature extraction unit 11 may extract features using a recursive neural network (RNN) such as a long short-term memory (LSTM) or a gated recurrent unit (GRU).
Furthermore, the feature extraction unit 11 may extract features using an attention such as a transformer or an MLP-Mixer that replaces it.

(Clustering Department)
The clustering unit 12 determines the cluster to which a feature belongs by clustering the feature in the feature space. Clustering is a process of grouping points with similar feature amounts. For example, the clustering unit 12 calculates the distance or similarity between points in the feature space and clusters the feature amounts based on the calculated distance or similarity. The distance can be Euclidean distance or Manhattan distance. The similarity can be cosine similarity.

Typical clustering methods include, for example, k-means, Gaussian mixture models (GMM), hierarchical (agglomerative) clustering, spectral clustering, DBSCAN, OPTICS, BIRCH, or Mean Shift. The feature extraction unit 11 uses the selected clustering method to assign each feature point to a cluster. At this time, clustering is performed based on the distance or similarity between points in the feature space.

(Pseudo label assignment unit)
The pseudo label assignment unit 13 sequentially sets ranges within a cluster, generates pseudo labels based on cluster information corresponding to feature quantities included within the ranges, and assigns pseudo labels to at least one feature quantity included within the ranges. The pseudo label assignment unit 13 assigns pseudo labels to all points in the feature quantity space by performing the above process for each cluster set in the feature quantity space. The range set within a cluster is a moving window that includes points of one or more feature quantities, and defines the cluster in the same dimension as the feature quantities, for example. This range is sequentially set so that all points included in the cluster are targets for assigning pseudo labels.

The pseudo label may be any one of the mode, median, average, minimum, and maximum values of the cluster information corresponding to the feature amount included in the range set in the cluster.
For example, if the range set within a cluster is a fixed-length moving window for one-dimensional data of features, and the cluster information corresponding to the features included in the moving window is a cluster number, the pseudo label assignment unit 13 calculates one of the mode, median, average, minimum, or maximum of the cluster number corresponding to the features included in the moving window as the pseudo label.
Since pseudo labels can be generated that statistically take into account cluster information corresponding to the features contained in the clusters, the information processing device 1 can more accurately represent the features or patterns of the data.

The pseudo label assignment unit 13 may change the size of the range (moving window) set within the cluster as needed. Changing the size of the range makes it possible to extract features over different ranges. For example, when the processing accuracy is set using an input device not shown in FIG. 1, the pseudo label assignment unit 13 sets a range size according to this processing accuracy and performs the pseudo label assignment process. In this case, if the size of the range is changed significantly, the range from which features are extracted becomes wider, and if the size of the range is decreased, more localized features are extracted. The pseudo label assignment unit 13 may also change the size of the range to the size set by the user using the input device.

The pseudo label assignment unit 13 may assign pseudo labels to one-dimensional features using a filter with a feedback loop, such as an IIR (Infinite Impulse Response) filter. For example, the pseudo label assignment unit 13 adjusts the parameters of the IIR filter to control the position and size of the moving window, sequentially setting the moving window within the cluster, and assigns the information obtained by the IIR filter as pseudo labels to the features within the moving window.

The pseudo label assignment unit 13 may assign pseudo labels to two-dimensional features using an image filter. An image filter is a filter that uses a convolution operation to identify a specific area on two-dimensional data and transform the information within that area. Image filters include smoothing filters, edge detection filters, and sharpening filters. For example, the pseudo label assignment unit 13 adjusts the parameters of the image filter to control the position and size of the moving window, sequentially setting the moving window within the cluster, and assigns the information obtained by the image filter as a pseudo label to the feature within the moving window.

The pseudo label assignment unit 13 may assign pseudo levels using an averaging filter, a weighted average filter, a median filter, or a Laplacian filter.
The averaging filter is, for example, a filter that replaces the value of each pixel in an image with the average value of the values of its surrounding pixels. When the value of each pixel in an image is a feature, the pseudo label assignment unit 13 uses the averaging filter to sequentially set moving windows on clusters in the image and assigns the average value of all pixel values included in the moving windows as a pseudo label to the feature to be assigned.

A weighted average filter is, for example, a filter that calculates an average value by weighting the values of pixels surrounding a pixel of interest. When the value of each pixel in an image is a feature, the pseudo label assignment unit 13 uses the weighted average filter to sequentially set a moving window within a cluster on the image, calculate the average value of all pixel values included in the moving window, and assign the calculated value as a pseudo label to the feature to be assigned.

A median filter is a filter that, for example, calculates the median of the pixel values included in a moving window and uses the calculated value as the value of the pixel at the center of the moving window. When the value of each pixel in an image is a feature, the pseudo label assignment unit 13 uses a median filter to sequentially set moving windows within clusters on the image, calculates the median of the values of all pixels included in the moving window, and assigns the calculated value as a pseudo label to the feature to be assigned.

A Laplacian filter, for example, is a filter that emphasizes edges or features in an image and detects sudden changes in the image using a two-dimensional differential operator. When the value of each pixel in an image is a feature, the pseudo label assignment unit 13 uses a Laplacian filter to sequentially set moving windows within clusters on the image, calculates the values of pixels included in the moving windows that have suddenly changed, and assigns the calculated values as pseudo labels to the feature to be assigned.

The pseudo label assignment unit 13 may assign pseudo labels to the feature quantities of the graph data using a label propagation method or a label diffusion method.
Label propagation is a method of updating the labels of nodes in a graph based on the labels of surrounding nodes. Similarly, label diffusion is a method of updating the labels of nodes in a graph, but in label diffusion, the labels are updated taking into account not only the node labels but also information such as edge weights and distances.
For example, the pseudo label assignment unit 13 assigns an initial label to each node and updates the label using the mode or average value of the labels of adjacent nodes. This allows the labels to be propagated or diffused throughout the graph. The pseudo label assignment unit 13 assigns pseudo labels based on the labels of the feature quantities included in the moving window. By repeating these processes until the label updates converge, it is possible to assign pseudo labels to each feature quantity point within the cluster.

(Inference Department)
The inference unit 14 calculates an inferred value based on the feature quantities extracted from the data by the feature quantity extraction unit 11. For example, the inference unit 14 uses a model for calculating an inferred value from the feature quantities. The model is a model of the relationship between the input feature quantities and the output inferred value. Examples of the model include a neural network, a generalized linear model, a generalized additive model, a decision tree, a decision tree ensemble, a support vector machine, and a nearest neighbor model. The parameter values of the model are updated by the parameter update unit 15 (described later), and the model is trained so that it can generate an appropriate inferred value from the input data. The inference unit 14 uses the trained model to perform processing such as prediction, classification, or regression on the input data.

For example, when a feature is input to the input layer of a neural network in the inference unit 14, the feature propagates from the input layer through the hidden layer to the output layer in the neural network, and a final inference value is generated. In the fully connected layer, a bias is added to the dot product of the input and weight, and an output value is calculated using an activation function. Activation functions include ReLU (Rectified Linear Unit), Sigmoid, Softmax, and Tanh. Note that it is also possible to output the value obtained by adding a bias to the dot product of the input and weight without using an activation function.

(Parameter update section)
The parameter update unit 15 updates the parameters of the feature extraction unit 11 and the inference unit 14 based on the pseudo label and the inference value. For example, the parameter update unit 15 updates the parameters of the feature extraction unit 11 and the inference unit 14 so that the inference value output from the inference unit 14 matches the pseudo label output from the pseudo label assignment unit 13.

For example, the parameter update unit 15 uses a loss function to evaluate the degree of agreement between the inferred value and the pseudo label. The loss function is an index that represents the difference between the inferred value and the pseudo label, and is used to evaluate the performance of the model. The value of the loss function becomes smaller as the inferred value and the pseudo label value become closer.
Note that loss functions include cross-entropy when the inferred values are categorical, and mean absolute error (MAE) or mean squared error (MSE) when the inferred values are numeric.

The parameter update unit 15 updates the parameters of the feature extraction unit 11 and the inference unit 14 so as to minimize the value of the loss function.
For example, if the feature extraction unit 11 and the inference unit 14 are capable of gradient calculation of a neural network or the like, the parameter update unit 15 calculates the gradient of the loss function using the gradient descent method and uses the calculated gradient to calculate the amount of update for each parameter of the feature extraction unit 11 and the inference unit 14. The parameter update unit 15 repeats the above procedure to perform learning until the value of the loss function converges. As a result, the parameters are updated in a direction that minimizes the loss function.
Gradient descent methods include stochastic gradient descent (SGD), mini-batch gradient descent, Adam, or RMSProp.

Next, the operation of the information processing device 1 will be described.
FIG. 2 is a flowchart showing the operation of the information processing device 1.
The parameter values of the feature extraction unit 11 are initialized (step ST1). For example, the feature extraction unit 11 initializes parameters such as the weights of a model for extracting features with random values or predefined values.

The information processing device 1 acquires data for learning (step ST2). The data acquired by the information processing device 1 is output to the feature extraction unit 11. The learning dataset is a dataset for unsupervised learning that does not have a correct answer label assigned.

The information processing device 1 then performs a learning process using a learning dataset (step ST3). For example, the feature extraction unit 11 extracts features from the data. The clustering unit 12 determines the cluster to which each feature belongs by clustering the features in feature space. The pseudo label assignment unit 13 sequentially sets ranges within the cluster, generates pseudo labels based on cluster information corresponding to the features included within the range, and assigns the generated pseudo label to at least one feature included within the range. The inference unit 14 calculates inference values based on the features, and the parameter update unit 15 updates the parameters of the feature extraction unit 11 and the inference unit 14 based on the pseudo labels and inference values.

It is determined whether the learning conditions have ended (step ST4). If the learning conditions have ended (step ST4; YES), proceed to step ST5. On the other hand, if the learning conditions have not ended (step ST4; NO), return to step ST3 and the learning process is repeated. That is, feature extraction by the feature extraction unit 11, clustering by the clustering unit 12, assignment of pseudo labels by the pseudo label assignment unit 13, inference by the inference unit 14, and parameter update by the parameter update unit 15 are repeatedly executed until the learning conditions end. Note that the learning conditions may be, for example, a condition of whether a predetermined number of repetitions has been reached, or a condition of whether a predetermined learning time has been reached, etc.

The learning condition may be the following condition (1) or (2).
(1) The condition is that learning is terminated when the number of samples whose pseudo labels change becomes equal to or less than a predetermined value.
(2) The acquired data is divided into data for learning and data for verification, and the information processing device 1 uses the learning data to update the parameters of the feature extraction unit 11 and the inference unit 14 in step ST3. The learning condition is that if the value of a loss function representing the difference between an inferred value and a pseudo label does not decrease a predetermined number of times consecutively using the verification data, the value of the loss function is deemed to have converged and learning is terminated.

Furthermore, the information processing device 1 may perform clustering and pseudo-labeling, and inference and parameter updating at a 1:1 ratio, or at different ratios.
For example, the information processing device 1 may perform clustering and pseudo-labeling once, and then repeat inference and parameter updating ten times according to the stochastic gradient descent method.

The inference unit 14 performs inference processing using the model trained in step ST4 (step ST5). Inference includes prediction, classification, regression, etc. for new data. When the features of new data are input, the trained model outputs a prediction or classification result.

Next, a learning method according to the first embodiment will be described.
FIG. 3 is a flowchart showing the learning method according to the first embodiment, illustrating the detailed processing of step ST3 in FIG.
The feature extraction unit 11 extracts features from the learning data (step ST1A).
Fig. 4 is a diagram showing an outline of a process for extracting a feature quantity Z from data X. In Fig. 4, data X is one-dimensional data composed of elements x _i . Feature quantity Z is a feature quantity vector composed of feature quantities z _i , where i is an integer equal to or greater than 1.
The feature extraction unit 11 calculates the feature z _i using the relational expression z _i = Feature(Neighbor(x _i )). The feature extraction unit 11 calculates the neighborhood of x _i included in the data X. The neighborhood is a set of other elements existing around the element x _i . For example, it is the k nearest neighbors of the element x _i or elements within a certain distance range.
The feature extraction unit 11 calculates a feature z _i using a feature function based on the neighborhood data. The feature z _i calculated using the feature function is each element of the feature corresponding to each element x _i of the data. In this way, a feature Z including the feature z _i is extracted from the one-dimensional data X.

The clustering unit 12 determines the cluster to which the feature belongs by clustering the feature in the feature space (step ST2A).
FIG. 5 is a diagram showing an overview of clustering of feature quantity Z. In FIG. 5, cluster C is a cluster into which feature quantity z _i is classified. Feature quantity z _i is assigned a label c _i . The clustering unit 12 clusters feature quantity z _i into cluster C according to a clustering algorithm set therein. Examples of clustering algorithms include k-means, hierarchical clustering, and DBSCAN. These algorithms determine which cluster C feature quantity z _i belongs to.
The clustering unit 12 assigns a label _c corresponding to the feature _z using the relational expression _c = Cluster( _z ), thereby determining the cluster C to which the feature _z belongs. Note that the value of _c is information representing the cluster C to which the feature _z belongs.

The pseudo label assignment unit 13 sequentially sets ranges (moving windows) within the cluster, generates pseudo labels based on cluster information corresponding to features included within the ranges, and assigns the pseudo labels to at least one feature included within the ranges (step ST3A).
Fig. 6 is a diagram showing an outline of the process of assigning a pseudo label a _i to each point in cluster C. In Fig. 6, pseudo label A is composed of pseudo labels a _i that are assigned pseudo based on labels c _i included in cluster C. The pseudo label assignment unit 13 sets a moving window W within cluster C according to the function Neighbor(c _i ) and calculates the label c _i of each element within the moving window W. As a result, a set of labels for each element included in the moving window W is obtained.
Next, the pseudo label assignment unit 13 calculates a pseudo label a _i based on a label obtained by aggregating the labels c _i of the elements within the moving window W according to the Aggregate function.
For example, the mode or average value of the aggregated labels c _i is calculated as the pseudo label a _i . The pseudo label assignment unit 13 assigns the calculated pseudo label a _i to the element to be assigned. As a result, a set of pseudo labels a _i included in the pseudo label A is obtained.

The inference unit 14 calculates an inference value based on the feature amount (step ST4A).
Fig. 7 is a diagram showing an outline of inference using the feature quantity Z. In Fig. 7, the inference result Y is the result of inference using the feature quantity Z, and is a set of elements _yi of the inference result based on the feature quantity _zi . The inference unit 14 performs inference according to the relational expression _yi = Predict( _zi ).
For example, when a feature value z _i is input, the inference unit 14 generates an inference result Y using an inference model that outputs an element y _i of an inference value. This makes it possible to infer target information from the feature value z _i .

The parameter update unit 15 updates the parameters of the feature extraction unit 11 and the inference unit 14 based on the pseudo label and the inference value (step ST5A).
FIG. 8 is a diagram showing an overview of parameter updates for the feature extraction unit 11 and the inference unit 14. In FIG. 8, the loss function Loss( _ai , _yi ) is a function that quantifies the difference between the pseudo label _ai and the element _yi of the inference result, and is used to evaluate the degree of match between them. The parameter update unit 15 updates each parameter of the feature extraction unit 11 and the inference unit 14 in a direction that minimizes the value of the loss function Loss( _ai , _yi ). When the value of the loss function becomes equal to or less than a threshold or a certain number of iterations is reached, the parameter values are considered to have converged, and learning ends.

Next, a hardware configuration for realizing the functions of the information processing device 1 will be described.
9 is a block diagram showing a hardware configuration that realizes the functions of the information processing device 1. The functions of the feature extraction unit 11, clustering unit 12, pseudo-label assignment unit 13, inference unit 14, and parameter update unit 15 provided in the information processing device 1 are realized by processing circuits. That is, the information processing device 1 includes a processing circuit for executing the processes of steps ST1A to ST5A shown in FIG. 3. The processing circuit may be a CPU (Central Processing Unit) that executes a program stored in a memory.

The feature extraction unit 11 acquires a learning dataset received from an external device by a communication unit included in the information processing device 1, for example, via the input interface 100. Furthermore, if the dataset is stored in a memory unit included in the information processing device 1, the feature extraction unit 11 reads and acquires the dataset from the memory unit via the input interface 100. In this case, the information processing device 1 does not need to be equipped with a communication unit.

The inference unit 14 outputs the inference result to an external device, for example, via the output interface 101. The inference unit 14 may also transmit the inference result to an external device by controlling the communication unit via the output interface 101.

The functions of the feature extraction unit 11, clustering unit 12, pseudo label assignment unit 13, inference unit 14, and parameter update unit 15 included in the information processing device 1 are realized by software, firmware, or a combination of software and firmware.
The software or firmware is written as a program and stored in the memory 103 .

The processor 102 reads and executes programs stored in the memory 103 to implement the functions of the feature extraction unit 11, clustering unit 12, pseudo label assignment unit 13, inference unit 14, and parameter update unit 15 provided in the information processing device 1. For example, the information processing device 1 includes a memory 103 for storing programs that, when executed by the processor 102, result in the processing of steps ST1A to ST5A shown in FIG. 3 being performed. These programs cause a computer to execute the procedures or methods of the processing performed by the feature extraction unit 11, clustering unit 12, pseudo label assignment unit 13, inference unit 14, and parameter update unit 15. The memory 103 may be a computer-readable storage medium that stores programs for causing a computer to function as the feature extraction unit 11, clustering unit 12, pseudo label assignment unit 13, inference unit 14, and parameter update unit 15.

Memory 103 may be, for example, non-volatile or volatile semiconductor memory such as RAM (Random Access Memory), ROM (Read Only Memory), flash memory, EPROM (Erasable Programmable Read Only Memory), EEPROM (Electrically-EPROM) (registered trademark), magnetic disk, flexible disk, optical disk, compact disk, mini disk, DVD, etc.

Some of the functions of the feature extraction unit 11, clustering unit 12, pseudo label assignment unit 13, inference unit 14 and parameter update unit 15 provided in the information processing device 1 may be realized by dedicated hardware, and other functions may be realized by software or firmware.
For example, the functions of the feature extraction unit 11, the clustering unit 12, and the pseudo-labeling unit 13 may be realized by a processing circuit that is dedicated hardware, and the functions of the inference unit 14 and the parameter update unit 15 may be realized by the processor 102 reading and executing a program stored in the memory 103. In this way, the processing circuit can realize the above functions by hardware, software, firmware, or a combination of these.

As described above, the information processing device 1 according to the first embodiment includes a feature extraction unit 11 that extracts features from data, a clustering unit 12 that determines a cluster to which a feature belongs by clustering the features in a feature space, a pseudo label assignment unit 13 that sequentially sets ranges within a cluster, generates pseudo labels based on cluster information corresponding to the features included within the ranges, and assigns the generated pseudo label to at least one feature included within the ranges, an inference unit 14 that calculates an inferred value based on the features, and a parameter update unit 15 that updates parameters of the feature extraction unit 11 and the inference unit 14 based on the pseudo label and the inferred value.
This enables the information processing device 1 to perform learning by assigning a pseudo label to each point of the feature of the cluster. For example, when the learning data is image data, the information processing device 1 can assign a pseudo label to each pixel in the image, and can be applied to image segmentation.

In the information processing device 1 according to embodiment 1, the feature extraction unit 11 extracts features by convolution operations. This enables the information processing device 1 to extract features that indicate local characteristics of data.

In the information processing device 1 according to embodiment 1, the pseudo label is one of the mode, median, mean, minimum, or maximum of the cluster information corresponding to the feature amounts included in the range. Because it is possible to generate pseudo labels that statistically consider the cluster information corresponding to the feature amounts included in the cluster, the information processing device 1 can more accurately represent the features or patterns of the data.

The learning method according to the first embodiment includes step ST1A in which a feature extraction unit 11 extracts features from data; step ST2A in which a clustering unit 12 determines a cluster to which a feature belongs by clustering the features in a feature space; step ST3A in which a pseudo label assignment unit 13 sequentially sets ranges within a cluster, generates pseudo labels based on cluster information corresponding to the features included in the ranges, and assigns the generated pseudo label to at least one feature included in the range; step ST4A in which an inference unit 14 calculates an inference value based on the feature; and step ST5A in which a parameter update unit 15 updates parameters of the feature extraction unit 11 and the inference unit 14 based on the pseudo label and the inference value.
By executing the above learning method, the information processing device 1 can perform learning in which pseudo labels are assigned to each point of the feature quantity of a cluster.

The computer that executes the program according to the first embodiment functions as a feature extraction unit 11 that extracts features from data, a clustering unit 12 that determines the cluster to which a feature belongs by clustering the features in the feature space, a pseudo label assignment unit 13 that sequentially sets ranges within the cluster, generates pseudo labels based on cluster information corresponding to the features included in the ranges, and assigns the generated pseudo label to at least one feature included in the range, an inference unit 14 that calculates an inferred value based on the feature, and a parameter update unit 15 that updates the parameters of the feature extraction unit 11 and the inference unit 14 based on the pseudo label and the inferred value.
A computer that executes the program according to the first embodiment is capable of performing learning by assigning pseudo labels to each point of the feature quantity of a cluster.

Embodiment 2.
Although the first embodiment does not mention the characteristics of the training data, the information processing device according to the second embodiment acquires data having spatial continuity as training data. As a result, the information processing device according to the second embodiment can assign pseudo labels taking into account the spatial continuity of the training data by extracting features from spatially continuous local data.

(Outline of information processing device)
10 is a block diagram showing an example configuration of an information processing device 1A according to embodiment 2. In FIG. 10, the information processing device 1A acquires data having spatial continuity and extracts features from the spatially continuous local data. The information processing device 1A then sequentially sets ranges within clusters obtained by clustering the features, generates pseudo labels based on cluster information corresponding to the features included within the ranges, and assigns the pseudo labels to at least one feature included within the ranges.

The information processing device 1A includes a feature extraction unit 11A, a clustering unit 12A, a pseudo-label assignment unit 13A, an inference unit 14A, a parameter update unit 15A, and a data acquisition unit 16.
For example, the information processing device 1A is realized by a computer. A memory included in the computer stores programs constituting information processing applications for realizing the functions of the feature extraction unit 11A, the clustering unit 12A, the pseudo label assignment unit 13A, the inference unit 14A, the parameter update unit 15A, and the data acquisition unit 16. A processor included in the computer executes the information processing application read from the memory, thereby realizing the functions of the feature extraction unit 11A, the clustering unit 12A, the pseudo label assignment unit 13A, the inference unit 14A, the parameter update unit 15A, and the data acquisition unit 16.

(Data acquisition section)
The data acquisition unit 16 acquires data having spatial continuity. For example, the data acquisition unit 16 acquires the data having spatial continuity by connecting to a sensor device or a server via a network using a communication unit not shown in FIG. 10 .
Examples of data with spatial continuity include time series data, audio data, image data, video data, spectrograms, text data, three-dimensional point clouds, range Doppler, and graph data. Time series data and audio data are one-dimensional data with temporal continuity. Image data is two-dimensional data that represents two-dimensional space. Video data is data that represents two-dimensional space and also has a time dimension. Spectrograms have the dimensions of time and frequency. Text data is one-dimensional data with an order dimension. Three-dimensional point clouds are data in three-dimensional space. Range Doppler is data with the dimensions of distance and velocity. Graph data is data represented by node connections.
Furthermore, feature points extracted from spatially continuous local data may be scalars or vectors. For example, features of univariate time-series data are represented by scalars, and features of multivariate time-series data are represented by vectors. Color image data is represented by a three-dimensional vector in which each pixel value has color information of R (red), G (green), and B (blue).

(Feature extraction unit)
The feature extraction unit 11A extracts features from spatially continuous local data acquired by the data acquisition unit 16. These features also have spatial continuity. When the data from which features are to be extracted by convolution calculation is one-dimensional data, two-dimensional data, three-dimensional data, or graph data, the feature extraction unit 11A extracts features by performing one-dimensional convolution, two-dimensional convolution, three-dimensional convolution, or graph convolution.

(Clustering Department)
The clustering unit 12A determines the cluster to which the feature value belongs by clustering the feature values having spatial continuity.
For example, the clustering unit 12A may perform hard clustering or soft clustering.
Hard clustering is a method of assigning each feature point to only a unique cluster, and assigns, as cluster information, for example, a single cluster number (an integer value) to each point in the cluster. Examples of hard clustering include k-means clustering.
Soft clustering is a method in which each point of a feature quantity has the possibility of belonging to multiple clusters, and a score (real value) is assigned to each cluster as cluster information. Examples of soft clustering include GMM.

(Pseudo label assignment unit)
The pseudo label assignment unit 13A sequentially sets ranges within a cluster, generates pseudo labels based on cluster information corresponding to feature amounts included in the ranges, and assigns the pseudo label to at least one feature amount included in the ranges.
For example, the pseudo labeling unit 13A aggregates information on spatially continuous local clusters within the above range and assigns a pseudo label to the clusters.
In the case of hard clustering, the pseudo-labeling unit 13A assigns information that aggregates cluster numbers (integer values) as pseudo labels. The aggregation may be the mode, median, mean, minimum, maximum, or quantile of the cluster numbers assigned to points of feature quantities included in a range set within the cluster.
In the case of soft clustering, the pseudo label assignment unit 13A assigns, for example, information that aggregates scores (real-valued values) for clusters as pseudo labels.

(Inference Department)
The inference unit 14A calculates an inferred value based on the feature extracted from the data by the feature extraction unit 11A. For example, the inference unit 14A calculates the inferred value using a neural network. When the feature is input to the input layer of the neural network, the feature propagates from the input layer to the output layer via the hidden layer, and a final inferred value is generated. In the fully connected layer, a bias is added to the inner product of the input and the weight, and an output value is calculated using an activation function. When the activation function is a sigmoid function, an inferred value for a binary category is output. When the activation function is a Softmax function, an inferred value for a multi-value category is output.

(Parameter update section)
The parameter update unit 15A updates the parameters of the feature extraction unit 11A and the inference unit 14A based on the pseudo label and the inference value. For example, the parameter update unit 15A updates the parameters of the feature extraction unit 11A and the inference unit 14A so that the inference value output from the inference unit 14A matches the pseudo label output from the pseudo label assignment unit 13A.

Next, a learning method according to the second embodiment will be described.
FIG. 11 is a flowchart showing a learning method according to the second embodiment, illustrating the detailed processing of step ST3 in FIG.
The data acquisition unit 16 acquires data having spatial continuity (step ST1B).
The feature extraction unit 11A extracts feature amounts from spatially continuous local data (step ST2B).
The clustering unit 12A determines the cluster to which the feature belongs by clustering the feature in the feature space (step ST3B).
The pseudo label assignment unit 13A sequentially sets moving windows within the cluster, generates pseudo labels based on cluster information corresponding to the features included in the moving window, and assigns the pseudo label to at least one feature included in the range (step ST4B).
The inference unit 14A calculates an inference value based on the feature amount (step ST5B).
The parameter update unit 15A updates the parameters of the feature extraction unit 11A and the inference unit 14A based on the pseudo label and the inference value (step ST6B).

FIG. 12A is a diagram showing an overview of the process of assigning pseudo labels to feature points within a cluster, which is time-series data. In FIG. 12A, the pseudo label assignment unit 13A sets a moving window W1 for cluster D1, which is time-series data. The moving window W1 includes point P1, which is the target of assigning a pseudo label, and its neighboring point PN. The pseudo label assignment unit 13A calculates a pseudo label by aggregating the label numbers of points P1 and PN included in the moving window W1. The pseudo label assignment unit 13A assigns the pseudo label to point P1, which is the target of assignment. By repeating this process, pseudo labels are assigned to all points included in cluster D1. This enables the information processing device 1A to perform learning by assigning pseudo labels to each feature point, which is time-series data.

FIG. 12B is a diagram showing an overview of the process of assigning pseudo labels to feature points within a cluster, which is two-dimensional data. In FIG. 12B, the pseudo label assignment unit 13A sets a moving window W2 within cluster D2, which is two-dimensional data. The moving window W2 includes point P2, which is the target of assigning a pseudo label, and its neighboring points. The pseudo label assignment unit 13A calculates a pseudo label by aggregating the label numbers of the points included in the moving window W2. The pseudo label assignment unit 13A assigns a pseudo label to point P2, which is the target of assignment. By repeating this process, pseudo labels are assigned to all points included in cluster D2. This enables the information processing device 1A to perform learning by assigning pseudo labels to each feature point, which is two-dimensional data.

Figure 12C is a diagram showing an overview of the process of assigning pseudo labels to feature points within a cluster, which is three-dimensional data. In Figure 12C, the pseudo label assignment unit 13A sets a moving window W3 within cluster D3. The moving window W3 includes point P3, which is the target of assigning a pseudo label, and its neighboring points. The pseudo label assignment unit 13A calculates a pseudo label by aggregating the label numbers of the points included in the moving window W3. The pseudo label assignment unit 13A assigns the pseudo label to point P3, which is the target of assignment. By repeating this process, pseudo labels are assigned to all points included in cluster D3. This enables the information processing device 1A to perform learning by assigning pseudo labels to each feature point, which is three-dimensional data.

FIG. 12D is a diagram showing an overview of the process of assigning pseudo labels to feature points within a cluster, which is graph data. In FIG. 12D, the pseudo label assignment unit 13A sets a moving window W4 within cluster D4. The moving window W4 includes point P4, which is the target of assigning a pseudo label, and its neighboring points. The pseudo label assignment unit 13A calculates a pseudo label by aggregating the label numbers of the points included in the moving window W4. The pseudo label assignment unit 13A assigns the pseudo label to point P4, which is the target of assignment. By repeating this process, pseudo labels are assigned to all points included in cluster D4. This enables the information processing device 1A to perform learning by assigning pseudo labels to each feature point, which is graph data.

(Variation 1)
In variant example 1, the information processing device 1A acquires time series data as a learning dataset, performs hard clustering on each point of the feature extracted from the time series data, and assigns cluster information, which is an integer value indicating a cluster number, to each point.
The time-series data is, for example, time-series data of acceleration and angular acceleration detected by sensors attached to a person's hands and feet. This time-series data is segmented by the type of movement of the person wearing the sensor, such as walking, running, or sitting. Since the true value of the segment, i.e., the type of movement, is unknown, no training data is provided.
The operation of the information processing device 1A according to the first modification will be described below with reference to the flowcharts of FIGS.

The feature extraction unit 11A extracts features by one-dimensional convolution calculation using a CNN.
First, the information processing device 1A randomly initializes parameters of a one-dimensional convolutional layer in a CNN (step ST1).
The data acquisition unit 16 acquires time-series data of acceleration and angular acceleration detected by sensors attached to the person's hands and feet (step ST2), which enables the information processing device 1A to perform learning by assigning pseudo-labels to each point of spatially continuous local feature quantities.
Thereafter, the information processing device 1A uses the time-series data acquired by the data acquisition unit 16 as learning data and repeatedly executes the following learning process (processing from step ST1A to step ST5A) a specified number of times (step ST3).

The feature extraction unit 11A extracts a time series of features from time series data using a one-dimensional convolutional layer of a CNN (step ST1A).
The clustering unit 12A performs hard clustering on each point of the time-series feature extracted by the feature extraction unit 11A using k-means with a cluster count of 10 (step ST2A). As a result, an integer value indicating a cluster number is assigned as cluster information to each point of the feature included in a cluster.

The pseudo label assignment unit 13A sequentially sets a moving window of size 5 within the cluster at each time point, takes a majority vote on the cluster information within the moving window to convert it into a pseudo label, and sequentially assigns the pseudo label to the feature at the target time point (step ST3A).
The inference unit 14A converts the feature quantity at each time point into a 10-dimensional vector using a fully connected layer of a CNN and a Softmax function (step ST4A).

The parameter update unit 15A updates each parameter of the one-dimensional convolutional layer of the CNN used by the feature extraction unit 11A and the fully connected layer of the CNN used by the inference unit 14A using stochastic gradient descent so as to minimize the value of the cross-entropy loss function calculated based on the pseudo label at each time point and the 10-dimensional vector (step ST5A).
The processes from step ST1A to step ST5A are repeatedly executed until the designated number of times is reached (step ST4).

The inference unit 14A performs segmentation on the inference result data (step ST5). For example, the inference unit 14A acquires a time series of features extracted from the time series data by a one-dimensional convolutional layer of a CNN, and converts the features at each time point into a 10-dimensional vector using a fully connected layer and a Softmax function. The inference unit 14A outputs, as a segment, the cluster in which the 10-dimensional vector is maximized at each time point.
This allows the information processing device 1A to segment the time series data of acceleration and angular acceleration detected by sensors attached to a person's hands and feet by the type of movement, such as walking, running, or sitting, of the person wearing the sensor.

(Variation 2)
In Modification 2, the information processing device 1A acquires image data, which is two-dimensional data, as a learning dataset, and performs soft clustering on each point of feature quantities extracted from the image data. The information processing device 1A segments the image data on a pixel-by-pixel basis.
The information processing device 1A uses a trained model trained in advance for segmentation, but performs segmentation in a target domain that is different from the source domain trained in advance. It is assumed that there is no training data in the target domain. The operation of the information processing device 1A in Variation 2 will be described below with reference to the flowcharts of FIGS. 2 and 3 .

The feature extraction unit 11A extracts features by two-dimensional convolution calculation using a CNN.
First, the information processing device 1A randomly initializes parameters of a two-dimensional convolutional layer in a CNN (step ST1).
The data acquisition unit 16 acquires image data (step ST2).
Thereafter, the information processing device 1A uses the image data acquired by the data acquisition unit 16 as learning data and repeatedly executes the following learning process (processing from step ST1A to step ST5A) a specified number of times (step ST3).

The feature extraction unit 11A extracts a time series of features from image data using a two-dimensional convolutional layer of a CNN (step ST1A).
The clustering unit 12A performs soft clustering on the feature of each pixel extracted from the image data by the feature extraction unit 11A using a GMM with 100 clusters (step ST2A). A cluster likelihood, which is a score for the cluster, is assigned to the feature of each pixel included in a cluster as cluster information.

The pseudo label assignment unit 13A sequentially sets a moving window, which is an averaging filter of size 3x3, for the cluster, averages the cluster likelihood of each pixel using the averaging filter, calculates the maximum cluster likelihood as a pseudo label, and sequentially assigns pseudo labels to the features of the pixels to be assigned (step ST3A).
The inference unit 14A converts the feature quantity of each pixel into a 100-dimensional vector using a fully connected layer of a CNN and a Softmax function (step ST4A).

The parameter update unit 15A updates each parameter of the two-dimensional convolutional layer of the CNN used by the feature extraction unit 11A and the fully connected layer of the CNN used by the inference unit 14A using stochastic gradient descent so as to minimize the value of the cross-entropy loss function calculated based on the pseudo label of each pixel and the 100-dimensional vector (step ST5A).
The processes from step ST1A to step ST5A are repeatedly executed until the designated number of times is reached (step ST4).

The inference unit 14A performs segmentation on the inference result data (step ST5). For example, the inference unit 14A acquires the feature of each pixel extracted from the image data by the two-dimensional convolutional layer of the CNN, and converts the feature of each pixel into a 100-dimensional vector using a fully connected layer and a Softmax function. The inference unit 14A outputs, as a segment, the cluster in which the 100-dimensional vector for each pixel is maximized.
This allows the information processing device 1A to segment image data in a target domain that is different from the pre-trained source domain.

As described above, the information processing device 1A according to embodiment 2 includes a data acquisition unit 16 that acquires spatially continuous data, and a feature extraction unit 11A that extracts feature vectors from spatially continuous local data among the acquired data. This enables the information processing device 1A to perform learning by assigning pseudo-labels to each point of spatially continuous local feature values.

In the information processing device 1A according to embodiment 2, the data acquisition unit 16 acquires time-series data. This enables the information processing device 1A to perform learning by assigning pseudo-labels to each point of the feature quantity, which is the time-series data.

In the information processing device 1A according to the second embodiment, the clustering unit 12A determines clusters by hard clustering. This enables the information processing device 1A to perform learning by assigning pseudo labels calculated based on cluster information that is assigned only to unique clusters to each feature point within a cluster.

In the information processing device 1A according to the second embodiment, the cluster information is an integer value indicating the cluster number. This enables the information processing device 1A to perform learning by assigning a pseudo label calculated based on the cluster number to each feature point within a cluster.

In the information processing device 1A according to the second embodiment, the clustering unit 12A determines clusters by soft clustering. This enables the information processing device 1A to perform learning by assigning pseudo labels calculated based on cluster information that can be assigned to multiple clusters to each feature point within a cluster.

In the information processing device 1A according to the second embodiment, the cluster information is a score for the cluster. This enables the information processing device 1A to perform learning by assigning a pseudo label calculated based on the score for the cluster to each feature point within the cluster.

It is possible to combine the various embodiments, modify any of the components in each embodiment, or omit any of the components in each embodiment.

The information processing device disclosed herein can be used for various information processing tasks, such as unsupervised learning.

1, 1A: Information processing device, 11, 11A: Feature extraction unit, 12, 12A: Clustering unit, 13, 13A: Pseudo label assignment unit, 14, 14A: Inference unit, 15, 15A: Parameter update unit, 16: Data acquisition unit, 100: Input interface, 101: Output interface, 102: Processor, 103: Memory.

Claims (11)

a feature extraction unit that extracts features from data;
a clustering unit that determines a cluster to which the feature value belongs by clustering the feature value in a feature value space;
a pseudo label assigning unit that sequentially sets ranges within the cluster, generates pseudo labels based on cluster information corresponding to the feature amounts included in the ranges, and assigns the generated pseudo labels to at least one of the feature amounts included in the ranges;
an inference unit that calculates an inference value based on the feature amount;
a parameter update unit that updates parameters of the feature extraction unit and the inference unit based on the pseudo label and the inference value;
An information processing device comprising: a data acquisition unit that acquires data having spatial continuity;
The information processing apparatus according to claim 1 , wherein the feature extraction unit extracts the feature from spatially continuous local data among the acquired data. The information processing apparatus according to claim 2 , wherein the data acquisition unit acquires time-series data. The information processing device according to claim 1 , wherein the clustering unit determines the clusters by hard clustering. The information processing apparatus according to claim 4 , wherein the cluster information is an integer value indicating a cluster number. The information processing device according to claim 1 , wherein the clustering unit determines the clusters by soft clustering. The information processing apparatus according to claim 6 , wherein the cluster information is a score for the cluster. The information processing apparatus according to claim 1 , wherein the feature extraction unit extracts the feature by a convolution operation. The information processing device according to claim 1 , wherein the pseudo label is one of a mode, a median, an average, a minimum value, and a maximum value of the cluster information corresponding to the feature included in the range. A learning method executed by an information processing device,
A feature extraction unit extracts features from data;
a clustering unit determining a cluster to which the feature value belongs by clustering the feature value in a feature value space;
a pseudo label assignment unit sequentially setting ranges within the cluster, generating pseudo labels based on cluster information corresponding to the feature amounts included in the ranges, and assigning the generated pseudo labels to at least one of the feature amounts included in the ranges;
an inference unit calculating an inferred value based on the feature amount;
a parameter updating unit updating parameters of the feature extraction unit and the inference unit based on the pseudo label and the inference value;
A learning method that includes: Computer,
a feature extraction unit that extracts features from the data;
a clustering unit that determines a cluster to which the feature value belongs by clustering the feature value in a feature value space;
a pseudo label assignment unit that sequentially sets ranges within the cluster, generates pseudo labels based on cluster information corresponding to the feature amounts included in the ranges, and assigns the generated pseudo labels to at least one of the feature amounts included in the ranges;
an inference unit that calculates an inference value based on the feature amount;
a program for causing the program to function as a parameter update unit that updates parameters of the feature extraction unit and the inference unit based on the pseudo label and the inference value;