JP7700542B2 - Information processing device, information processing method, and program - Google Patents

Description

The present invention relates to an information processing device, an information processing method, and a program.

Neural networks (hereafter referred to as "NNs") have high performance in areas such as image recognition. It is known that in order to improve the learning accuracy of NNs, a huge amount of input data and corresponding teacher labels are required. However, teacher labels are often assigned manually. As a result, the burden of assigning teacher labels to huge amounts of input data falls on the user.

In recent years, in order to solve this problem, research has been actively conducted into semi-supervised learning, in which teacher labels are assigned to a small amount of collected input data and the remaining data is trained by a neural network without teacher labels. Semi-supervised learning can significantly reduce the burden on users. In general, the loss function used in semi-supervised learning is defined as the weighted sum of the loss function corresponding to the dataset with teacher labels (labeled dataset) and the loss function corresponding to the dataset without teacher labels (unlabeled dataset).

The method described in Non-Patent Document 1 is one of the semi-supervised learning methods in image recognition, in which two types of data augmentation are applied to an image without teacher labels, and learning is performed based on comparing the two types of images obtained by the two types of data augmentation. This makes it possible to improve the learning accuracy. In the method described in Non-Patent Document 1, a constant that does not depend on input data without teacher labels (unlabeled data) is multiplied by the loss function corresponding to each piece of unlabeled data to calculate the loss function corresponding to the entire unlabeled data.

The method described in Non-Patent Document 2 is a type of semi-supervised learning method that calculates a coefficient indicating the degree of influence for each piece of unlabeled data, and calculates a loss function corresponding to the entire unlabeled data by a weighted sum of the coefficient for each piece of unlabeled data and the loss function. However, the coefficients are calculated using a dataset with teacher labels for validation that is prepared separately from the training data.

Kihyuk Sohn and 8 others, "FixMatch: Simplifying Semi-Supervised Learning with Consistencyand Confidence", [online], [Retrieved June 18, 2021], Internet <https://arxiv.org/abs/2001.07685> Zhongzheng Ren and 2 others, "Not AllUnlabeled Data are Equal: Learning to Weight Data in Semi-supervised Learning", [online], [Retrieved June 18, 2021], Internet <https://arxiv.org/abs/2007.01293>

However, according to the technique described in Non-Patent Document 1, the loss functions corresponding to each piece of unlabeled data are treated equally, and then the NN is updated based on the loss functions. As a result, the loss functions corresponding to unlabeled data that interferes with learning are also treated equally, making learning unstable and making it difficult to increase the accuracy of the constructed NN.

In addition, according to the method described in Non-Patent Document 2, the influence of unlabeled data can be determined for each sample. However, since it is necessary to prepare a dataset with teacher labels for validation in addition to the training data, this may impose a heavy burden on users.

The present invention has been proposed to solve these problems, and it is hoped that a technology will be provided that makes it possible to build a highly accurate neural network without having to prepare a data set with teacher labels for validation purposes in addition to the training data.

In order to solve the above problem, according to an aspect of the present invention, there is provided an information processing device comprising: an input unit that acquires unlabeled data and labeled data to which a teacher label has been attached; an estimation unit that outputs an inference result corresponding to the unlabeled data and an inference result corresponding to the labeled data based on the unlabeled data, the labeled data, and a neural network; a coefficient generation unit that calculates a coefficient indicating a degree of influence of the unlabeled data based on the inference result corresponding to the unlabeled data; an unlabeled data evaluation unit that outputs an evaluation result corresponding to the unlabeled data based on the inference result corresponding to the unlabeled data and the coefficient; a labeled data evaluation unit that outputs an evaluation result corresponding to the labeled data based on the inference result corresponding to the labeled data and the teacher label; and an update unit that updates weight parameters of the neural network based on the evaluation result corresponding to the unlabeled data and the evaluation result corresponding to the labeled data.
In the technology according to the embodiment of the present invention, unlabeled data and labeled data are used for learning. Therefore, the technology according to the embodiment of the present invention corresponds to a technology related to semi-supervised machine learning.

The inference result corresponding to the unlabeled data may include an inference value for each class, and the coefficient generation unit may calculate a difference between the inference value of a first class having a maximum inference value for each unlabeled data and the inference values of one or more classes different from the first class, and calculate the coefficient for each unlabeled data based on the difference.

The coefficient generation unit may calculate the difference between the inferred value of the first class and the inferred value of the second class having the second largest inferred value for each of the unlabeled data.

The coefficient generation unit may calculate the difference between the inferred value of the first class and the average value of inferred values of multiple classes different from the first class for each of the unlabeled data.

The coefficient generation unit may calculate, for each of the unlabeled data, a number that has a positive correlation with the difference as the coefficient.

The inference result corresponding to the unlabeled data may include an inference value for each class, and the coefficient generation unit may identify a predicted class based on the inference value for each unlabeled data, calculate the number of unlabeled data corresponding to the predicted class as a frequency of the predicted class, and calculate the coefficient based on the frequency of the predicted class for each unlabeled data.

The coefficient generation unit may identify, for each of the unlabeled data, the class for which the inference value is maximum as the predicted class.

The coefficient generation unit may calculate, for each of the unlabeled data, a number that has a negative correlation with the frequency of the predicted class as the coefficient.

The unlabeled data evaluation unit may output an evaluation result corresponding to the unlabeled data based on multiplying a loss based on an inference result corresponding to the unlabeled data by the coefficient.

According to another aspect of the present invention, there is provided an information processing method executed by a computer, comprising: acquiring unlabeled data and labeled data to which a teacher label has been attached; outputting an inference result corresponding to the unlabeled data and an inference result corresponding to the labeled data based on the unlabeled data, the labeled data, and a neural network; calculating a coefficient indicating a degree of influence of the unlabeled data based on the inference result corresponding to the unlabeled data; outputting an evaluation result corresponding to the unlabeled data based on the inference result corresponding to the unlabeled data and the coefficient; outputting an evaluation result corresponding to the labeled data based on the inference result corresponding to the labeled data and the teacher label; and updating weight parameters of the neural network based on the evaluation result corresponding to the unlabeled data and the evaluation result corresponding to the labeled data.

According to another aspect of the present invention, a program is provided that causes a computer to function as an information processing device including an input unit that acquires unlabeled data and labeled data to which a teacher label has been attached, an estimation unit that outputs an inference result corresponding to the unlabeled data and an inference result corresponding to the labeled data based on the unlabeled data, the labeled data, and a neural network, a coefficient generation unit that calculates a coefficient indicating the influence of the unlabeled data based on the inference result corresponding to the unlabeled data, an unlabeled data evaluation unit that outputs an evaluation result corresponding to the unlabeled data based on the inference result corresponding to the unlabeled data and the coefficient, a labeled data evaluation unit that outputs an evaluation result corresponding to the labeled data based on the inference result corresponding to the labeled data and the teacher label, and an update unit that updates weight parameters of the neural network based on the evaluation result corresponding to the unlabeled data and the evaluation result corresponding to the labeled data.

As described above, the present invention provides technology that makes it possible to build a highly accurate neural network without having to prepare a dataset with teacher labels for validation in addition to the training data.

1 is a diagram illustrating an example of a functional configuration of a learning device according to a first embodiment of the present invention. 13 is a flowchart showing an example of an operation in a learning stage executed by the learning device according to the embodiment. FIG. 2 is a diagram illustrating a hardware configuration of an information processing device as an example of a learning device.

The preferred embodiment of the present invention will be described in detail below with reference to the attached drawings. Note that in this specification and the drawings, components having substantially the same functional configuration are designated by the same reference numerals to avoid redundant description.

In addition, in this specification and drawings, multiple components having substantially the same functional configuration may be distinguished by adding different numbers after the same reference symbol. However, if there is no particular need to distinguish between multiple components having substantially the same functional configuration, only the same reference symbol will be used. In addition, similar components in different embodiments may be distinguished by adding different letters after the same reference symbol. However, if there is no particular need to distinguish between similar components in different embodiments, only the same reference symbol will be used.

(0. Overview of the embodiment)
An overview of an embodiment of the present invention will be described. In the embodiment of the present invention, an information processing device that trains a neural network (hereinafter, also referred to as a "learning device") will be described. In the learning device, the neural network is trained based on training data (learning stage). After that, in the classification device, an estimated label is output based on the trained neural network and classification data (test data).

In the embodiment of the present invention, it is mainly assumed that the learning device and the identification device are realized by the same computer. However, the learning device and the identification device may be realized by different computers. In such a case, the trained neural network generated by the learning device is provided to the identification device. For example, the trained neural network may be provided from the learning device to the identification device via a recording medium or via communication. The "learning stage" executed in the learning device is described below.

1. First embodiment
First, a first embodiment of the present invention will be described. In the first embodiment of the present invention, semi-supervised learning is performed by a learning device.

(Configuration of the learning device)
An example of the configuration of a learning device according to a first embodiment of the present invention will be described with reference to Fig. 1. Fig. 1 is a diagram showing an example of the functional configuration of a learning device 10 according to a first embodiment of the present invention. As shown in Fig. 1, the learning device 10 according to the first embodiment of the present invention includes an input unit 111, an estimation unit 122, a coefficient generation unit 131, an unlabeled data evaluation unit 132, a labeled data evaluation unit 133, and an update unit 134.

In the first embodiment of the present invention, it is mainly assumed that the estimation unit 122 is included in the neural network 120. That is, the estimation unit 122 is configured by connecting computation graphs constructed by neurons in the order of processing, and can be regarded as one neural network as a whole. In the following, the neural network is also written as "NN". In more detail, the estimation unit 122 may mainly include a convolution layer and a pooling layer. In the following, it is mainly assumed that a two-dimensional convolution layer is used as the convolution layer, but a three-dimensional convolution layer may also be used.

In addition to the estimation unit 122, the input unit 111, the coefficient generation unit 131, the unlabeled data evaluation unit 132, the labeled data evaluation unit 133, and the update unit 134 include a calculation device such as a CPU (Central Processing Unit) or a GPU (Graphics Processing Unit), and the functions can be realized by the calculation device expanding a program stored in a ROM (Read Only Memory) into a RAM and executing it. At this time, a computer-readable recording medium on which the program is recorded can also be provided. Alternatively, these blocks may be composed of dedicated hardware or a combination of multiple hardware components. Data necessary for the calculation by the calculation device is appropriately stored in a storage unit (not shown).

The unlabeled dataset 101, the labeled dataset 102, and the weight parameters 121 are stored in a storage unit (not shown). Such a storage unit may be configured with a memory such as a RAM (Random Access Memory), a hard disk drive, or a flash memory.

In the initial state, an initial value is set for the weight parameter 121. For example, the initial value set for the weight parameter 121 may be a random value, but may be any value. For example, the initial value set for the weight parameter 121 may be a learned value obtained in advance by learning.

(Unlabeled Dataset 101)
The unlabeled dataset 101 includes a plurality of learning data (input data) each of which is not associated with a teacher label. Hereinafter, learning data that is not associated with a teacher label is also referred to as "unlabeled data". Note that in the embodiment of the present invention, it is mainly assumed that the unlabeled data is image data (particularly, still image data). However, the type of unlabeled data is not particularly limited, and data other than still image data may be used as unlabeled data. For example, the unlabeled data may be video data including a plurality of frames, or may be time-series data or audio data.

(Labeled Dataset 102)
The labeled data set 102 includes a plurality of pieces of learning data (input data) and teacher labels associated with each of the plurality of pieces of learning data. Hereinafter, the learning data associated with the teacher labels is also referred to as "labeled data." In addition, the combination of the teacher labels and the labeled data is also referred to as "labeled data." The teacher labels are assigned manually or by a function not shown. Note that, like the types of unlabeled data, the types of labeled data are not particularly limited.

(Input unit 111)
The input unit 111 sequentially acquires unlabeled data from the unlabeled dataset 101, creates mini-batches based on the acquired unlabeled data, and outputs the created mini-batches to the estimation unit 122 of the neural network 120. Furthermore, the input unit 111 sequentially acquires labeled data (combinations of teacher labels and labeled data) from the labeled dataset 102, creates mini-batches based on the acquired labeled data, and outputs the created mini-batches to the estimation unit 122 of the neural network 120. The size of the mini-batches is not particularly limited.

(Estimation unit 122)
The estimation unit 122 obtains an inference result corresponding to the unlabeled data based on the unlabeled data included in the mini-batch output from the input unit 111 and the neural network 120. More specifically, the estimation unit 122 obtains data output from the neural network 120 as an inference result corresponding to the unlabeled data based on inputting the unlabeled data to the neural network 120 to which the weight parameters 121 are set. The estimation unit 122 outputs the inference result corresponding to the unlabeled data to the coefficient generation unit 131.

At this time, the estimation unit 122 may output two types of labels based on a semi-supervised learning framework to the coefficient generation unit 131 as an inference result corresponding to the unlabeled data. Here, the algorithm for obtaining the two types of labels is not limited to a specific algorithm, and an algorithm used in semi-supervised learning may be used.

For example, the input unit 111 may obtain two types of unlabeled data based on the unlabeled data acquired from the unlabeled dataset 101. As an example, the input unit 111 may obtain two types of unlabeled data by performing two types of data extension on the unlabeled data. In this case, the input unit 111 outputs the two types of unlabeled data to the estimation unit 122, and the estimation unit 122 outputs labels corresponding to the two types of unlabeled data as two types of labels to the coefficient generation unit 131.

Alternatively, one type of unlabeled data may be output from the input unit 111 to the estimation unit 122, and two types of weight parameters may be used in the estimation unit 122. As an example, the estimation unit 122 may obtain two types of labels, namely data obtained by applying all of the weight parameters 121 to the unlabeled data output from the input unit 111, and data obtained by applying part of the weight parameters 121. In this case, the estimation unit 122 outputs the two types of labels to the coefficient generation unit 131.

For example, the estimation unit 122 outputs one of the two types of labels to the coefficient generation unit 131 as a pseudo teacher label corresponding to unlabeled data, and the other as an estimated label corresponding to unlabeled data. Note that there is no limitation on which of the two types of labels is to be the pseudo teacher label. For example, the label obtained by weaker data augmentation may be the pseudo teacher label. Alternatively, the label obtained by applying all of the weight parameters 121 may be the pseudo teacher label.

Furthermore, the estimation unit 122 obtains an inference result corresponding to the labeled data based on the labeled data included in the mini-batch output from the input unit 111 and the neural network 120. More specifically, the estimation unit 122 obtains the data output from the neural network 120 as an inference result corresponding to the labeled data based on inputting the labeled data to the neural network 120 to which the weight parameter 121 is set. The estimation unit 122 outputs the inference result corresponding to the labeled data to the labeled data evaluation unit 133.

The format of the inference result output from the estimation unit 122 is not particularly limited. However, it is preferable that the format of the inference result output from the estimation unit 122 is set in accordance with the format of the teacher label. For example, if the teacher label indicates the class of the classification problem and is a one-hot vector having a length equal to the number of classes, the format of the inference result output from the estimation unit 122 may also be a vector having a length equal to the number of classes. In this case, the inference result output from the estimation unit 122 may include a value for each class (hereinafter also referred to as an "inference value").

As an example, when the inference values of all classes are adjusted by the inference unit 122 to sum to 1, the inference value corresponding to each class may correspond to the probability corresponding to each class. However, the sum of the inference values of all classes does not have to be adjusted by the inference unit 122 to sum to 1. In either case, the inference value output from the inference unit 122 may be larger the higher the likelihood of the class.

(Coefficient Generation Unit 131)
The coefficient generation unit 131 calculates a coefficient indicating the influence of the unlabeled data based on the inference result corresponding to the unlabeled data output from the estimation unit 122. More specifically, the coefficient generation unit 131 specifies a predicted class based on the inference value for each piece of unlabeled data included in the mini-batch. If the batch size is B, the predicted class based on the inference value of the unlabeled data x ^u = {x ₁ ^u , ..., x _B ^u } is specified as y ^u = {y ₁ ^u , ..., y _B ^u }. Each element of y ^u may be a predicted class number.

Here, the predicted class may be identified in any manner. As an example, the class with the largest inference value is considered to be the most likely class. Therefore, the coefficient generation unit 131 may identify the class with the largest inference value for each piece of unlabeled data included in the mini-batch as the predicted class. For example, either of two types of labels may be used as the inference value used to identify the predicted class, but it is preferable to use a pseudo teacher label.

Then, the coefficient generating unit 131 calculates the number of unlabeled data corresponding to the predicted class as the frequency of the predicted class. For example, when the neural network 120 solves a classification problem into N classes (i.e., when the inference result corresponding to the unlabeled data includes inference values for N classes), the frequency c _i of the predicted class can be expressed as the following formula (1).

The coefficient generating unit 131 calculates a coefficient t for each unlabeled data x ^u included in the mini-batch based on the frequency c of the predicted class. As an example, it is preferable that the coefficient generating unit 131 calculates a number having a negative correlation with the frequency c of the predicted class for each unlabeled data x ^u included in the mini-batch as the coefficient t. As a result, the unlabeled data corresponding to a predicted class with a smaller frequency c is treated as having a higher degree of influence.

For example, if f is a function that returns an output value that shows a negative correlation with respect to an input value, then the function f that returns an output value that shows a negative correlation with respect to an input value can be expressed as in the following equation (2).

For example, the following formula (3) is an example of a function f that returns an output value that shows a negative correlation with an input value.

The coefficient generation unit 131 outputs the inference results and coefficients for each unlabeled data included in the mini-batch to the unlabeled data evaluation unit 132.

(Unlabeled Data Evaluation Unit 132)
The unlabeled data evaluation unit 132 obtains an evaluation result corresponding to the unlabeled data based on the inference result and the coefficient for each piece of unlabeled data included in the mini-batch. More specifically, the unlabeled data evaluation unit 132 calculates a loss for each piece of unlabeled data included in the mini-batch based on the inference result, and obtains an evaluation result corresponding to the unlabeled data based on the loss and the coefficient for each piece of unlabeled data.

First, the unlabeled data evaluation unit 132 evaluates the estimated labels for each piece of unlabeled data based on the pseudo teacher labels and calculates the loss.

Here, the loss function used to calculate the loss is not limited to a specific function, and a loss function similar to that used in a general neural network may be used. For example, the loss function may be a mean square error based on the difference between the pseudo teacher label corresponding to the unlabeled data and the estimated label corresponding to the unlabeled data, or a cross entropy error based on the difference between the pseudo teacher label corresponding to the unlabeled data and the estimated label corresponding to the unlabeled data.

Next, the unlabeled data evaluation unit 132 obtains an evaluation result corresponding to the unlabeled data based on the loss and coefficient for each piece of unlabeled data. More specifically, the unlabeled data evaluation unit 132 obtains an evaluation result corresponding to the unlabeled data based on multiplying the loss and the coefficient for each piece of unlabeled data.

As an example, if the unlabeled data is x ^u , the weight parameter 121 is θ, and the coefficient calculated by the coefficient generation unit 131 is t, the evaluation result l _u corresponding to the unlabeled data can be calculated by summing up the multiplication results of the loss and the coefficient for each unlabeled data in a mini-batch, as shown in the following equation (4).

The unlabeled data evaluation unit 132 outputs the evaluation results corresponding to the unlabeled data to the update unit 134.

(Labeled Data Evaluation Unit 133)
The labeled data evaluation unit 133 evaluates each piece of labeled data included in the mini-batch based on the teacher label corresponding to the labeled data to obtain an evaluation result for each piece of labeled data. More specifically, the labeled data evaluation unit 133 calculates a loss based on the teacher label corresponding to the labeled data and the labeled data, and obtains an evaluation result corresponding to the labeled data based on the loss for each piece of labeled data.

First, the labeled data evaluation unit 133 evaluates each piece of labeled data based on the teacher label corresponding to the labeled data and calculates the loss.

Here, the loss function is not limited to a specific function, and a loss function similar to that used in a general neural network may be used. For example, the loss function may be a mean square error based on the difference between the labeled data and the teacher label corresponding to the labeled data, or a cross entropy error based on the difference between the labeled data and the teacher label corresponding to the labeled data.

Next, the labeled data evaluation unit 133 obtains an evaluation result corresponding to the labeled data based on the loss for each labeled data. More specifically, the labeled data evaluation unit 133 obtains an evaluation result corresponding to the labeled data by summing the losses for each labeled data in a mini-batch.

As an example, if labeled data is ^xt , a teacher label corresponding to the labeled data is ^xt , and a weight parameter 121 is θ, then the evaluation result l _s corresponding to the labeled data can be expressed as shown in the following equation (5).

The labeled data evaluation unit 133 outputs the evaluation result corresponding to the labeled data to the update unit 134.

(Update unit 134)
The update unit 134 updates the weighting parameters 121 based on the evaluation result corresponding to the unlabeled data output from the unlabeled data evaluation unit 132 and the evaluation result corresponding to the labeled data output from the labeled data evaluation unit 133. This allows the weighting parameters 121 to be trained so that the estimated labels corresponding to the unlabeled data approach the pseudo teacher labels corresponding to the unlabeled data and so that the labeled data approach the teacher labels corresponding to the labeled data.

For example, the update unit 134 may update the weight parameter 121 based on a weighted sum of the evaluation results corresponding to the labeled data and the evaluation results corresponding to the unlabeled data (hereinafter, simply referred to as the "weighted sum"). More specifically, the update unit 134 may update the weight parameter 121 by backpropagation based on the weighted sum of the evaluation results corresponding to the labeled data and the evaluation results corresponding to the unlabeled data.

The weighted sum may be expressed in any way. As an example, if the evaluation result corresponding to the labeled data is l _s as shown in formula (5), the evaluation result corresponding to the unlabeled data is l _u as shown in formula (4), and the hyperparameter for taking the weighted sum is λ, the weighted sum L can be expressed as shown in formula (6) below.

Note that the update unit 134 judges whether or not the learning termination condition has been satisfied each time the weight parameter 121 is updated. If it is judged that the learning termination condition has not been satisfied, the input unit 111 acquires the next input data (a combination of labeled data and teacher label, and unlabeled data), and the estimation unit 122, the coefficient generation unit 131, the unlabeled data evaluation unit 132, the labeled data evaluation unit 133, and the update unit 134 each re-execute their own processing based on the next input data. On the other hand, if it is judged that the learning termination condition has been satisfied, the learning is terminated.

The learning end condition is not particularly limited, and may be any condition that indicates that the neural network 120 has been learned to a certain extent. Specifically, the learning end condition may include a condition that the value of the weighted sum is smaller than a threshold value. Alternatively, the learning end condition may include a condition that the change in the value of the weighted sum is smaller than a threshold value (a condition that the value of the weighted sum has converged). Alternatively, the learning end condition may include a condition that the weight parameter 121 has been updated a predetermined number of times. Alternatively, when the accuracy of the neural network 120 (e.g., accuracy rate, etc.) is calculated, the learning end condition may include a condition that the accuracy exceeds a predetermined percentage (e.g., 90%).

The above describes an example of the configuration of a learning device according to the first embodiment of the present invention.

(Learning stage operation)
Next, a flow of operations in a "learning stage" executed by the learning device 10 according to the first embodiment of the present invention will be described with reference to Fig. 2. Fig. 2 is a flowchart showing an example of operations in a learning stage executed by the learning device 10 according to the first embodiment of the present invention.

First, the input unit 111 creates a mini-batch by acquiring unlabeled data of the batch size from the unlabeled dataset 101, and outputs the created mini-batch to the estimation unit 122 of the neural network 120 (S101).

Then, the estimation unit 122 obtains an inference result corresponding to the unlabeled data based on the unlabeled data included in the mini-batch created by the input unit 111 and the neural network 120 (S102). The estimation unit 122 outputs the inference result corresponding to the unlabeled data to the coefficient generation unit 131.

At this time, the estimation unit 122 may output two types of labels to the coefficient generation unit 131 as inference results corresponding to the unlabeled data. For example, the estimation unit 122 outputs one of the two types of labels to the coefficient generation unit 131 as a pseudo teacher label corresponding to the unlabeled data, and the other as an estimated label corresponding to the unlabeled data.

The coefficient generation unit 131 calculates a coefficient indicating the influence of the unlabeled data based on the inference result corresponding to the unlabeled data output from the estimation unit 122 (S103). More specifically, the coefficient generation unit 131 identifies a predicted class based on the inference value for each piece of unlabeled data included in the mini-batch. As an example, the coefficient generation unit 131 may identify the class with the maximum inference value for each piece of unlabeled data included in the mini-batch as the predicted class. For example, either of two types of labels may be used as the inference value used to identify the predicted class, but it is preferable to use a pseudo teacher label.

Then, the coefficient generation unit 131 calculates the number of unlabeled data corresponding to the predicted class as the frequency of the predicted class. The coefficient generation unit 131 calculates a coefficient for each unlabeled data included in the mini-batch based on the frequency of the predicted class. As an example, it is desirable for the coefficient generation unit 131 to calculate a number that has a negative correlation with the frequency of the predicted class for each unlabeled data included in the mini-batch as the coefficient. As a result, the unlabeled data corresponding to a predicted class with a smaller frequency is treated as having a higher degree of influence.

The coefficient generation unit 131 outputs the inference results and coefficients for each unlabeled data included in the mini-batch to the unlabeled data evaluation unit 132.

The unlabeled data evaluation unit 132 obtains an evaluation result corresponding to the unlabeled data based on the inference result and coefficient for each piece of unlabeled data included in the mini-batch (S104). More specifically, the unlabeled data evaluation unit 132 calculates a loss for each piece of unlabeled data included in the mini-batch based on the inference result, and obtains an evaluation result corresponding to the unlabeled data based on the loss and coefficient for each piece of unlabeled data. The unlabeled data evaluation unit 132 outputs the evaluation result corresponding to the unlabeled data to the update unit 134.

Then, the input unit 111 creates a mini-batch by acquiring labeled data of the batch size from the labeled dataset 102 as a combination of teacher labels and labeled data, and outputs the created mini-batch to the estimation unit 122 of the neural network 120 (S105).

Then, the estimation unit 122 obtains an inference result corresponding to the labeled data based on the labeled data included in the mini-batch created by the input unit 111 and the neural network 120 (S106). The estimation unit 122 outputs the inference result corresponding to the labeled data to the labeled data evaluation unit 133.

For each piece of labeled data included in the mini-batch, the labeled data evaluation unit 133 evaluates the labeled data based on the teacher label corresponding to the labeled data to obtain an evaluation result for each piece of labeled data (S107). In more detail, the labeled data evaluation unit 133 calculates a loss based on the teacher label corresponding to the labeled data and the labeled data, and obtains an evaluation result corresponding to the labeled data based on the loss for each piece of labeled data. The labeled data evaluation unit 133 outputs the evaluation result corresponding to the labeled data to the update unit 134.

The update unit 134 updates the weighting parameters 121 based on the evaluation results corresponding to the unlabeled data output from the unlabeled data evaluation unit 132 and the evaluation results corresponding to the labeled data output from the labeled data evaluation unit 133 (S108). This allows the weighting parameters 121 to be trained so that the estimated labels corresponding to the unlabeled data approach the pseudo teacher labels corresponding to the unlabeled data, and so that the labeled data approach the teacher labels corresponding to the labeled data.

For example, the update unit 134 may update the weight parameter 121 based on a weighted sum of the evaluation result corresponding to the labeled data and the evaluation result corresponding to the unlabeled data. More specifically, the update unit 134 may update the weight parameter 121 by backpropagation based on the weighted sum of the evaluation result corresponding to the labeled data and the evaluation result corresponding to the unlabeled data.

Each time the update unit 134 finishes updating the weight parameter 121, it determines whether the learning termination condition has been satisfied (S109). If it is determined that the learning termination condition has not been satisfied ("NO" in S109), the operation proceeds to S101, the input unit 111 acquires the next input data, and the estimation unit 122, the coefficient generation unit 131, the unlabeled data evaluation unit 132, the labeled data evaluation unit 133, and the update unit 134 each execute their own processing based on the next input data again. On the other hand, if it is determined that the learning termination condition has been satisfied, the learning is terminated.

The above describes the flow of operations in the "learning stage" performed by the learning device 10 according to the first embodiment of the present invention.

(Summary of the first embodiment)
As described above, according to the first embodiment of the present invention, a pseudo predicted class is identified for each piece of unlabeled data based on the inference result corresponding to the unlabeled data. Then, the degree of influence on the loss is automatically determined for each piece of unlabeled data based on the pseudo predicted class.

This makes it possible to reduce the impact of a phenomenon that can occur during the learning stage (especially the early stages of learning) (i.e., the phenomenon in which inference results are concentrated in a specific class) on the loss of the class in which the inference results are concentrated. As a result, the effect of stable learning can be enjoyed.

In addition, according to the first embodiment of the present invention, the impact of unlabeled data on loss can be determined without relying on a semi-supervised learning algorithm. Furthermore, according to the first embodiment of the present invention, even if there is unlabeled data whose impact on loss is relatively small, the impact is used as a coefficient for learning, making it possible to perform stable learning while eliminating the need for manual data selection using thresholds, etc.

The above describes the first embodiment of the present invention.

2. Second embodiment
Next, a second embodiment of the present invention will be described. In the second embodiment of the present invention, semi-supervised learning is also performed by a learning device.

As shown in FIG. 1, the learning device 10 according to the second embodiment of the present invention includes an input unit 111, an estimation unit 122, a coefficient generation unit 131, an unlabeled data evaluation unit 132, a labeled data evaluation unit 133, and an update unit 134, similar to the learning device 10 according to the first embodiment of the present invention.

The learning device 10 according to the second embodiment of the present invention differs from the learning device 10 according to the first embodiment of the present invention mainly in the function of the coefficient generation unit 131. Therefore, the following mainly describes the function of the coefficient generation unit 131, and a detailed description of the functions of the other blocks is omitted.

(Coefficient Generation Unit 131)
The coefficient generating unit 131 calculates a coefficient indicating the influence of the unlabeled data based on the inference result corresponding to the unlabeled data output from the estimation unit 122. More specifically, the coefficient generating unit 131 identifies a first class having a maximum inference value for each piece of unlabeled data included in the mini-batch. Then, the coefficient generating unit 131 calculates a difference based on the inference value of the first class and the inference values of one or more classes different from the first class. For example, either of two types of labels may be used as the inference value used to calculate the difference, but it is preferable to use a pseudo teacher label. The coefficient generating unit 131 calculates a coefficient based on the difference for each piece of unlabeled data.

As an example, the coefficient generating unit 131 may identify a second class having the second largest inference value. In this case, the coefficient generating unit 131 may calculate the difference between the inference value of the first class and the inference value of the second class for each piece of unlabeled data. For example, when the neural network 120 solves a classification problem into N classes (i.e., when the inference result corresponding to the unlabeled data includes inference values for N classes), the difference d _i may be expressed as in the following formula (7).

In equation (7), v represents the inference result corresponding to the unlabeled data and represents a set of inference values corresponding to each class. argmax is a function that returns the class number corresponding to the inference value passed as an argument as an output value.

The method of calculating the difference based on the inferred value of the first class and the inferred values of one or more classes different from the first class is not limited. For example, the coefficient generating unit 131 may calculate the difference between the inferred value of the first class and the average value of the inferred values of multiple classes different from the first class for each piece of unlabeled data. In this case, if the function that takes the average value is ave, the difference d _i can be expressed as the following formula (8).

The coefficient generation unit 131 calculates a coefficient t for each piece of unlabeled data included in the mini-batch based on the difference d. As an example, it is preferable that the coefficient generation unit 131 calculates a number that has a positive correlation with the difference d as the coefficient t for each piece of unlabeled data included in the mini-batch. This allows unlabeled data that corresponds to a larger difference to be treated as having a higher degree of influence.

For example, if a function that returns an output value that shows a positive correlation with an input value is g, the function g that returns an output value that shows a positive correlation with an input value can be expressed as the following equation (9).

For example, the following formula (10) is an example of a function g that returns an output value that shows a positive correlation with an input value.

The coefficient generation unit 131 outputs the inference results and coefficients for each unlabeled data included in the mini-batch to the unlabeled data evaluation unit 132.

The above describes an example of the configuration of a learning device according to the second embodiment of the present invention.

(Learning stage operation)
Next, the flow of operations in the "learning stage" performed by the learning device 10 according to the second embodiment of the present invention will be described with reference to Fig. 2. The operations in the "learning stage" performed by the learning device 10 according to the second embodiment of the present invention differ from the operations in the "learning stage" performed by the learning device 10 according to the first embodiment of the present invention mainly in the operations of the coefficient generation unit 131. Therefore, the following mainly describes the operations of the coefficient generation unit 131, and detailed descriptions of other operations are omitted.

In the second embodiment of the present invention, steps S101 to S102 are executed in the same manner as in the first embodiment of the present invention. Next, the coefficient generation unit 131 calculates a coefficient indicating the degree of influence of the unlabeled data based on the inference result corresponding to the unlabeled data output from the estimation unit 122 (S103).

More specifically, the coefficient generation unit 131 identifies a first class that has the maximum inference value for each piece of unlabeled data included in the mini-batch. Then, the coefficient generation unit 131 calculates a difference based on the inference value of the first class and the inference values of one or more classes different from the first class. For example, either of two types of labels may be used as the inference value used to calculate the difference, but it is preferable to use a pseudo teacher label. The coefficient generation unit 131 calculates a coefficient based on the difference for each piece of unlabeled data.

As an example, the coefficient generation unit 131 may identify a second class having the second largest inferred value. In this case, the coefficient generation unit 131 may calculate the difference between the inferred value of the first class and the inferred value of the second class for each piece of unlabeled data. Alternatively, the coefficient generation unit 131 may calculate the difference between the inferred value of the first class and the average value of the inferred values of multiple classes different from the first class for each piece of unlabeled data.

The coefficient generation unit 131 calculates a coefficient based on the difference for each piece of unlabeled data included in the mini-batch. As an example, it is preferable that the coefficient generation unit 131 calculates a number that has a positive correlation with the difference for each piece of unlabeled data included in the mini-batch as the coefficient. This allows unlabeled data that corresponds to a larger difference to be treated as having a higher degree of influence.

The coefficient generation unit 131 outputs the inference results and coefficients for each unlabeled data included in the mini-batch to the unlabeled data evaluation unit 132.

In the second embodiment of the present invention, steps S104 to S109 are executed in the same manner as in the first embodiment of the present invention.

The above describes the flow of operations in the "learning stage" performed by the learning device 10 according to the second embodiment of the present invention.

(Summary of the second embodiment)
As described above, according to the second embodiment of the present invention, a difference is calculated for each piece of unlabeled data based on an inference result corresponding to the unlabeled data, between an inferred value of a first class having a maximum inferred value and inferred values of one or more classes different from the first class. Then, a degree of influence on loss is automatically determined for each piece of unlabeled data based on the difference.

This allows the same effects as those of the first embodiment of the present invention to be obtained. Furthermore, according to the second embodiment of the present invention, even if the batch size is relatively small, the degree of influence on the loss can be determined with high accuracy. Therefore, according to the second embodiment of the present invention, stable learning is possible even if the batch size is relatively small.

The second embodiment of the present invention has been described above.

(3. Hardware Configuration Example)
Next, a hardware configuration example of the learning device 10 according to the first embodiment of the present invention will be described. Note that the hardware configuration of the learning device 10 according to the second embodiment of the present invention can be realized in the same manner as the hardware configuration of the learning device 10 according to the first embodiment of the present invention.

Below, an example of the hardware configuration of an information processing device 900 will be described as an example of the hardware configuration of the learning device 10 according to the first embodiment of the present invention. Note that the example of the hardware configuration of the information processing device 900 described below is merely one example of the hardware configuration of the learning device 10. Therefore, the hardware configuration of the learning device 10 may be such that unnecessary components are deleted from the hardware configuration of the information processing device 900 described below, or new components may be added.

Figure 3 is a diagram showing the hardware configuration of an information processing device 900 as an example of a learning device 10 according to the first embodiment of the present invention. The information processing device 900 includes a CPU (Central Processing Unit) 901, a ROM (Read Only Memory) 902, a RAM (Random Access Memory) 903, a host bus 904, a bridge 905, an external bus 906, an interface 907, an input device 908, an output device 909, a storage device 910, and a communication device 911.

The CPU 901 functions as an arithmetic processing device and control device, and controls the overall operation of the information processing device 900 in accordance with various programs. The CPU 901 may also be a microprocessor. The ROM 902 stores programs and arithmetic parameters used by the CPU 901. The RAM 903 temporarily stores programs used in the execution of the CPU 901 and parameters that change appropriately during the execution. These are interconnected by a host bus 904 consisting of a CPU bus, etc.

The host bus 904 is connected to an external bus 906, such as a PCI (Peripheral Component Interconnect/Interface) bus, via a bridge 905. Note that the host bus 904, bridge 905, and external bus 906 do not necessarily need to be configured separately, and these functions may be implemented on a single bus.

The input device 908 is composed of input means for the user to input information, such as a mouse, keyboard, touch panel, button, microphone, switch, and lever, and an input control circuit that generates an input signal based on the user's input and outputs it to the CPU 901. A user who operates the information processing device 900 can input various data to the information processing device 900 and instruct processing operations by operating this input device 908.

The output device 909 includes, for example, display devices such as a CRT (Cathode Ray Tube) display device, a Liquid Crystal Display (LCD) device, an OLED (Organic Light Emitting Diode) device, and a lamp, as well as audio output devices such as a speaker.

The storage device 910 is a device for storing data. The storage device 910 may include a storage medium, a recording device for recording data on the storage medium, a reading device for reading data from the storage medium, and a deleting device for deleting data recorded on the storage medium. The storage device 910 is configured, for example, with a HDD (Hard Disk Drive). This storage device 910 drives the hard disk and stores the programs executed by the CPU 901 and various data.

The communication device 911 is, for example, a communication interface configured with a communication device for connecting to a network. The communication device 911 may support either wireless communication or wired communication.

The above describes an example of the hardware configuration of the learning device 10 according to the first embodiment of the present invention.

(4. Summary)
Although the preferred embodiment of the present invention has been described in detail above with reference to the accompanying drawings, the present invention is not limited to such an example. It is clear that a person having ordinary knowledge in the technical field to which the present invention pertains can conceive of various modified or altered examples within the scope of the technical ideas described in the claims, and it is understood that these also naturally belong to the technical scope of the present invention.

In the first embodiment and the second embodiment of the present invention, the case where the learning data is image data (particularly, still image data) has been mainly described. However, the type of learning data is not particularly limited. For example, as long as features that match the type of learning data are extracted, data other than still image data may also be used as learning data. For example, the learning data may be video data including multiple frames, or may be audio data.

In this case, when the learning data is still image data, a two-dimensional convolutional layer is generally used as the convolutional layer included in the estimation unit 122. On the other hand, when a three-dimensional convolutional layer is used as the convolutional layer included in the estimation unit 122, video image data can be applied as the learning data.

In the first embodiment of the present invention, the following formula (3) has been given as an example of a function f that returns an output value that shows a negative correlation with an input value. However, the function f that returns an output value that shows a negative correlation with an input value is not limited to this example. For example, the following formulas (3-A) and (3-B) are also examples of a function f that returns an output value that shows a negative correlation with an input value.

In the second embodiment of the present invention, the following formula (9) has been given as an example of function g that returns an output value that shows a positive correlation with an input value. However, function g that returns an output value that shows a positive correlation with an input value is not limited to this example. For example, the following formulas (9-A) and (9-B) are also examples of function g that returns an output value that shows a positive correlation with an input value.

REFERENCE SIGNS LIST 10 Learning device 101 Unlabeled data set 102 Labeled data set 111 Input section 120 Neural network 121 Weight parameters 122 Estimation section 131 Coefficient generation section 132 Unlabeled data evaluation section 133 Labeled data evaluation section 134 Update section

Claims (11)

An input unit that acquires unlabeled data and labeled data to which a teacher label is attached;
an inference unit that outputs an inference result corresponding to the unlabeled data and an inference result corresponding to the labeled data based on the unlabeled data, the labeled data, and a neural network;
a coefficient generation unit that calculates a coefficient indicating an influence of the unlabeled data based on an inference result corresponding to the unlabeled data;
an unlabeled data evaluation unit that outputs an evaluation result corresponding to the unlabeled data based on an inference result corresponding to the unlabeled data and the coefficients;
a labeled data evaluation unit that outputs an evaluation result corresponding to the labeled data based on an inference result corresponding to the labeled data and the teacher label;
an update unit that updates weight parameters of the neural network based on an evaluation result corresponding to the unlabeled data and an evaluation result corresponding to the labeled data;
An information processing device comprising: the inference results corresponding to the unlabeled data include an inference value for each class;
the coefficient generation unit calculates a difference between an inferred value of a first class having a maximum inferred value for each unlabeled data and an inferred value of one or more classes different from the first class, and calculates the coefficient for each unlabeled data based on the difference;
The information processing device according to claim 1 . the coefficient generation unit calculates, for each of the unlabeled data, a difference between an inferred value of the first class and an inferred value of a second class having a second largest inferred value;
The information processing device according to claim 2 . the coefficient generation unit calculates, for each of the unlabeled data, a difference between the inferred value of the first class and an average value of inferred values of a plurality of classes different from the first class;
The information processing device according to claim 2 . the coefficient generation unit calculates, for each of the unlabeled data, a number having a positive correlation with the difference as the coefficient;
The information processing device according to any one of claims 2 to 4. the inference results corresponding to the unlabeled data include an inference value for each class;
the coefficient generation unit specifies a predicted class based on an inference value for each unlabeled data, calculates a number of unlabeled data corresponding to the predicted class as a frequency of the predicted class, and calculates the coefficient based on the frequency of the predicted class for each of the unlabeled data.
The information processing device according to claim 1 . The coefficient generation unit identifies a class for which the inference value is maximum as a predicted class for each of the unlabeled data.
The information processing device according to claim 6 . the coefficient generation unit calculates, for each of the unlabeled data, a number having a negative correlation with a frequency of the predicted class as the coefficient;
8. The information processing device according to claim 6 or 7. the unlabeled data evaluation unit outputs an evaluation result corresponding to the unlabeled data by multiplying a loss based on an inference result corresponding to the unlabeled data by the coefficient.
The information processing device according to any one of claims 2 to 8. Obtaining unlabeled data and labeled data to which a teacher label is attached;
outputting an inference result corresponding to the unlabeled data and an inference result corresponding to the labeled data based on the unlabeled data, the labeled data, and a neural network;
calculating a coefficient indicating an influence of the unlabeled data based on an inference result corresponding to the unlabeled data;
outputting an evaluation result corresponding to the unlabeled data based on the inference result corresponding to the unlabeled data and the coefficients;
outputting an evaluation result corresponding to the labeled data based on an inference result corresponding to the labeled data and the teacher label;
updating weight parameters of the neural network based on the evaluation results corresponding to the unlabeled data and the evaluation results corresponding to the labeled data;
13. A computer-implemented information processing method comprising: Computer,
An input unit that acquires unlabeled data and labeled data to which a teacher label is attached;
an inference unit that outputs an inference result corresponding to the unlabeled data and an inference result corresponding to the labeled data based on the unlabeled data, the labeled data, and a neural network;
a coefficient generation unit that calculates a coefficient indicating an influence of the unlabeled data based on an inference result corresponding to the unlabeled data;
an unlabeled data evaluation unit that outputs an evaluation result corresponding to the unlabeled data based on an inference result corresponding to the unlabeled data and the coefficients;
a labeled data evaluation unit that outputs an evaluation result corresponding to the labeled data based on an inference result corresponding to the labeled data and the teacher label;
an update unit that updates weight parameters of the neural network based on an evaluation result corresponding to the unlabeled data and an evaluation result corresponding to the labeled data;
A program that causes the information processing device to function as an information processing device having the above-mentioned.

Images