Machine learning device, machine learning method, and machine learning program

Abstract

To provide a machine learning technique capable of enhancing classification accuracy.SOLUTION: An inference unit 40 infers a label of unlabeled data using a trained model stored in a trained model storage unit 30. An evaluation unit 50 evaluates a probability indicative of likelihood of each candidate of an inference result and, if the inference result is split between a first candidate and a second candidate, attaches a label of the first candidate to the unlabeled data to generate labeled data. A learning unit 20 retrains the trained model using the labeled data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to machine learning techniques. [Background technology]

[0002] Unsupervised learning has the advantage that it does not require labeling of data, so data preparation is not time-consuming and large amounts of data can be used for learning, but it has the disadvantage that learning efficiency is lower than that of supervised learning. Therefore, semi-supervised learning has been proposed, which allows simple and efficient learning by utilizing a small amount of labeled data and a large amount of unlabeled data.

[0003] Patent Document 1 discloses an information processing device that performs semi-supervised learning of an image recognizer, which is an information converter that includes a feature extractor but does not include a generator and performs recognition processing of an input image, while further reducing the possibility of competition between learning processes. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2019-207561 A Summary of the Invention [Problem to be solved by the invention]

[0005] The classifier-based bootstrap method in semi-supervised learning is a technique in which a trained classifier is used to make inferences on unlabeled data, and the results of inferences with high confidence are added to labeled data to proceed with learning. Here, the inference data with high confidence is highly likely to be data similar to the labeled data used during learning. It is known that learning more data that is difficult to classify is effective in improving classification accuracy, and using inference data with high confidence similar to the labeled data for learning does not improve classification accuracy very much.

[0006] The present invention has been made in view of the above circumstances, and has an object to provide a machine learning technique capable of improving classification accuracy. [Means for solving the problem]

[0007] In order to solve the above problem, a machine learning device according to one embodiment of the present invention includes an inference unit that infers a label for unlabeled data using a trained model, an evaluation unit that evaluates a probability indicating the likelihood of each candidate in the inference result, and, when the inference result is divided between a first candidate and a second candidate, assigns a label of the first candidate to the unlabeled data to generate labeled data, and a learning unit that re-trains the trained model using the labeled data.

[0008] Another aspect of the present invention is a machine learning method, which includes an inference step of inferring a label of unlabeled data using a trained model, an evaluation step of evaluating a probability indicating the likelihood of each candidate of the inference result, and if the inference result is dichotomous between a first candidate and a second candidate, generating labeled data by assigning a label of the first candidate to the unlabeled data, and a training step of retraining the trained model using the labeled data.

[0009] Any combination of the above components, and any transformation of the present invention into a method, device, system, recording medium, computer program, etc., are also effective as aspects of the present invention. Effect of the Invention

[0010] According to the present invention, it is possible to provide a machine learning technique that can improve classification accuracy. [Brief description of the drawings]

[0011] [Figure 1] FIG. 1 is a configuration diagram of a machine learning device according to an embodiment. [Diagram 2] 2 is a flowchart illustrating a machine learning procedure performed by the machine learning device of FIG. 1. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] 1 is a configuration diagram of a machine learning device 100 according to an embodiment. The machine learning device 100 includes an input unit 10, a learning unit 20, a trained model storage unit 30, an inference unit 40, an evaluation unit 50, and an output unit 60.

[0013] The input unit 10 supplies unlabeled data to the inference unit 40, and supplies labeled data supplied from the evaluation unit 50 to the learning unit 20. An example of image recognition will be described here. The data input to the input unit 10 is an image in which an object is to be recognized, such as an image of a person, an image of a motorbike, an image of a car, or an image of a bicycle.

[0014] The trained model storage unit 30 stores an object recognition model with excellent recognition accuracy that has been trained in advance using a labeled pre-training data set.

[0015] The inference unit 40 recognizes objects in an input image, which is unlabeled data, using the trained object recognition model stored in the trained model storage unit 30. The output unit 60 outputs the recognition result of the input image.

[0016] The evaluation unit 50 evaluates the probability indicating the likelihood of each candidate of the inference result of the unlabeled data by the inference unit 40. The evaluation unit 50 generates labeled data by assigning the label of the first candidate to difficult-to-classify data whose inference result is divided into two, the first candidate and the second candidate, and supplies the labeled data to the input unit 10. In the case of image recognition, the evaluation unit 50 evaluates the probability of each candidate of the recognition result of the input image, extracts input images whose recognition result is divided into two, and assigns the label of the first candidate to the images to generate labeled data. For example, if the recognition result of the input image is divided into two, the first candidate bicycle and the second candidate motorbike, the evaluation unit 50 assigns the label of the first candidate bicycle to the input image to generate a labeled image.

[0017] An inference result is said to be split between the first and second candidates, with the probability of the third and subsequent candidates being very low.

[0018] More specifically, the condition for dividing the inference result into two is that the sum of the probability of the first candidate and the probability of the second candidate is greater than a predetermined lower limit A, and the difference or ratio between the probability of the first candidate and the probability of the second candidate is smaller than a predetermined upper limit B. Furthermore, a condition may be added that the probability of the third candidate and below is smaller than a predetermined upper limit C. Here, A, B, and C are generally different values, but B and C may be the same value. For example, the inference result may be that the probability that the input image is a bicycle is 57%, the probability that it is a motorbike is 38%, the probability that it is a car is 5%, and the probability that it is a person is 0%.

[0019] Alternatively, the condition for dividing the inference result into two is that the probability of the first candidate is in the range from a predetermined lower limit D to a predetermined upper limit E, and the sum of the probability of the first candidate and the probability of the second candidate is greater than a predetermined lower limit F. Furthermore, a condition may be added that the probability of the third candidate and below is smaller than a predetermined upper limit G. Here, D, E, F, and G are generally different values ​​from each other, but E and F may be the same value. As an example, the condition may be 50%<probability of the first candidate<70% and the probability of the first candidate+the probability of the second candidate>80%. Alternatively, the condition may be set by determining the range of the probability of the first candidate and the range of the probability of the second candidate, such as 55%<probability of the first candidate<65% and 35%<probability of the second candidate<45%.

[0020] If the inference result of the unlabeled data satisfies the condition for dichotomy, the evaluation unit 50 assigns a first candidate label to the data to generate labeled data and supplies it to the input unit 10. However, if the inference result of the unlabeled data does not satisfy the condition for dichotomy, the data is not effective training data for improving classification accuracy, and therefore is not supplied to the input unit 10.

[0021] The learning unit 20 adds the labeled data supplied from the input unit 10 to the dataset used for pre-learning, re-learns the trained object recognition model stored in the trained model memory unit 30, updates the weights of the object recognition model, and generates a new object recognition model.

[0022] FIG. 2 is a flowchart illustrating a machine learning procedure performed by the machine learning device 100.

[0023] The inference unit 40 inputs an input image of a recognition target provided from the input unit 10 into a trained object recognition model, and recognizes an object in the image (S10).

[0024] The evaluation unit 50 evaluates the inference result by the inference unit 40, and if the inference result is divided between the first candidate and the second candidate (Y in S20), it assigns a label of the first candidate to the input image and supplies it as labeled data to the input unit 10 (S30). If the inference result is not divided (N in S20), it does not generate labeled data, returns to step S10, and inputs the next image.

[0025] The procedure from step S10 to step S30 is repeated until a predetermined number of input images with the same label, for example 30, are accumulated (N in S40). When a predetermined number of input images with the same label, here 30, are accumulated (Y in S40), the learning unit 20 generates a new dataset by adding a predetermined number of labeled images to the dataset used for pre-learning, re-learns the trained object recognition model using the new dataset, and updates the weights of the trained object recognition model (S50). The learning unit 20 generates a new object recognition model and stores it in the trained model storage unit 30 (S60).

[0026] In the above description, transfer learning is performed after a predetermined number of input images with the same label have been accumulated, but learning may be performed each time a labeled image is generated.

[0027] The various processes of the machine learning device 100 described above can be realized not only as a device using hardware such as a CPU and a memory, but also as firmware stored in a ROM (read-only memory) or a flash memory, or as software for a computer, etc. The firmware program or software program can be provided by recording it on a recording medium readable by a computer, etc., or can be transmitted and received with a server via a wired or wireless network, or can be transmitted and received as a data broadcast of terrestrial or satellite digital broadcasting.

[0028] As described above, according to the machine learning device 100 of this embodiment, the classification accuracy can be improved by additionally learning data that is difficult to classify using the inference result of the classifier. In order to improve the classification accuracy, it is effective to learn more data that is difficult to classify. In this embodiment, the classification accuracy is improved by using data whose inference result is divided between the first candidate and the second candidate as labeled data for learning.

[0029] The present invention has been described above based on the embodiments. The embodiments are merely examples, and it will be understood by those skilled in the art that various modifications are possible in the combination of each component and each treatment process, and that such modifications are also within the scope of the present invention. [Explanation of symbols]

[0030] 10 input unit, 20 learning unit, 30 trained model memory unit, 40 inference unit, 50 evaluation unit, 60 output unit, 100 machine learning device.

Claims

1. an inference unit that infers labels for unlabeled data using the trained model; an evaluation unit that evaluates a probability indicating the likelihood of each candidate of the inference result, and when the inference result is divided between the first candidate and the second candidate, generates labeled data by assigning a label of the first candidate to unlabeled data; A learning unit that re-learns the trained model using labeled data; The machine learning device is further characterized in that the evaluation unit adds as a condition for determining that the inference result is divided between the first candidate and the second candidate that the probability of the third candidate and below in the inference result is smaller than a predetermined upper limit value.

2. The machine learning device according to claim 1, characterized in that the evaluation unit determines that the inference result is divided between the first candidate and the second candidate when the sum of the probability of the first candidate and the probability of the second candidate is greater than a predetermined lower limit value and the difference or ratio between the probability of the first candidate and the probability of the second candidate is smaller than a predetermined upper limit value.

3. The machine learning device according to claim 1, characterized in that the evaluation unit determines that the inference result is divided between the first candidate and the second candidate when the probability of the first candidate is in the range from a predetermined lower limit value to a predetermined upper limit value and the sum of the probability of the first candidate and the probability of the second candidate is greater than a predetermined lower limit value.

4. an inference step of inferring labels of unlabeled data using the trained model; an evaluation step of evaluating a probability indicating the likelihood of each candidate of the inference result, and when the inference result is divided between the first candidate and the second candidate, generating labeled data by labeling the unlabeled data with the label of the first candidate; A learning step of re-learning the trained model using labeled data; The machine learning method is characterized in that the evaluation step further includes adding as a condition for determining that the inference result is divided between the first and second candidates that the probability of the third candidate and below in the inference result is smaller than a predetermined upper limit value.

5. an inference step of inferring labels of unlabeled data using the trained model; an evaluation step of evaluating a probability indicating the likelihood of each candidate of the inference result, and when the inference result is divided between the first candidate and the second candidate, generating labeled data by labeling the unlabeled data with the label of the first candidate; A learning step of re-learning the trained model using labeled data; The machine learning program is characterized in that the evaluation step further adds, as a condition for determining that the inference result is divided between the first and second candidates, that the probability of the third candidate and below in the inference result is smaller than a predetermined upper limit value.

Images