Machine leaning 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 re-learns 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 does not require labeling of data, so data preparation is not required. It has the advantage that it takes little time and a large amount of data can be used for learning, but However, it has the disadvantage that the learning efficiency is lower than that of the learning with a small number of labeled data. Semi-supervised learning that can easily and efficiently learn using a large amount of unlabeled data Practice is proposed.

[0003] Patent Document 1 includes a feature extractor but does not include a generator, and performs recognition processing of input images. The semi-supervised learning of image recognizers, which are information converters, is performed by reducing the possibility of competition between learning. An information processing device that reduces the number of [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] Classifier-based bootstrap methods in semi-supervised learning use a trained classifier to Perform inference on unlabeled data and add the results with high confidence to the labeled data. In this method, the inference data with high confidence is based on the linear regression data used during learning. There is a high probability that the data is similar to the data with the belt. It is known that learning more data is effective, and the number of labeled data and the classification Even if similar inference data with high confidence is used for learning, classification accuracy does not improve significantly.

[0006] The present invention has been made in view of the above circumstances, and its object is to improve classification accuracy. The objective of the present invention is to provide a machine learning technology that can [Means for solving the problem]

[0007] In order to solve the above problems, a machine learning device according to an embodiment of the present invention uses a trained model. The inference part infers the labels of unlabeled data, and the inference part estimates the likelihood of each candidate of the inference result. If the inference result is dichotomous between the first and second candidates, the unlabeled data is An evaluation unit generates labeled data by assigning the first candidate label to the data, and and a learning unit that re-learns the trained model using the trained data.

[0008] Another aspect of the present invention is a machine learning method. The method includes: We show the inference step of inferring the labels of unlabeled data and the likelihood of each candidate inference result. If the inference result is split between the first and second candidates, the unlabeled data is An evaluation step of generating labeled data by labeling the first candidate data; and a learning step of re-learning the trained model using new data.

[0009] Any combination of the above components and the expression of the present invention may be used as a method, an apparatus, a system, a recording medium, Conversions between the body, computer program, etc. are also valid aspects of the present invention. be. Effect of the Invention

[0010] According to the present invention, it is possible to provide a machine learning technique that can improve classification accuracy. Cut. [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] FIG. 1 is a configuration diagram of a machine learning device 100 according to an embodiment. The machine learning device 100 , an input unit 10, a learning unit 20, a trained model storage unit 30, an inference unit 40, an evaluation unit 50, and and an output section 60.

[0013] The input unit 10 supplies unlabeled data to the inference unit 40, and the label provided by the evaluation unit 50 The data with a bell is supplied to the learning unit 20. Here, an example of image recognition will be described. The data input to the system is an image of the object to be recognized. For example, an image of a person, a motor vehicle, These include images of Iku, images of cars, images of bicycles, etc.

[0014] The trained model storage unit 30 stores a trained model using a pre-training data set with labels. The object recognition model with excellent recognition accuracy is stored.

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

[0016] The evaluation unit 50 evaluates the likelihood of each candidate of the inference result of the inference unit 40 on the unlabeled data. The evaluation unit 50 evaluates the probability that the inference result is divided between the first candidate and the second candidate. The label of the first candidate is assigned to the data that is difficult to classify, to generate labeled data, and the data is input to the input unit 10. In the case of image recognition, the probability of each candidate recognition result of the input image is evaluated, and the recognition result is Extract input images that are difficult to classify, and assign the first candidate label to the image to label it. For example, if the input image recognition result is a bicycle as the first candidate and a bicycle as the second candidate, If the classification is between two motorcycles, the input image is labeled as the first candidate bicycle. and generate a labeled image.

[0017] The inference result is split into two parts: the inference result is split between the first and second candidates, and the third and subsequent candidates The lower probability is very low.

[0018] The condition for the inference result to be split into two is, more specifically, the sum of the probability of the first candidate and the probability of the second candidate. The total 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 The probability of the third or lower candidate being the first candidate is smaller than the upper limit B. Furthermore, the probability of the third or lower candidate being the second candidate is smaller than the upper limit C. Here, A, B, and C are generally different values. However, B and C may have the same value. For example, the probability that the input image is a bicycle is 57%, The probability that it is a motorcycle is 38%, the probability that it is a car is 5%, and the probability that it is a person is 0%. This is the result of inference.

[0019] Alternatively, the condition for the inference result to be divided into two is when the probability of the first candidate falls between a predetermined lower limit D and a predetermined The range is an upper limit value E, and the sum of the probability of the first candidate and the probability of the second candidate is less than a predetermined lower limit value F. Furthermore, the condition that the probability of the third candidate and below is smaller than a predetermined upper limit G is satisfied. Here, D, E, F, and G are generally different values, but E and F are As an example, 50%<probability of the first candidate<70% and the first candidate You can also set the condition that the probability of the first candidate + the probability of the second candidate > 80%. Alternatively, you can set the condition that the probability of the first candidate < 55%. The probability range of the first candidate is 35% < 65% and 35% < probability of the second candidate < 45%. A range of probabilities for each of the second candidates may be determined as a condition.

[0020] If the inference result of the unlabeled data satisfies the condition for dichotomy, the evaluation unit 50 classifies the data as The label of the first candidate is assigned to the data to generate labeled data, which is then supplied to the input unit 10. If the inference result of the unlabeled data does not meet the dichotomous condition, the data has a classification accuracy of Since this is not effective training data for improvement, it is not supplied to the input unit 10.

[0021] The learning unit 20 uses the labeled data supplied from the input unit 10 as data used for pre-learning. The trained object recognition model stored in the trained model storage unit 30 is added to the trained model set. Retraining is performed, the weights of the object recognition model are updated, and a new object recognition model is generated.

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

[0023] The inference unit 40 processes the input image of the recognition target provided from the input unit 10 based on the trained object recognition model. The images are then input into the system to recognize objects in the images (S10).

[0024] The evaluation unit 50 evaluates the inference result by the inference unit 40 and determines whether the inference result is a first candidate or a second candidate. If the result is Y in S20, the input image is labeled with the first candidate label. If the inference result is not dichotomous (S20), the data is supplied as labeled data (S30). N), no labeled data is generated, and the process returns to step S10 to input the next image.

[0025] Step 4: The procedure from step S10 to step S30 is repeated. When the input images of the group are accumulated (Y in S40), the learning unit 20 uses the data used for the pre-learning Generate a new dataset by adding a given number of labeled images to the set, and then The trained object recognition model is retrained using the dataset, and the weights of the trained object recognition model are The learning unit 20 generates a new object recognition model and updates the learned model record (S50). The result is stored in the memory unit 30 (S60).

[0026] In the above explanation, transfer learning was performed after a certain number of input images with the same label had been accumulated. Learning may be performed each time a labeled image is generated.

[0027] The various processes of the machine learning device 100 described above are performed using hardware such as a CPU and a memory. It can be realized as a device using ROM (Read Only Memory) firmware stored in hard disk drives, flash memory, etc., and computer software, etc. The firmware program, software The program may be provided by recording it on a computer-readable recording medium, or by wired communication. Alternatively, the data can be transmitted to and from the server via a wireless network, or via terrestrial or satellite broadcasting. It is also possible to transmit and receive it as digital broadcasting data broadcasting.

[0028] As described above, according to the machine learning device 100 of the present embodiment, the inference result of the classifier is By using the learning algorithm to learn additional data that is difficult to classify, the accuracy of classification can be improved. To improve classification accuracy, it is effective to learn more data that is difficult to classify. In this embodiment, data for which the inference result is divided between the first candidate and the second candidate is labeled. By using these as valid data for learning, classification accuracy is improved.

[0029] The present invention has been described above based on the embodiments. The embodiments are merely examples, and each of the structures The fact that various modifications are possible in the combination of components and each treatment process, and that such modifications It will be understood by those skilled in the art 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. The probability of each candidate inference result when inferring the label of unlabeled data using a trained model If the inference result is split between the first and second candidates, the label is an evaluation unit that generates labeled data by assigning a first candidate label to the unlabeled data; Including, The evaluation unit further determines whether the probability of the third or lower candidate in the inference result is smaller than a predetermined upper limit value. and as a condition for determining that the inference result is divided between the first and second candidates. A learning device characterized by:

2. The probability of each candidate inference result when inferring the label of unlabeled data using a trained model If the inference result is split between the first and second candidates, the label is an evaluation step of generating labeled data by labeling the first candidate label to the unlabeled data; Including, The evaluation step further comprises determining whether the probability of the third or subsequent candidate in the inference result is smaller than a predetermined upper limit value. Add this as a condition for determining whether the inference result is split between the first and second candidates. A learning method comprising:

3. The probability of each candidate inference result when inferring the label of unlabeled data using a trained model If the inference result is split between the first and second candidates, the label is an evaluation step of generating labeled data by labeling the first candidate label to the unlabeled data; Run the following on your computer: The evaluation step further comprises determining whether the probability of the third or subsequent candidate in the inference result is smaller than a predetermined upper limit value. Add this as a condition for determining whether the inference result is split between the first and second candidates. A learning program characterized by:

Images