METHOD OF OUTPUTTING PREDICTION RESULT USING NEURAL NETWORK, METHOD OF GENERATING NEURAL NETWORK, AND APPARATUS THEREFOR

• US 20200134427A1
• Filed: 10/22/2019
• Published: 04/30/2020
• Est. Priority Date: 10/30/2018
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: OH, Do-kwan, Jang, Cheol-hun, Ji, Dae-hyun
• Granted Patent: US 11,681,921 B2
• Time in Patent Office: Days
• CPC Class Codes: G06F 18/2155 characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
• CPC Class Codes: G06F 18/24143 Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN]
• CPC Class Codes: G06N 20/20 Ensemble learning
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G06V 10/454 Integrating the filters into a hierarchical structure, e.g. convolutional neural networks [CNN]
• CPC Class Codes: G06V 10/764 using classification, e.g. of video objects
• CPC Class Codes: G06V 10/7788 the supervisor being a human, e.g. interactive learning with a human teacher
• CPC Class Codes: G06V 10/82 using neural networks
• CPC Class Codes: G06V 20/56 exterior to a vehicle by using sensors mounted on the vehicle

Abstract

A method of generating a second neural network model according to an example embodiment includes: inputting unlabeled input data to a first neural network model; obtaining prediction results corresponding to the unlabeled input data based on the first neural network model; and generating a second neural network model based on the prediction results of the first neural network model and a degree of distribution of the prediction results.

Claims

1. A method of generating a neural network model, the method comprising:

inputting unlabeled input data to a first neural network model;

obtaining prediction results corresponding to the unlabeled input data based on the first neural network model; and

generating a second neural network model based on the prediction results and a degree of distribution of the prediction results.

2. The method of claim 1, wherein the first neural network model comprises a plurality of nodes, and

the obtaining comprises:

randomly dropping out at least a portion of the plurality of nodes of the first neural network model to repeatedly obtain the prediction results corresponding to the unlabeled input data.

3. The method of claim 2, wherein the obtaining comprises:

obtaining an average of the repeatedly obtained prediction results; and

determining a pseudo label of the unlabeled input data based on the average of the prediction results.

4. The method of claim 1, wherein the generating the second neural network model comprises:

training the second neural network model based on a value obtained by multiplying the prediction results of the first neural network model by distribution information of the prediction results.

5. The method of claim 1, wherein the training comprises:

training the second neural network model by applying a weight to the prediction results, the weight being proportional to distribution information of the prediction results.

6. The method of claim 1, wherein the obtaining comprises, in response to the unlabeled input data being an image:

predicting a class of an object included in the unlabeled input data; and

predicting a bounding box for detecting the object included in the unlabeled input data.

7. The method of claim 1, wherein the obtaining comprises, in response to the unlabeled input data being voice:

recognizing the voice included in the unlabeled input data.

8. The method of claim 1, wherein the generating comprises generating the second neural network model that corresponds to a result of on-device learning based on the first neural network model or a result of domain adaptation based on the first neural network model.

11. A non-transitory computer readable storage medium storing computer program, which, when executed by at least one processor, causes the at least one processor to execute the method of claim 1.

9. A method of outputting prediction results using a neural network, the method comprising:

receiving object data;

outputting second prediction results corresponding to the object data using a second neural network model, the second neural network model being generated based on a first neural network model, the second prediction results comprising;

a first section indicating that the object data corresponds to a pseudo label corresponding to first prediction results of the first neural network model, a second section in which it is unknown that the object data corresponds to the pseudo label corresponding to the first prediction results, and a third section indicating that the object data does not correspond to the pseudo label corresponding to the first prediction results, wherein the first section, the second section, and the third section are determined based on a degree of distribution of the second prediction results,

wherein the outputting the second prediction results comprising outputting the second prediction results corresponding to the first section, the second section, and the third section.

10. The method of claim 9, wherein the second neural network model is trained by a weight according to a degree of distribution of the first prediction results obtained by the first neural network model.

12. An apparatus for generating a neural network model, the apparatus comprising:

a communication interface configured to receive unlabeled input data; and

at least one processor configured to input the unlabeled input data to a first neural network model, obtain prediction results corresponding to the unlabeled input data based on the first neural network model, and generate a second neural network model based on the prediction results and a degree of distribution of the prediction results.

13. The apparatus of claim 12, wherein the first neural network model comprises a plurality of nodes, and

the at least one processor is configured to randomly drop out at least a portion of the plurality of nodes of the first neural network model to repeatedly obtain the prediction results corresponding to the unlabeled input data.

14. The apparatus of claim 13, wherein the at least one processor is configured to obtain an average of the repeatedly obtained prediction results and to determine a pseudo label of the unlabeled input data based on the average of the prediction results.

15. The apparatus of claim 12, wherein the at least one processor is configured to train the second neural network model based on a value obtained by multiplying the prediction results of the first neural network model by distribution information of the prediction results.

16. The apparatus of claim 15, wherein the at least one processor is configured to train the second neural network model by applying a weight to the prediction results, the weight being proportional to the distribution information of the prediction results.

17. The apparatus of claim 12, wherein the at least one processor, in response to the unlabeled input data being an image, is configured to:

predict a class of an object included in the unlabeled input data; and

predict a bounding box for detecting the object included in the unlabeled input data, or predict the class of the object included in the unlabeled input data and the bounding box for detecting the object included in the unlabeled input data.

18. The apparatus of claim 12, wherein the at least one processor, in response to the unlabeled input data being voice, is configured to recognize the voice included in the unlabeled input data.

19-20. -20. (canceled)