Apparatus and method for training deep neural network

1. A method which is performed by a computing device comprising one or more processors and a memory storing one or more programs to be executed by the one or more processors, the method comprising:

training a deep neural network model using a first data set including a plurality of labeled data and a second data set including a plurality of unlabeled data;

assigning a ground-truth label value to some of the plurality of unlabeled data;

updating the first data set and the second data set such that the data to which the ground-truth label value is assigned is included in the first data set; and

further training the deep neural network model using the updated first data set and the updated second data set,

wherein the deep neural network model comprises a feature extractor configured to generate a feature vector for input data of the deep neural network model and a plurality of classifiers configured to each generate an estimated label value for the input data based on the feature vector, and

wherein the training comprises training the deep neural network by performing a first global update to update training parameters of the feature extractor and the plurality of classifiers using the first data set and a first local update to update training parameters of each of the plurality of classifiers using the second data set.

2. The method of claim 1 , wherein the training comprises repeating the first global update and the first local update alternately.

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

performing the first global update such that a ground-truth label value assigned to each data included in the first data set matches the estimated label value of each of the plurality of classifiers for each data included in the first data set; and

performing the first local update such that each of the plurality of classifiers generates a different estimated label value for each data included in the second data set.

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

performing the first global update using a first loss function based on the ground-truth label value assigned to each data included in the first data set and the estimated label value of each of the plurality of classifiers for each data included in the first data set; and

performing the first local update using a second loss function based on the estimated label value of each of the plurality of classifiers for each data included in the second data set and an ensemble-estimated label value based on the estimated label value of each of the plurality of classifiers for each data included in the second data set.

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

determining some of the data included in the first data set to be target data to be labeled based on an estimated label value of each of the plurality of classifiers for each data included in the first data set;

receiving the ground-truth label value to be assigned to each target data to be labeled from a user; and

assigning the received ground-truth label value to each target data to be labeled.

6. The method of claim 5 , wherein the determining comprises:

calculating a score for each data included in the first data set such that the more the estimated label values of the plurality of classifiers differ, the higher the score of each data included in the first data set is; and

determining, as the target data to be labeled, a predetermined number of data or a predetermined proportion of data, among the data included in the first data set, based on descending order of the scores calculated.

7. The method of claim 1 , wherein the further training comprises further training the deep neural network model by performing second global update to update training parameters of the deep neural network model using the updated first data set and second local update to update training parameters of the feature extractor using the updated second data set.

8. The method of claim 7 , wherein the further training comprises repeating the second global update and the second local update alternately.

9. The method of claim 7 , wherein the further training comprises:

performing the second global update such that a ground-truth label value assigned to each data included in the updated first data set matches an estimated label value of each of the plurality of classifiers for each data included in the updated first data set; and

performing the second local update such that each of the plurality of classifiers generates the same estimated label value for each data included in the updated first data set.

10. The method of claim 7 , wherein the further training comprises:

performing the second global update using a first loss function based on a ground-truth label value assigned to each data included in the updated first data set and an estimated label value of each of the plurality of classifiers for each data included in the updated first data set and

performing the second local update using a third loss function based on an estimated label value of each of the plurality of classifiers for each data included in the updated second data set and an ensemble-estimated label value based on the estimated label value of each of the plurality of classifiers for each data included in the updated second data set.

11. An apparatus comprising:

one or more processors; and

a memory storing one or more programs to be executed by the one or more processors,

wherein the programs include commands to perform operations comprising:

training a deep neural network model using a first data set including a plurality of labeled data and a second data set including a plurality of unlabeled data,

assigning a ground-truth label value to some of the plurality of unlabeled data,

updating the first data set and the second data set such that the data to which the ground-truth label value is assigned is included in the first data set, and

further training the deep neural network model using the updated first data set and the updated second data set,

wherein the deep neural network model comprises a feature extractor configured to generate a feature vector for input data of the deep neural network model and a plurality of classifiers configured to each generate an estimated label value for the input data based on the feature vector, and

wherein the training comprises training the deep neural network by performing a first global update to update training parameters of the feature extractor and the plurality of classifiers using the first data set and a first local update to update training parameters of each of the plurality of classifiers using the second data set.

12. The apparatus of claim 11 , wherein the training comprises repeating the first global update and the first local update alternately.

13. The apparatus of claim 11 , wherein the training comprises:

performing the first global update such that a ground-truth label value assigned to each data included in the first data set matches the estimated label value of each of the plurality of classifiers for each data included in the first data set; and

performing the first local update such that each of the plurality of classifiers generates a different estimated label value for each data included in the second data set.

14. The apparatus of claim 11 , wherein the training comprises:

performing the first global update using a first loss function based on the ground-truth label value assigned to each data included in the first data set and the estimated label value of each of the plurality of classifiers for each data included in the first data set; and

performing the first local update using a second loss function based on the estimated label value of each of the plurality of classifiers for each data included in the second data set and an ensemble-estimated label value based on the estimated label value of each of the plurality of classifiers for each data included in the second data set.

15. The apparatus of claim 11 , wherein the assigning comprises:

determining some of the data included in the first data set to be target data to be labeled based on an estimated label value of each of the plurality of classifiers for each data included in the first data set;

receiving the ground-truth label value to be assigned to each target data to be labeled from a user; and

assigning the received ground-truth label value to each target data to be labeled.

16. The apparatus of claim 15 , wherein the determining comprises:

calculating a score for each data included in the first data set such that the more the estimated label values of the plurality of classifiers differ, the higher the score of each data included in the first data set is; and

determining, as the target data to be labeled, a predetermined number of data or a predetermined proportion of data, among the data included in the first data set, based on descending order of the scores calculated.

17. The apparatus of claim 11 , wherein the further training comprises further training the deep neural network model by performing second global update to update training parameters of the deep neural network model using the updated first data set and second local update to update training parameters of the feature extractor using the updated second data set.

18. The apparatus of claim 17 , wherein the further training comprises repeating the second global update and the second local update alternately.

19. The apparatus of claim 17 , wherein the further training comprises:

performing the second global update such that a ground-truth label value assigned to each data included in the updated first data set matches an estimated label value of each of the plurality of classifiers for each data included in the updated first data set; and

performing the second local update such that each of the plurality of classifiers generates the same estimated label value for each data included in the updated first data set.

20. The apparatus of claim 17 , wherein the further training comprises:

performing the second global update using a first loss function based on a ground-truth label value assigned to each data included in the updated first data set and an estimated label value of each of the plurality of classifiers for each data included in the updated first data set and

performing the second local update using a third loss function based on an estimated label value of each of the plurality of classifiers for each data included in the updated second data set and an ensemble-estimated label value based on the estimated label value of each of the plurality of classifiers for each data included in the updated second data set.