Method and device for recognition and method and device for constructing recognition model

• US 10,475,442 B2
• Filed: 10/24/2016
• Issued: 11/12/2019
• Est. Priority Date: 11/25/2015
• Status: Active Grant

• Current Assignee: Samsung Electronics Co. Ltd.
• Original Assignee: Samsung Electronics Co. Ltd.
• Inventors: Lee, Ho Shik
• Primary Examiner(s): Colucci, Michael C
• Application Number:
  US15/332,166
• Publication Number:
  US 20170148430A1
• Time in Patent Office: 1,114 Days
• Field of Search: 704 9, 704232, 704211, 704234, 704235, 704239, 706 12, 706 16
• CPC Class Codes: G10L 15/01 Assessment or evaluation of speech recognition systems
• CPC Class Codes: G10L 15/063 Training
• CPC Class Codes: G10L 15/16 using artificial neural networks
• CPC Class Codes: G10L 2015/0635 updating or merging of old and new templates; Mean values; Weighting

Abstract

A method and a device for recognition, and a method and a device for constructing a recognition model are disclosed. A device for constructing a recognition model includes a training data inputter configured to receive additional training data, a model learner configured to train a first recognition model constructed based on basic training data to learn the additional training data, and a model constructor configured to construct a final recognition model by integrating the first recognition model with a second recognition model generated by the training of the first recognition model.

Claims

1. A device for constructing a recognition model, the device comprising:

one or more hardware processors configured to;

obtain a first recognition model trained based on basic training data;

obtain new additional training data separate from the basic training data;

generate a second recognition model by further training the first recognition model to learn the additional training data; and

construct a final recognition model by integrating the first recognition model with the second recognition model,

wherein, in the integrating of the first recognition model with the second recognition model, a select weight of the second recognition model is set to emphasize learning of the additional training data over maintaining of the first recognition model, or is set to emphasize the maintaining of the first recognition model over the learning of the additional training data.

2. The device of claim 1, wherein the basic training data and the additional training data comprise at least one of voice data, sound data, image data, text data, and bio-signal data.

3. The device of claim 1, wherein the first recognition model is a neural network-based recognition model.

4. The device of claim 3, wherein, for the generating of the second recognition model, the one or more hardware processors are configured to use either an error back-propagation algorithm or a gradient descent algorithm in training the first recognition model with the additional training data.

5. The device of claim 1, further comprising a memory configured to store one or more first recognition models, including the first recognition model, respectively constructed based on one or more sets of basic training data in advance,

wherein, for the generating of the second recognition model, the one or more hardware processors are configured to acquire the first recognition model related to the obtained additional training data from the memory.

6. The device of claim 1, wherein, for the constructing of the final recognition model, the one or more hardware processors are configured to obtain a linear sum of the first recognition model and the second recognition model.

7. The device of claim 6, wherein, for the integrating of the first recognition model with the second recognition model, the one or more hardware processors are configured to select one or more weights, including the select weight, from a preset weight candidate group, generate one or more recognition models for each of the selected one or more weights by assigning the selected one or more weights and performing linear summations, and determine at least one of the generated one or more recognition models as a final recognition model.

8. The device of claim 7, wherein the selected one or more weights are preset based on at least one of a comparison between quantity of the basic training data and quantity of the additional training data, a purpose of additional learning, and importance of additional learning.

9. The device of claim 1, wherein the one or more hardware processors are further configured to evaluate a constructed final recognition model using evaluation data in response to the final recognition model being constructed.

10. The device of claim 9, wherein, in response to an evaluation that the final recognition model does not satisfy a predetermined standard, the one or more hardware processors are configured to train the final recognition model to re-learn the additional training data.

11. The device of claim 1, wherein the weight is selected based on a difference between a quantity of the basic training data and a quantity of the additional training data.

12. A method of constructing a recognition model, the method comprising:

obtaining a first recognition model trained based on basic training data;

obtaining new additional training data separate from the basic training data;

generating a second recognition model by further training, using one or more hardware processors, the first recognition model to learn the additional training data; and

constructing, using the one or more hardware processors, a final recognition model by integrating the first recognition model with the second recognition model,

wherein, in the integrating of the first recognition model with the second recognition model a select weight of the second recognition model is set to emphasize learning of the additional training data over maintaining of the first recognition model, or is set to emphasize the maintaining of the first recognition model over the learning of the additional training data.

13. The method of claim 12, wherein the first recognition model is a neural network-based recognition model.

14. The method of claim 12, wherein the training of the first recognition model comprises acquiring a first recognition model related to the obtained additional training data from a model storage in response to obtaining the additional training data.

15. The method of claim 12, wherein the constructing of the final recognition model comprises obtaining a linear sum of the first recognition model and the second recognition model.

16. The method of claim 15, wherein the constructing of the final recognition model comprises selecting one or more weights from a preset weight candidate group, generating one or more recognition models for each weight by assigning the selected weights and performing linear summations, and determining at least one of the generated one or more recognition models as a final recognition model.

17. The method of claim 12, further comprising evaluating a constructed final recognition model using evaluation data in response to the final recognition model being constructed.

18. The method of claim 17, further comprising, in response to an evaluation that the final recognition model does not satisfy a predetermined standard, training the final recognition model to re-learn the additional training data.

19. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, cause the processor to perform the method of claim 12.

20. A recognition device comprising:

one or more hardware processors configured to;

receive data to be recognized from a user;

determine, based on the received data, whether to update a recognition model constructed in advance based on basic training data;

in response to a determination that an update is required, train the recognition model to learn new additional training data separate from the basic training data to update the recognition model, the updating of the recognition model being performed to emphasize learning of the received data over maintaining of the recognition model, or performed to emphasize the maintaining of the recognition model over the learning of the received data; and

recognize the received data using the recognition model constructed in advance or the updated recognition model.

21. The device of claim 20, wherein the one or more hardware processors comprise:

a recognition data inputter configured to obtain the data;

a model update determiner configured to determine whether to update the recognition model constructed in advance;

a model updater configured to train and update the recognition model; and

a data recognizer configured to recognize the received data.

22. The recognition device of claim 20, wherein the model update determiner is configured to determine whether to update the recognition model constructed in advance based on either relevance between basic training data used to construct the recognition model constructed in advance and the received data or accuracy of a recognition result obtained using the recognition model constructed in advance.

23. The recognition device of claim 20, further comprising an additional data inputter configured to receive additional training data related to the received data in response to a determination that an update is required.

24. The recognition device of claim 20, wherein the model updater is configured to perform a weighted linear summation of the recognition model constructed in advance and a recognition model generated as a result of learning of the recognition model constructed in advance and update the recognition model constructed in advance based on the result of the weighted linear summation.

25. The recognition device of claim 20, wherein the one or more hardware processors are configured to process a recognition result based on a determined purpose of the data recognition and provide the processed recognition result to the user.

26. The device of claim 20, wherein a weight of the updated recognition model is selected based on a difference between a quantity of basic training data in the recognition model constructed in advance and a quantity of the additional training data.

27. A recognition model construction device, comprising:

one or more hardware processors configured to;

obtain a first recognition model trained based on basic training data;

obtain a second recognition model by training the first recognition model to learn new additional training data separate from the basic training data; and

construct a third recognition model by integrating the first recognition model with the second recognition model based on a select weight, the select weight being set to emphasize learning of the additional training data over maintaining of the first recognition model, or is set to emphasize the maintaining of the first recognition model over the learning of the additional training data.

28. The device of claim 27, wherein the first recognition model is a model constructed based on basic training data, and the additional training data comprise data not included in the basic training data.

29. The device of claim 27, wherein the one or more hardware processors is configured to construct the third recognition model by using the weight to obtain a linear sum of the first recognition model and the second recognition model.

30. The device of claim 27, wherein the one or more hardware processors is further configured to select a model among recognition models stored in a memory as the first recognition model based on a relevance of the selected model to the additional training data.