ARTIFICIAL INTELLIGENCE SERVER

[0000]
An artificial intelligence server is disclosed. The artificial intelligence server includes an input unit to which input data is inputted, and a processor, when a first output value outputted by an artificial intelligence model with respect to first input data is correct and a second output value outputted by the artificial intelligence model with respect to second input data is incorrect, configured to use the first input data and the second input data to obtain a first domain causing an incorrect answer, and train the artificial intelligence model to be domain-adapted for the first domain.

[00000]

1 . Artificial intelligence server comprising:

an input interface to which input data is inputted; and

a processor, when a first output value outputted by an artificial intelligence model with respect to first input data is correct and a second output value outputted by the artificial intelligence model with respect to second input data is incorrect, configured to use the first input data and the second input data to obtain a first domain causing an incorrect answer, and train the artificial intelligence model to be domain-adapted for the first domain.

2 . The artificial intelligence server of
claim 1
, wherein when a third output value outputted by the trained artificial intelligence model with respect to third input data is correct and a fourth output value outputted by the trained artificial intelligence model with respect to fourth input data is incorrect, the processor obtains a second domain causing an incorrect answer using the third input data and the fourth input data; and re-trains the trained artificial intelligence model to be domain-adapted for the second domain,

wherein the second domain is different from the first domain.

3 . The artificial intelligence server of
claim 2
, wherein when the artificial intelligence model outputs an output value using features corresponding to a plurality of domains, the processor obtains the first domain causing the most incorrect answer among the plurality of domains.

4 . The artificial intelligence server of
claim 3
, wherein when the artificial intelligence model outputs an output value using features corresponding to the plurality of domains, the processor obtains the first domain causing the most incorrect answer among the plurality of domains by using a distribution of the first input data and a distribution of the second input data for each of the plurality of domains.

5 . The artificial intelligence server of
claim 3
, wherein when the trained artificial intelligence model outputs an output value using features corresponding to the plurality of domains, the processor obtains the second domain causing the most incorrect answer among the plurality of domains.

6 . The artificial intelligence server of
claim 5
, wherein when the trained artificial intelligence model outputs an output value using features corresponding to the plurality of domains, the processor obtains the second domain causing the most incorrect answer among the plurality of domains by using a distribution of the first input data and a distribution of the second input data for each of the plurality of domains.

7 . The artificial intelligence server of
claim 2
, wherein the first domain comprises a 1 - 1 domain and a 1 - 2 domain,

wherein the processor trains the artificial intelligence model to allow a feature extracted by the artificial intelligence model with respect to input data corresponding to the 1 - 1 domain and a feature extracted by the artificial intelligence model with respect to input data corresponding to the 1 - 2 domain to be mapped to the same area.

8 . The artificial intelligence server of
claim 7
, wherein the second domain comprises a 2 - 1 domain and a 2 - 2 domain,

wherein the processor re-trains the trained artificial intelligence model to allow a feature extracted by the trained artificial intelligence model with respect to input data corresponding to the 2 - 1 domain and a feature extracted by the artificial intelligence model with respect to input data corresponding to the 2 - 2 domain to be mapped to the same area.

9 . The artificial intelligence server of
claim 7
, wherein the artificial intelligence model comprises:

a feature extractor configured to extract the feature using input data;

a class classifier configured to classify classes using the extracted features; and

a domain classifier configured to classify domains using the extracted features.

10 . The artificial intelligence server of
claim 9
, wherein the processor trains the artificial intelligence model to allow the class classifier to classify the classes and prevent the domain classifier from classifying the 1 - 1 domain and the 1 - 2 domain.

11 . The artificial intelligence server of
claim 1
, wherein the processor obtains a second domain causing a second largest incorrect answer among a plurality of domains by using the first input data and the second input data, and re-trains the trained artificial intelligence model to be domain-adapted for the second domain.

12 . The artificial intelligence server of
claim 1
, wherein the processor selects an artificial intelligence model with the highest performance among a plurality of artificial intelligence models in which at least one of the number of domain adaptation, a target domain of domain adaptation, or the order of domain adaptation is different.

13 . The artificial intelligence server of
claim 12
, wherein the processor

trains the artificial intelligence model to be domain-adapted for the first domain so as to generate a second artificial intelligence model,

trains the second artificial intelligence model to be domain-adapted for a second domain so as to generate a third artificial intelligence model, and

selects an artificial intelligence model with a higher performance among the second artificial intelligence model and the third artificial intelligence model.

14 . The artificial intelligence server of
claim 12
, wherein the processor

trains the artificial intelligence model to be domain-adapted for the first domain so as to generate a second artificial intelligence model, and trains the second artificial intelligence model to be domain-adapted for the second domain so as to generate a third artificial intelligence model,

trains the artificial intelligence model to be domain-adapted for the second domain so as to generate a fourth artificial intelligence mode, and

selects an artificial intelligence model with higher performance among the third artificial intelligence model and the fourth artificial intelligence model.

15 . The artificial intelligence server of
claim 12
, wherein the processor

trains the artificial intelligence model to be domain-adapted for the first domain so as to generate a second artificial intelligence model,

trains the second artificial intelligence model to be domain-adapted for the second domain so as to generate a third artificial intelligence mode, and

deletes the third artificial intelligence model from memory when a performance of the second artificial intelligence model among the second artificial intelligence model and the third artificial intelligence model is higher.

16 . The artificial intelligence server of
claim 12
, wherein the processor

trains the artificial intelligence model to be domain-adapted for the first domain so as to generate a second artificial intelligence model,

trains the second artificial intelligence model to be domain-adapted for the second domain so as to generate a third artificial intelligence mode, and

does not additionally trains the third artificial intelligence model when a performance of the third artificial intelligence model is increased by less than a predetermined value compared to a performance of the second artificial intelligence model.

17 . The artificial intelligence server of
claim 12
, wherein the processor does not additionally train an artificial intelligence model that is not selected as an artificial intelligence model with the highest performance for more than a predetermined period among the plurality of artificial intelligence models.

18 . A domain adaptation method comprising:

when a first output value outputted by an artificial intelligence model with respect to first input data is correct and a second output value outputted by the artificial intelligence model with respect to second input data is incorrect, using the first input data and the second input data to obtain a first domain causing an incorrect answer; and

training the artificial intelligence model to be domain-adapted for the first domain.

19 . The method of
claim 18
, further comprising:

when a third output value outputted by the trained artificial intelligence model with respect to third input data is correct and a fourth output value outputted by the trained artificial intelligence model with respect to fourth input data is incorrect, obtaining a second domain causing an incorrect answer using the third input data and the fourth input data; and

re-training the trained artificial intelligence model to be domain-adapted for the second domain,

wherein the second domain is different from the first domain

TECHNICAL FIELD
[0001]
The present invention relates to an artificial intelligence server that can improve the performance of an artificial intelligence model by training the artificial intelligence model to be domain adaptation (domain adaptation) to the various domains that caused the incorrect answer.

BACKGROUND ART
[0002]
Artificial intelligence is a field of computer science and information technology that studies a method for computers to do thinking, learning, and self-development that human intelligence can do and means enabling computers to imitate human intelligent behavior.

[0003]
In addition, artificial intelligence does not exist by itself, but is directly or indirectly related to other fields of computer science. Especially in modern days, artificial intelligence elements are introduced in various fields of information technology so that attempts are being actively made to solve problems in the field.

[0004]
Meanwhile, technologies for recognizing and learning the surrounding situation using artificial intelligence, providing information desired by a user in a desired form, or performing a desired operation or function have been actively studied.

[0005]
Then, an electronic device providing such various operations and functions may be referred to as an artificial intelligence device.

[0006]
Meanwhile, the AI model is trained in a lab environment and released as a product.

[0007]
However, since the laboratory environment and the actual use environment of the artificial intelligence model may be different, the performance of the artificial intelligence model may be lower than that of the laboratory environment.

[0008]
For example, the designer of the artificial intelligence model has trained a speech recognition model using speech data collected in a quiet environment (i.e., a low noise environment). However, when a product equipped with a speech recognition model is used in a noisy environment (high noise environment), the performance of the speech recognition model may be lowered because a loud noise data is inputted to the speech recognition model.

[0009]
Therefore, the need of improving the performance by detecting the difference between the environment in which the artificial intelligence model is trained and the actual use environment and training the deep learning model according to this difference has emerged.

DISCLOSURE OF THE INVENTION Technical Problem
[0010]
The present invention relates to an artificial intelligence server that can improve the performance of an artificial intelligence model by training the artificial intelligence model to be domain adaptation (domain adaptation) to the various domains that caused the incorrect answer.

Technical Solution
[0011]
According to an embodiment of the present invention, an artificial intelligence server includes an input unit to which input data is inputted, and a processor, when a first output value outputted by an artificial intelligence model with respect to first input data is correct and a second output value outputted by the artificial intelligence model with respect to second input data is incorrect, configured to use the first input data and the second input data to obtain a first domain causing an incorrect answer, and train the artificial intelligence model to be domain-adapted for the first domain.

Advantageous Effects
[0012]
The present invention has the advantage of constantly improving the performance of the artificial intelligence model by repeatedly performing domain adaptation.

[0013]
In addition, since the present invention determines the domain causing the most incorrect answer and first performs domain adaptation on the domain causing the most incorrect answer, there is an advantage to improve the performance of the artificial intelligence model faster.

[0014]
In addition, according to the present invention, since domain adaptation is repeatedly performed while changing a domain that is to be a target of domain adaptation, various domains are domain-adapted. Therefore, there is an advantage of improving the performance of the artificial intelligence model more quickly.

[0015]
In addition, according to the present invention, each time the domain adaptation is repeatedly performed, the domain adaptation is performed by selecting a domain causing the most incorrect answer. Therefore, there is an advantage of improving the performance of the artificial intelligence model more quickly.

[0016]
According to the present invention, the performance of the AI model can be improved by performing domain adaptation in various combinations and selecting the artificial intelligence model having the highest performance.

[0017]
According to the present invention, some artificial intelligence models of the plurality of artificial intelligence models are not additionally trained or some artificial intelligence models are deleted from the memory, thereby reducing the amount of computation and storage space.

BRIEF DESCRIPTION OF THE DRAWINGS
[0018]
FIG. 1 illustrates an AI apparatus 100 according to an embodiment of the present invention.

[0019]
FIG. 2 illustrates an AI server 200 according to an embodiment of the present invention.

[0020]
FIG. 3 illustrates an AI system 1 according to an embodiment of the present invention.

[0021]
FIG. 4 is a view illustrating an operation method of an AI server according to an embodiment of the present invention.

[0022]
FIGS. 5 and 7 are views for describing a method of acquiring a domain causing an incorrect answer according to an embodiment of the present invention.

[0023]
FIG. 8 is a view illustrating a domain adaptation method.

[0024]
FIG. 9 is a view for describing domain adaptation using Domain Adversarial Training of Neural Networks (DANN) according to an embodiment of the present invention.

[0025]
FIG. 10 is a view for describing a method of selecting an artificial intelligence model having optimal performance while repeatedly performing domain adaptation and then, managing a history.

[0026]
FIG. 11 is a view for describing a method of extracting an important word from a spoken text and acquiring a domain causing an incorrect answer using a feature extracted from the important word according to an embodiment of the present invention.

[0027]
FIG. 12 is a view for describing a method of acquiring a low confidence word and distinguishing the low confidence word using the importance of the low confidence word according to an embodiment of the present invention.

MODE FOR CARRYING OUT THE INVENTION
[0028]
Hereinafter, embodiments of the present disclosure are described in more detail with reference to accompanying drawings and regardless of the drawings symbols, same or similar components are assigned with the same reference numerals and thus overlapping descriptions for those are omitted. The suffixes “module” and “unit” for components used in the description below are assigned or mixed in consideration of easiness in writing the specific…