Artificial intelligence apparatus for generating training data, artificial intelligence server, and method for the same

1. An artificial intelligence apparatus for generating training data, comprising:

a memory configured to store a target artificial intelligence model; and

a processor configured to:

receive sensor data;

determine whether the received sensor data is relevant or irrelevant to a learning of the target artificial intelligence model, the received sensor data being relevant based on the received sensor data including an object of interest of the target artificial intelligence model and the received sensor data being irrelevant based on the received sensor data not including the object of interest;

generate a first result from the received sensor data by using the target artificial intelligence model;

calculate an aleatoric uncertainty regarding noise embedded in the received sensor data by determining Gaussian noise observed in the generated first result as the aleatoric uncertainty;

generate a plurality of second results from the received data by randomly activating or deactivating one or more hidden nodes included in a hidden layer of the target artificial intelligence model;

calculate an epistemic uncertainty by determining a variance of the generated second results as the epistemic uncertainty using the target artificial intelligence model with combinations of activation or deactivation of the one or more hidden nodes;

determine whether the calculated aleatoric uncertainty is less than a first reference value and the calculated epistemic uncertainty is greater than or equal to a second reference value;

extract a label from the received sensor data by using a label extractor based on a determination that aleatoric uncertainty is less than the first reference value and the epistemic uncertainty is greater than or equal to the second reference value;

determine a confidence level of the extracted label; and

generate training data including the received sensor data and the extracted label based on the determined confidence level exceeding a third reference value.

2. The artificial intelligence apparatus according to claim 1 ,

wherein the object of interest is an object that is identifiable or recognizable by the target artificial intelligence model.

3. The artificial intelligence apparatus according to claim 2 , wherein the processor is configured to determine whether the object of interest is included in the received sensor data or whether an unknown object is included in the received sensor data, by using an open set recognition model corresponding to the target artificial intelligence model.

4. The artificial intelligence apparatus according to claim 3 , wherein the processor is configured to determine that only a portion of the received sensor data, in which the unknown object is included, is irrelevant to the learning.

5. The artificial intelligence apparatus according to claim 1 , wherein the processor is configured to:

determine that the received sensor data is useful for the learning if the calculated aleatoric uncertainty for the received sensor data is less than the first reference value and the calculated epistemic uncertainty for the received sensor data is greater than or equal to the second reference value.

6. The artificial intelligence apparatus according to claim 5 , wherein the processor is configured to:

determine that the received sensor data is not useful for the learning if the calculated aleatoric uncertainty for the received sensor data is greater than or equal to the first reference value, and the calculated epistemic uncertainty for the received sensor data is less than the second reference value.

7. The artificial intelligence apparatus according to claim 1 , wherein the processor is configured to:

calculate a distance between a position where sensor data which has the same label as the extracted label to which a label is set by a user is mapped to a feature space and a position where the received sensor data is mapped to the feature space; and

determine a confidence level of the extracted label to be higher as the distance is smaller.

8. The artificial intelligence apparatus according to claim 1 , wherein the processor is configured to store, in the memory, a generative model for reproducing the generated training data, instead of the generated training data.

9. The artificial intelligence apparatus according to claim 8 , wherein the processor is configured to learn the generative model by using the generated training data, and

wherein the generative model includes one of a generative adversarial network (GAN), a conditional GAN (cGAN), an auto-encoder, or a variational auto-encoder (VAE).

10. The artificial intelligence apparatus according to claim 8 , wherein the processor is configured to:

reproduce previous training data corresponding to a previous learning task of the generative model by using the generative model; and

learn the generative model by using the reproduced previous training data and the generated training data.

11. The artificial intelligence apparatus according to claim 1 , further comprising a sensor unit including at least one sensor,

wherein the processor is configured to receive the sensor data through the sensor unit.

12. The artificial intelligence apparatus according to claim 1 , further comprising a communication unit configured to communicate with at least one external device,

wherein the processor is configured to receive the sensor data from the at least one external device through the communication unit.

13. The artificial intelligence apparatus according to claim 1 , wherein the processor is configured to:

determine whether the aleatoric uncertainty of the received sensor data is greater than or equal to the first reference value; and

delete the received sensor data.

14. The artificial intelligence apparatus according to claim 1 , wherein the processor is configured to:

determine whether the aleatoric uncertainty of the received sensor data is less than the first reference value and the epistemic uncertainty of the received sensor data is less than the second reference value; and

store the received sensor data in the memory as non-training data.

15. A method for generating training data, comprising:

receiving sensor data;

determining whether the received sensor data is relevant or irrelevant to a learning of a target artificial intelligence model, the received sensor data being relevant based on the received sensor data including an object of interest of the target artificial intelligence model and the received sensor data being irrelevant based on the received sensor data not including the object of interest;

generating a first result from the received sensor data by using the target artificial intelligence model;

calculating an aleatoric uncertainty regarding noise embedded in the received sensor data by determining Gaussian noise observed in the generated first result as the aleatoric uncertainty;

generating a plurality of second results from the received data by randomly activating or deactivating one or more hidden nodes included in a hidden layer of the target artificial intelligence model;

calculating an epistemic uncertainty by determining a variance of the generated second results as the epistemic uncertainty using the target artificial intelligence model with combinations of activation or deactivation of the one or more hidden nodes;

determining whether the calculated aleatoric uncertainty is less than a first reference value and the calculated epistemic uncertainty is greater than or equal to a second reference value;

extracting a label from the received sensor data by using a label extractor based on a determination that aleatoric uncertainty is less than the first reference value and the epistemic uncertainty is greater than or equal to the second reference value;

determining a confidence level of the extracted label; and

generating training data including the received sensor data and the extracted label based on the determined confidence level exceeding a third reference value.

16. A non-transitory recording medium having recorded thereon a program for performing a method for generating training data, the method comprising:

receiving sensor data;

determining whether the received sensor data is relevant or irrelevant to a learning of a target artificial intelligence model, the received sensor data being relevant based on the received sensor data including an object of interest of the target artificial intelligence model and the received sensor data being irrelevant based on the received sensor data not including the object of interest;

generating a first result from the received sensor data by using the target artificial intelligence model;

calculating an aleatoric uncertainty regarding noise embedded in the received sensor data by determining Gaussian noise observed in the generated first result as the aleatoric uncertainty;

generating a plurality of second results from the received data by randomly activating or deactivating one or more hidden nodes included in a hidden layer of the target artificial intelligence model;

calculating an epistemic uncertainty by determining a variance of the generated second results as the epistemic uncertainty using the target artificial intelligence model with combinations of activation or deactivation of the one or more hidden nodes;

determining whether the calculated aleatoric uncertainty is less than a first reference value and the calculated epistemic uncertainty is greater than or equal to a second reference value;

extracting a label from the received sensor data by using a label extractor based on a determination that the aleatoric uncertainty is less than the first reference value and the epistemic uncertainty is greater than or equal to the second reference value;

determining a confidence level of the extracted label; and

generating training data including the received sensor data and the extracted label based on the determined confidence level exceeding a third reference value.