Machine learning systems and methods with source-target adaptation

1. An artificial intelligence method for domain adaptation between a plurality of source domains and a target domain, the artificial intelligence method comprising:

receiving labeled data from the plurality of source domains and unlabeled data from the target domain;

separately training, by a processor, a plurality of source classifiers each corresponding to a source domain using the labeled data received from the respective source domains;

selecting a subset of the labeled data received from each source domain based on a similarity between the selected labeled data and the unlabeled data of the target domain;

refining, by the processor, each source classifier using the selected subset of the labeled data; and

predicting labels of the unlabeled data using the refined source classifiers.

2. The artificial intelligence method of claim 1 , further comprising:

separately training target encoders that map target features extracted from the unlabeled data to the respective source domains.

3. The artificial intelligence method of claim 2 , wherein each target encoder is trained to optimizing a distance between the mapped target features and source features extracted from the labeled data of the respective source domain.

4. The artificial intelligence method of claim 3 , wherein the similarity between the selected labeled data of the respective source domain and the target domain is determined by:

mapping target features extracted from the unlabeled data to the respective source domain using the trained target encoders; and

calculating the distance between the source features extracted from each unlabeled data of the respective source domain and the mapped target features.

5. The artificial intelligence method of claim 4 , wherein the selected subset of the labeled data include a predetermined percentage of the labeled data received from the respective source domain, and the predetermined percentage of the labeled data are associated with calculated distances smaller than those of remaining labeled data.

6. The artificial intelligence method of claim 3 , wherein the distance is a Wasserstein distance.

7. The artificial intelligence method of claim 3 , wherein optimizing the distance is further balanced by a gradient penalty of a discriminator adversarily trained to counter the corresponding target encoder.

8. The artificial intelligence method of claim 1 , wherein the learning network further comprises a plurality of feature extractors each corresponding a source domain, wherein the feature extractors extract features from the labeled data of the respective source domains, wherein the artificial intelligence method further comprises:

jointly training each feature extractor with the source classifier corresponding to the same source domain using the labeled data received from that source domain.

9. The artificial intelligence method of claim 1 , wherein jointly refining the source classifiers further comprises optimizing a cross-entropy loss determined using the selected subsets of labeled data from the plurality of source domains.

10. The artificial intelligence method of claim 1 , wherein predicting labels of the unlabeled data using the refined source classifiers further comprises:

performing predictions by applying the refined source classifiers to the unlabeled data of the target domain; and

aggregating the predictions weighted by respective domain weights each corresponding to a source domain.

11. The artificial intelligence method of claim 10 , wherein each domain weight is indicative of a discrepancy between the corresponding source domain and the target domain.

12. A system for domain adaptation between a plurality of source domains and a target domain, the system comprising:

a communication interface configured to labeled data from the plurality of source domains and unlabeled data from the target domain; and

at least one processor coupled to the communication interface and configured to:

separately train a plurality of source classifiers each corresponding to a source domain using the labeled data received from the respective source domains;

select a subset of the labeled data received from each source domain based on a similarity between the selected labeled data and the unlabeled data of the target domain;

jointly refine the source classifiers using the selected subsets of labeled data from the plurality of source domains; and

predict labels of the unlabeled data using the refined source classifiers.

13. The system of claim 12 , wherein the at least one processor is further configured to:

separately train target encoders that map target features extracted from the unlabeled data to the respective source domains, wherein each target encoder is trained to optimizing a distance between the mapped target features and source features extracted from the labeled data of the respective source domain.

14. The system of claim 13 , wherein to determine the similarity between the selected labeled data of the respective source domain and the target domain, the at least one processor is further configured to:

map target features extracted from the unlabeled data to the respective source domain using the trained target encoders; and

calculate the distance between the source features extracted from each unlabeled data of the respective source domain and the mapped target features.

15. The system of claim 14 , wherein the selected subset of the labeled data include a predetermined percentage of the labeled data received from the respective source domain, that the predetermined percentage of the labeled data are associated with calculated distances smaller than those of remaining labeled data.

16. The system of claim 13 , wherein the distance is a Wasserstein distance.

17. The system of claim 12 , wherein the learning network further comprises a plurality of feature extractors each corresponding a source domain, wherein the feature extractors extract features from the labeled data of the respective source domains, wherein the at least one processor is further configured to:

jointly train each feature extractor with the source classifier corresponding to the same source domain using the labeled data received from that source domain.

18. The system of claim 12 , wherein to jointly refine the source classifiers, the at least one processor is further configured to optimize a cross-entropy loss determined using the selected subsets of labeled data from the plurality of source domains.

19. The system of claim 12 , wherein to predict labels of the unlabeled data using the refined source classifiers, the at least processor is further configured to:

perform predictions by applying the refined source classifiers to the unlabeled data of the target domain; and

aggregate the predictions weighted by respective domain weights each corresponding to a source domain, wherein each domain weight is indicative of a discrepancy between the corresponding source domain and the target domain.

20. A non-transitory computer-readable medium storing instructions that, when executed by a processor, cause the processor to perform an artificial intelligence method for domain adaptation between a plurality of source domains and a target domain, the artificial intelligence method comprising:

receiving labeled data from the plurality of source domains and unlabeled data from the target domain;

separately training a plurality of source classifiers each corresponding to a source domain using the labeled data received from the respective source domains;

selecting a subset of the labeled data received from each source domain based on a similarity between the selected labeled data and the unlabeled data of the target domain;

refining each source classifier using the selected subset of the labeled data; and

predicting labels of the unlabeled data using the refined source classifiers.