Granted Patent B2

US 12,299,070 · App. 17/745,462 · Granted May 13, 2025

Inventors: Zijia Wang (WeiFang, CN); Jiacheng Ni (Shanghai, CN); Zhen Jia (Shanghai, CN)

Assignee: Dell Products L.P.

G06F18/21G06F18/241G06F2218/12

Abstract

Claims (81)

1. A method of evaluating samples, comprising:

receiving, at an edge device, a classification model from a cloud server, wherein the classification model is trained at the cloud server using a first sample set comprising a first number of samples;

receiving, at the edge device from the cloud server, a second sample set comprising a second number of samples less than the first number of samples, wherein the second sample set comprises a plurality of distilled samples obtained from the first sample set by a data distillation operation performed at the cloud server;

acquiring, at the edge device, a sample distribution corresponding to each class in a plurality of classes of the classification model;

acquiring, at the edge device from a terminal device having at least one Internet-of-Things (IoT) sensor, an input sample which is marked as a first class in the plurality of classes, wherein the input sample is captured by the IoT sensor of the terminal device and is marked with a label indicating the first class;

determining, at the edge device, whether the input sample conforms to a first sample distribution corresponding to the first class;

identifying, at the edge device and in response to the input sample conforming to the first sample distribution, the input sample as a trusted sample, the trusted sample indicating that the input sample is correctly marked with the label indicating the first class; and

retraining the classification model, at the edge device, utilizing the distilled samples and the trusted sample.

2. The method according to claim 1 , wherein the determining whether the input sample conforms to a first sample distribution corresponding to the first class comprises:

determining an input sample distribution of the input sample;

determining a similarity between the input sample distribution and the first sample distribution; and

determining, on the basis of the similarity, whether the input sample conforms to the first sample distribution.

3. The method according to claim 1 , further comprising:

receiving, at the edge device, a given one of the distilled samples from the cloud server, the given distilled sample being obtained by distilling initial samples at the cloud server, and the classification model being trained on the basis of the initial samples; and

retraining, by the edge device, the classification model by using the given distilled sample and the trusted sample.

4. The method according to claim 3 , further comprising:

sending, in response to determining that at least one of the initial samples at the cloud server is lost, the given distilled sample to the cloud server.

5. The method according to claim 1 , further comprising:

sending the trusted sample from the edge device to the cloud server; and

receiving an updated classification model from the cloud server, the updated classification model being retrained by the cloud server on the basis of initial samples and the trusted sample.

6. The method according to claim 5 , further comprising:

identifying, in response to the input sample not conforming to the first sample distribution, the input sample as an untrusted sample, the untrusted sample indicating that the input sample is not correctly marked; and

sending the untrusted sample to the cloud server; and

receiving an updated classification model from the cloud server, the updated classification model being retrained on the basis of the untrusted sample, the trusted sample, and the initial samples, and the untrusted sample being correctly marked using a self-learning model before the retraining.

7. The method according to claim 1 , further comprising:

determining a center of a distribution region of a sample that has the same class as the trusted sample in a feature space;

determining a distance between the trusted sample and the center in the feature space; and

identifying, in response to the distance being greater than a predetermined threshold, the trusted sample as a useful sample.

8. The method according to claim 7 , further comprising:

sending the useful sample to the cloud server; and

receiving an updated classification model from the cloud server, the updated classification model being trained on the basis of the useful sample and initial samples.

9. An electronic device, comprising:

at least one processor; and

memory coupled to the at least one processor, wherein the memory has instructions stored therein, and the instructions, when executed by the at least one processor, cause the electronic device to execute actions comprising:

receiving, at an edge device, a classification model from a cloud server, wherein the classification model is trained at the cloud server using a first sample set comprising a first number of samples;

receiving, at the edge device from the cloud server, a second sample set comprising a second number of samples less than the first number of samples, wherein the second sample set comprises a plurality of distilled samples obtained from the first sample set by a data distillation operation performed at the cloud server;

acquiring, at the edge device, a sample distribution corresponding to each class in a plurality of classes of the classification model;

acquiring, at the edge device from a terminal device having at least one Internet-of-Things (IoT) sensor, an input sample which is marked as a first class in the plurality of classes, wherein the input sample is captured by the IoT sensor of the terminal device and is marked with a label indicating the first class;

determining, at the edge device, whether the input sample conforms to a first sample distribution corresponding to the first class;

identifying, at the edge device and in response to the input sample conforming to the first sample distribution, the input sample as a trusted sample, the trusted sample indicating that the input sample is correctly marked with the label indicating the first class; and

retraining the classification model, at the edge device, utilizing the distilled samples and the trusted sample.

10. The electronic device according to claim 9 , wherein the determining whether the input sample conforms to a first sample distribution corresponding to the first class comprises:

determining an input sample distribution of the input sample;

determining a similarity between the input sample distribution and the first sample distribution; and

determining, on the basis of the similarity, whether the input sample conforms to the first sample distribution.

11. The electronic device according to claim 9 , further comprising:

receiving, at the edge device, a given one of the distilled samples from the cloud server, the given distilled sample being obtained by distilling initial samples at the cloud server, and the classification model being trained on the basis of the initial samples; and

retraining, by the edge device, the classification model by using the given distilled sample and the trusted sample.

12. The electronic device according to claim 11 , further comprising:

sending, in response to determining that at least one of the initial samples at the cloud server is lost, the given distilled sample to the cloud server.

13. The electronic device according to claim 9 , further comprising:

sending the trusted sample from the edge device to the cloud server; and

receiving an updated classification model from the cloud server, the updated classification model being retrained by the cloud server on the basis of initial samples and the trusted sample.

14. The electronic device according to claim 13 , further comprising:

identifying, in response to the input sample not conforming to the first sample distribution, the input sample as an untrusted sample, the untrusted sample indicating that the input sample is not correctly marked; and

sending the untrusted sample to the cloud server; and

receiving an updated classification model from the cloud server, the updated classification model being retrained on the basis of the untrusted sample, the trusted sample, and the initial samples, and the untrusted sample being correctly marked using a self-learning model before the retraining.

15. The electronic device according to claim 9 , further comprising:

determining a center of a distribution region of a sample that has the same class as the trusted sample in a feature space;

determining a distance between the trusted sample and the center in the feature space; and

identifying, in response to the distance being greater than a predetermined threshold, the trusted sample as a useful sample.

16. The electronic device according to claim 15 , further comprising:

sending the useful sample to the cloud server; and

receiving an updated classification model from the cloud server, the updated classification model being trained on the basis of the useful sample and initial samples.

17. A computer program product that is tangibly stored on a non-transitory computer-readable medium and comprises machine-executable instructions, wherein the machine-executable instructions, when executed by a machine, cause the machine to perform actions comprising:

receiving, at an edge device, a classification model from a cloud server, wherein the classification model is trained at the cloud server using a first sample set comprising a first number of samples;

receiving, at the edge device from the cloud server, a second sample set comprising a second number of samples less than the first number of samples, wherein the second sample set comprises a plurality of distilled samples obtained from the first sample set by a data distillation operation performed at the cloud server;

acquiring, at the edge device, a sample distribution corresponding to each class in a plurality of classes of the classification model;

acquiring, at the edge device from a terminal device having at least one Internet-of-Things (IoT) sensor, an input sample which is marked as a first class in the plurality of classes, wherein the input sample is captured by the IoT sensor of the terminal device and is marked with a label indicating the first class;

determining, at the edge device, whether the input sample conforms to a first sample distribution corresponding to the first class;

identifying, at the edge device and in response to the input sample conforming to the first sample distribution, the input sample as a trusted sample, the trusted sample indicating that the input sample is correctly marked with the label indicating the first class; and

retraining the classification model, at the edge device, utilizing the distilled samples and the trusted sample.

18. The computer program product according to claim 17 , wherein the determining whether the input sample conforms to a first sample distribution corresponding to the first class comprises:

determining an input sample distribution of the input sample;

determining a similarity between the input sample distribution and the first sample distribution; and

determining, on the basis of the similarity, whether the input sample conforms to the first sample distribution.

19. The computer program product according to claim 17 , further comprising:

receiving, at the edge device, a given one of the distilled samples from the cloud server, the given distilled sample being obtained by distilling initial samples at the cloud server, and the classification model being trained on the basis of the initial samples; and

retraining, by the edge device, the classification model by using the given distilled sample and the trusted sample.

20. The computer program product according to claim 19 , further comprising:

sending, in response to determining that at least one of the initial samples at the cloud server is lost, the given distilled sample to the cloud server.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (14)

Cited By (1)

US 12,548,309