Method and apparatus with input data classification
A processor-implemented method with input data classification includes: extracting an input embedding vector including a feature of biometric information of a user from input data including the biometric information; determining an adaptive embedding vector adaptive to the input embedding vector, based on a combination of a plurality of enrollment embedding vectors that are based on enrollment data; and classifying the input data based on a similarity between the input embedding vector and the adaptive embedding vector.
1. A processor-implemented method with input data classification, the method comprising:
extracting an input embedding vector including a feature of biometric information of a user from input data including the biometric information;
determining an adaptive embedding vector adaptive to the input embedding vector, wherein a value of the adaptive embedding vector is determined by combining a plurality of enrollment embedding vectors that are based on enrollment data; and
classifying the input data based on a similarity between the input embedding vector and the adaptive embedding vector.
2. The method of claim 1 , wherein the determining of the adaptive embedding vector comprises:
determining a combination embedding vector by combining enrollment embedding vectors adaptive to the input embedding vector among the plurality of enrollment embedding vectors; and
generating the adaptive embedding vector by updating the combination embedding vector based on a loss function between the combination embedding vector and the input embedding vector.
3. The method of claim 2 , wherein the generating of the adaptive embedding vector comprises generating the adaptive embedding vector by iteratively updating the combination embedding vector so that the loss function converges to a first value.
4. The method of claim 2 , wherein
the determining of the combination embedding vector comprises determining the combination embedding vector by combining the enrollment embedding vectors adaptive to the input embedding vector among the plurality of enrollment embedding vectors, through linear interpolation, and
the generating of the adaptive embedding vector comprises:
updating the combination embedding vector by applying a weight for the linear interpolation to the combination embedding vector; and
generating the adaptive embedding vector by updating the weight based on an error between the updated combination embedding vector and the input embedding vector.
5. The method of claim 4 , further comprising:
receiving a hyperparameter indicating a degree to which the weight is updated,
wherein the generating of the adaptive embedding vector comprises:
updating the weight by applying the hyperparameter to the error; and
generating the adaptive embedding vector based on a loss function between the input embedding vector and the updated combination embedding vector to which the updated weight is applied.
6. The method of claim 5 , wherein the generating of the adaptive embedding vector comprises generating the adaptive embedding vector by iteratively updating the weight so that the loss function between the input embedding vector and the updated combination embedding vector to which the updated weight is applied converges to a first value.
7. The method of claim 4 , further comprising:
deriving a hyperparameter indicating a degree to which the weight is updated, using a power iteration algorithm,
wherein the generating of the adaptive embedding vector comprises generating the adaptive embedding vector by iteratively updating the weight by applying the hyperparameter to the error.
8. The method of claim 7 , wherein the deriving of the hyperparameter comprises:
determining a correlation matrix corresponding to the plurality of enrollment embedding vectors; and
deriving the hyperparameter based on the correlation matrix.
9. The method of claim 8 , wherein the deriving of the hyperparameter based on the correlation matrix comprises:
determining an eigenvalue corresponding to the correlation matrix using the power iteration algorithm; and
deriving the hyperparameter based on the eigenvalue.
10. The method of claim 9 , wherein the determining of the eigenvalue comprises:
determining an eigenvector corresponding to the correlation matrix using the power iteration algorithm;
iteratively updating the eigenvector until the eigenvector converges; and
determining the eigenvalue corresponding to the updated eigenvector.
11. The method of claim 7 , wherein the generating of the adaptive embedding vector by iteratively updating the weight comprises:
differently scheduling the hyperparameter each time the weight is iteratively updated; and
generating the adaptive embedding vector by updating the weight by differently scheduled hyperparameters.
12. The method of claim 1 , wherein the classifying of the input data comprises either one or both of:
authenticating the input data based on the similarity; and
determining whether the input data is spoofed based on the similarity.
13. A non-transitory computer-readable storage medium storing instructions that, when executed by a processor, configure the processor to perform the method of claim 1 .
14. A processor-implemented method with input data classification, the method comprising:
extracting an input embedding vector including a feature of biometric information of a user from input data including the biometric information;
determining a combination embedding vector by combining enrollment embedding vectors adaptive to the input embedding vector among a plurality of enrollment embedding vectors included in a cluster based on enrolled images, through linear interpolation;
determining an error between the input embedding vector and the combination embedding vector that is updated by applying a weight for the linear interpolation to the combination embedding vector;
deriving a hyperparameter indicating a degree to which the weight is updated, based on a correlation matrix corresponding to the plurality of enrollment embedding vectors;
generating an embedding vector adaptive to the input embedding vector by iteratively updating the weight by applying the hyperparameter to the error; and
classifying whether the input data is spoofed based on a similarity between the input embedding vector and the adaptive embedding vector.
15. The method of claim 14 , wherein the generating of the adaptive embedding vector by iteratively updating the weight comprises:
differently scheduling the hyperparameter for each iteration at which the weight is iteratively updated; and
generating the adaptive embedding vector by updating the weight by differently scheduled hyperparameters.
16. An apparatus with input data classification, the apparatus comprising:
a sensor configured to capture input data including biometric information of a user;
a memory configured to store a cluster including a plurality of enrollment embedding vectors that are based on enrollment data; and
a processor configured to:
extract an input embedding vector including a feature of the biometric information from the input data;
determine an adaptive embedding vector adaptive to the input embedding vector, wherein a value of the adaptive embedding vector is determined by combining the plurality of enrollment embedding vectors; and
classify the input data based on a similarity between the input embedding vector and the adaptive embedding vector.
17. The apparatus of claim 16 , wherein, for the determining of the adaptive embedding vector, the processor is configured to:
determine a combination embedding vector by combining enrollment embedding vectors adaptive to the input embedding vector among the plurality of enrollment embedding vectors; and
generate the adaptive embedding vector by updating the combination embedding vector based on a loss function between the combination embedding vector and the input embedding vector.
18. The apparatus of claim 17 , wherein, for the generating of the adaptive embedding vector, the processor is configured to generate the adaptive embedding vector by iteratively updating the combination embedding vector so that the loss function converges to a first value.
19. The apparatus of claim 17 , wherein the processor is configured to:
for the determining of the combination embedding vector, determine the combination embedding vector by combining the enrollment embedding vectors adaptive to the input embedding vector among the plurality of enrollment embedding vectors, through linear interpolation; and
for the generating of the adaptive embedding vector,
update the combination embedding vector by applying a weight for the linear interpolation to the combination embedding vector; and
generate the adaptive embedding vector by updating the weight based on an error between the updated combination embedding vector and the input embedding vector.
20. The apparatus of claim 19 , wherein the processor is configured to:
derive a hyperparameter indicating a degree to which the weight is updated, based on a correlation matrix corresponding to the plurality of enrollment embedding vectors; and
for the generating of the adaptive embedding vector, generate the adaptive embedding vector by iteratively updating the weight by applying the hyperparameter to the error.
21. A processor-implemented method with input data classification, the method comprising:
extracting an input embedding vector from input data including biometric information;
determining a loss between the input embedding vector and a combination embedding vector generated by combining a plurality of enrollment embedding vectors;
generating an adaptive embedding vector by updating the combination embedding vector until the loss converges to a predetermined value; and
classifying the input data based on a similarity between the input embedding vector and the adaptive embedding vector.
22. The method of claim 21 , wherein the combination embedding vector is a combination of the enrollment embedding vectors among a plurality of combinations of the enrollment embedding vectors.
23. The method of claim 22 , wherein the adaptive embedding vector is the updated combination vector for which the loss converges to the predetermined value.
24. The method of claim 21 , further comprising obtaining the biometric information using one or more sensors.
25. The method of claim 21 , further comprising controlling access to one or more functions of an electronic apparatus based on a result of the classifying.