Computer-implemented method, device, and computer program product for training a feature extraction model

1 . A computer-implemented method, comprising:

determining, in a classifier of a processor-based machine learning system, based on a set of sample features extracted from an input sample by a feature extraction model implemented utilizing processor and memory resources of the processor-based machine learning system, a confidence level of the input sample and a similarity degree among the set of sample features, wherein the classifier has an input coupled to an output of the feature extraction model, and an output coupled to an input of the feature extraction model, the classifier being arranged in a feedback loop of the processor-based machine learning system between the output of the feature extraction model and the input of the feature extraction model;

determining, in the classifier of the processor-based machine learning system, a first loss based on the confidence level, the set of sample features, and label information for the input sample, the first loss being related to the quality of the label information;

determining, in the classifier of the processor-based machine learning system, a second loss based on the similarity degree among the set of sample features, the second loss being related to the quality of the set of sample features, wherein the second loss is computed, utilizing a potential energy function, by iteratively combining results of application of the potential energy function to respective pairs of feature values in the set of sample features; and

training, via the feedback loop of the processor-based machine learning system, the feature extraction model based on the first loss and the second loss, the training taking into account confidence level and similarity degree for each of a plurality of samples including the input sample at least in part through weights assigned to respective ones of at least a subset of the plurality of samples in at least the first loss.

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

determining, based on a total similarity degree among points in a feature space and the set of sample features, an aggregation point for the set of sample features in the feature space; and

determining an aggregated similarity degree between the set of sample features and the aggregation point.

3 . The method according to claim 2 , further comprising:

determining the confidence level based on the aggregated similarity degree and a similarity degree threshold.

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

training the classifier based on the first loss and the second loss.

5 . The method according to claim 1 , wherein the input sample comprises at least one of an image, audio, or video.

6 . The method according to claim 1 , wherein the first loss is determined utilizing a weighted cross-entropy function.

7 . The method according to claim 1 , wherein the potential energy function implements a difference between a first term having a squared variable in its denominator and a second term having a cubed variable in its denominator.

8 . The method according to claim 1 , wherein the confidence level is determined at least in part as a function of the similarity degree.

9 . The method according to claim 8 , wherein the confidence level is set to a first value responsive to an aggregated similarity degree being above a similarity degree threshold, and is set to a second value different than the first value responsive to the aggregated similarity degree being at or below the similarity degree threshold.

10 . The method according to claim 8 , wherein the function comprises a linear function between a first value of the confidence level corresponding to an aggregated similarity degree of zero and a second value of the confidence level corresponding to an aggregated similarity degree equal to a similarity degree threshold.

11 . An electronic device, comprising:

at least one processor; and

at least one memory storing computer program instructions, wherein the at least one memory and the computer program instructions are configured to cause, together with the at least one processor, the electronic device to perform actions comprising:

determining, in a classifier of a processor-based machine learning system, based on a set of sample features extracted from an input sample by a feature extraction model implemented utilizing processor and memory resources of the processor-based machine learning system, a confidence level of the input sample and a similarity degree among the set of sample features, wherein the classifier has an input coupled to an output of the feature extraction model, and an output coupled to an input of the feature extraction model, the classifier being arranged in a feedback loop of the processor-based machine learning system between the output of the feature extraction model and the input of the feature extraction model;

determining, in the classifier of the processor-based machine learning system, a first loss based on the confidence level, the set of sample features, and label information for the input sample, the first loss being related to the quality of the label information;

determining, in the classifier of the processor-based machine learning system, a second loss based on the similarity degree among the set of sample features, the second loss being related to the quality of the set of sample features, wherein the second loss is computed, utilizing a potential energy function, by iteratively combining results of application of the potential energy function to respective pairs of feature values in the set of sample features; and

training, via the feedback loop of the processor-based machine learning system, the feature extraction model based on the first loss and the second loss, the training taking into account confidence level and similarity degree for each of a plurality of samples including the input sample at least in part through weights assigned to respective ones of at least a subset of the plurality of samples in at least the first loss.

12 . The electronic device according to claim 11 , wherein the actions further comprise:

determining, based on a total similarity degree among points in a feature space and the set of sample features, an aggregation point for the set of sample features in the feature space; and

determining an aggregated similarity degree between the set of sample features and the aggregation point.

13 . The electronic device according to claim 12 , wherein the actions further include:

determining the confidence level based on the aggregated similarity degree and a similarity degree threshold.

14 . The electronic device according to claim 11 , wherein the actions further comprise:

training the classifier based on the first loss and the second loss.

15 . The electronic device according to claim 11 , wherein the input sample includes at least one of an image, audio, or video.

16 . A computer program product tangibly stored on a non-volatile computer-readable medium and including machine-executable instructions, wherein the machine- executable instructions, when executed, cause a device to execute a method, the method comprising:

determining, in a classifier of a processor-based machine learning system, based on a set of sample features extracted from an input sample by a feature extraction model implemented utilizing processor and memory resources of the processor-based machine learning system, a confidence level of the input sample and a similarity degree among the set of sample features, wherein the classifier has an input coupled to an output of the feature extraction model, and an output coupled to an input of the feature extraction model, the classifier being arranged in a feedback loop of the processor-based machine learning system between the output of the feature extraction model and the input of the feature extraction model;

determining, in the classifier of the processor-based machine learning system, a first loss based on the confidence level, the set of sample features, and label information for the input sample, the first loss being related to the quality of the label information;

determining, in the classifier of the processor-based machine learning system, a second loss based on the similarity degree among the set of sample features, the second loss being related to the quality of the set of sample features, wherein the second loss is computed, utilizing a potential energy function, by iteratively combining results of application of the potential energy function to respective pairs of feature values in the set of sample features; and

training, via the feedback loop of the processor-based machine learning system, the feature extraction model based on the first loss and the second loss, the training taking into account confidence level and similarity degree for each of a plurality of samples including the input sample at least in part through weights assigned to respective ones of at least a subset of the plurality of samples in at least the first loss.

17 . The computer program product according to claim 16 , wherein the method further comprises:

determining, based on a total similarity degree among points in a feature space and the set of sample features, an aggregation point for the set of sample features in the feature space; and

determining an aggregated similarity degree between the set of sample features and the aggregation point.

18 . The computer program product according to claim 17 , wherein the method further comprises:

determining the confidence level based on the aggregated similarity degree and a similarity degree threshold.

19 . The computer program product according to claim 16 , wherein the method further comprises:

training the classifier based on the first loss and the second loss.

20 . The computer program product according to claim 16 , wherein the input sample comprises at least one of an image, audio, or video.