Ensemble based labeling

1. A method comprising:

obtaining, by a processor, a plurality of unlabeled samples of an object;

estimating, by the processor, for each of the plurality of unlabeled samples, a probability that a label applies to the unlabeled sample, for each of a plurality of labels;

determining, by the processor, a candidate label among the plurality of labels based on the estimated probabilities of the plurality of unlabeled samples for each of the plurality of labels, determining the candidate label including:

calculating an average of the estimated probabilities of the plurality of unlabeled samples for each of the plurality of labels, and

determining a label that has a largest average among the plurality of labels, as the candidate label;

calculating, by the processor, a dispersion of the estimated probabilities of the plurality of unlabeled samples for the candidate label; and

identifying, by the processor, a target label among the plurality of labels based on an average of the estimated probabilities of the plurality of unlabeled samples for the candidate label, the dispersion for the candidate label, and a number of the plurality of unlabeled samples.

2. The method according to claim 1 , wherein the identifying a target label among the plurality of labels includes:

judging whether to identify the target label, based on the estimated probabilities of the plurality of unlabeled samples for the candidate label, the dispersion for the candidate label, and the number of the plurality of unlabeled samples, and

identifying the candidate label as the target label, in response to judging to identify the target label.

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

obtaining one or more additional unlabeled samples of the object in response to judging not to determine the target label;

estimating a probability that a label applies to each of the one or more additional unlabeled samples for each of a plurality of labels;

determining a candidate label among the plurality of labels, based on the estimated probabilities of the plurality of unlabeled samples and the one or more additional unlabeled samples for each of the plurality of labels; and

calculating a distribution of the estimated probabilities of the plurality of unlabeled samples and the one or more additional unlabeled samples for the candidate label;

wherein the identifying a target label among the plurality of labels is based on the estimated probabilities of the plurality of unlabeled samples and the one or more additional unlabeled samples for the candidate label, the distribution of the plurality of unlabeled samples and the one or more additional unlabeled samples for the candidate label, and the number of the plurality of unlabeled samples and the one or more additional unlabeled samples.

4. The method according to claim 1 , wherein the identifying a target label among the plurality of labels is based on a confidence interval for the candidate label.

5. The method according to claim 2 , wherein the judging whether to identify the target label includes judging whether a confidence interval of the candidate label is above a confidence interval of all other labels without an overlap, and

the identifying the candidate label as the target label is in response to judging that the confidence interval of the candidate label is above the confidence interval of all other labels without overlaps.

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

generating a reduced label set by removing at least one of the plurality of labels from the plurality of labels,

wherein the identifying a target label among the plurality of labels includes identifying a target label among a reduced label set.

7. The method according to claim 3 , wherein the one or more additional unlabeled samples are a plurality of additional unlabeled samples.

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

obtaining an initial unlabeled sample of the object;

estimating a probability that a label applies to the initial sample for each of a plurality of labels;

determining a candidate label among the plurality of labels, based on the estimated probabilities of the initial unlabeled sample for each of the plurality of labels;

judging whether an estimated probability of the initial unlabeled sample for the candidate label is above a threshold; and

identifying the candidate label as the target label, in response to judging that the estimated probability of the initial unlabeled sample for the candidate label is above the threshold.

9. The method according to claim 1 , wherein the estimating, for each of the plurality of unlabeled samples, a probability that a label applies to the unlabeled sample, for each of a plurality of labels, is performed by utilizing one or more neural networks.