Retraining a machine-learning classifier using re-labeled training samples
View Patent ↗Provided are systems, methods and techniques for machine learning. In one representative embodiment, a training set that includes training samples and corresponding assigned classification labels is obtained, and an automated classifier is trained against the training set. At least one of the training samples is selected and confirmation/re-labeling of it is requested. In response, a reply classification label is received and is used to retrain the automated classifier.
1. A method of machine learning, comprising:
(a) obtaining a training set that includes training samples and corresponding assigned classification labels;
(b) training an automated classifier against the training set;
(c) selecting at least one of the training samples;
(d) requesting confirmation/re-labeling of said at least one training sample;
(e) in response to step (d), receiving a reply classification label for said at least one training sample;
(f) retraining the automated classifier using the reply classification label; and, predicting a label for said at least one training sample,
wherein said at least one training sample is selected in step (c) based on a comparison between the assigned classification label and the predicted label for said at least one training sample.
2. A method according to claim 1 , wherein said at least one training sample is selected in step (c) after considering for possible selection a group of samples that includes a plurality of the training samples and a plurality of unlabeled samples.
3. A method according to claim 1 , wherein the reply classification label has been assigned by a human being.
4. A method according to claim 1 , wherein said requesting step (d) comprises providing the predicted label for said at least one training sample to a user and asking the user to indicate whether the predicted label for said at least one training sample is correct.
5. A method according to claim 1 , wherein in said requesting step (d) plural training samples are presented for confirmation/re-labeling as a group, arranged based on predicted labels for said plural training samples.
6. A method according to claim 1 , said requesting step (d) comprises providing the assigned classification label and the predicted label for said at least one training sample to a user and allowing the user to select one of the assigned classification label and the predicted label for said at least one training sample.
7. A method according to claim 1 , wherein plural similar training samples are selected and presented for confirmation/re-labeling, as a group of similar training samples, in said requesting step (d).
8. A method according to claim 1 , further comprising a step of tagging the reply classification label for said at least one training sample as having a stronger presumption of accuracy than the assigned classification label for said at least one training sample.
9. A method according to claim 1 , wherein said at least one training sample is selected based on at least one of: (i) an estimation of an effect of changing the assigned classification label for said at least one training sample; (ii) a result of a cross- validation calculation; (iii) a score representing a fit of the at least one training sample to a model based on the training set; (iv) an active-learning calculation; (v) a consistency evaluation; (vi) a prediction-confidence evaluation; (vii) a variety evaluation; and (viii) a calculated measure indicating confidence in the assigned classification label for said at least one training sample.
10. A method of supplementing training for a machine-learning classifier, comprising:
(a) obtaining a training set that includes training samples and corresponding assigned classification labels, the training set having been used to train a machine-learning classifier;
(b) obtaining a set of unlabeled samples;
(c) selecting a sample from among the training set and the set of unlabeled samples based on a specified selection criterion that involves evaluation of a plurality of both the training samples and the unlabeled samples;
(d) submitting the selected sample for labeling and, in response, receiving a reply classification label for the selected sample; and
(e) retraining the machine-learning classifier using the selected sample and the reply classification label for the selected sample,
wherein the specified selection criterion comprises predicting labels for a group of the training samples and comparing the predicted labels to the assigned classification labels for the group of training samples.
11. A method according to claim 10 , further comprising a step of repeating steps (c)-(e) a plurality of times, and wherein in the repetitions of step (c) samples are selected from both the training set and from the set of unlabeled samples.
12. A method according to claim 10 , wherein the specified selection criterion further comprises calculating a score for each training sample and for each unlabeled sample using an active-learning technique.
13. A method according to claim 12 , wherein the selected sample is selected in step (c) based on calculated scores from among samples in the training set and samples in the set of unlabeled samples.
14. A method according to claim 10 , wherein the specified selection criterion comprises a variety constraint which biases toward selecting additional training samples in subspaces where the unlabeled samples are inadequately represented by the training samples.
15. A method according to claim 10 , wherein plural similar samples are selected in step (c) and submitted, as a group of similar training samples, in step (d).
16. A method according to claim 10 , wherein plural selected samples are selected in step (c) and submitted in step (d) for labeling as a group, arranged based on predicted labels for said selected samples.
17. A method according to claim 10 , wherein in step (d) the selected sample is submitted together with a designated label and a request to confirm whether the designated label is correct.
18. A method of supplementing training for a machine-learning classifier, comprising:
(a) obtaining a training set that includes training samples and corresponding assigned classification labels, the training set having been used to train a machine-learning classifier;
(b) obtaining a set of unlabeled samples;
(c) selecting a sample from among the training set and the set of unlabeled samples based on a specified selection criterion that involves evaluation of a plurality of both the training samples and the unlabeled samples;
(d) submitting the selected sample for labeling and, in response, receiving a reply classification label for the selected sample; and
(e) retraining the machine-learning classifier using the selected sample and the reply classification label for the selected sample,
wherein plural selected samples are selected in step (c) and submitted in step (d) for labeling as a group, arranged based on predicted labels for said selected samples.