Machine learning using categorical uncertainty sampling
A system identifies and selects the best image to use to retrain a machine learning algorithm, thereby creating the best model. The best image is identified by determining the image, in an object identification, that has the most uncertainty. In addition to determining uncertainties for images, the system uses priority scores, groupings and orderings of subsets of images by the priority scores, computing a complement of recall or difficulty measure, and selecting images based on the difficulty measure.
1 . A process comprising:
receiving a model trained using a plurality of images;
receiving a plurality of unlabeled images;
providing the plurality of unlabeled images to the model;
receiving uncertainty scores from the model for each of the plurality of unlabeled images and for each occurrence of a particular class in each of the plurality of unlabeled images;
summing the uncertainty scores for each occurrence of the particular class in each of the plurality of unlabeled images;
assigning to each unlabeled image a priority score, the priority score comprising a largest uncertainty score of the particular classes in each unlabeled image;
associating with each of the plurality of unlabeled images the particular class having a largest priority score;
grouping the plurality of unlabeled images into subsets, each subset associated with a same particular class;
ordering the images of subsets by the priority score;
computing a complement of recall for each subset, thereby generating a difficulty score for each subset;
selecting a number of images from each of the subsets based on the difficulty score for each of the subsets; and
retraining the model with the selected number of images from each of the subsets.
2 . The process of claim 1 , wherein the uncertainty of each unlabeled image comprises an uncertainty of predictions for each of the particular classes of the unlabeled images.
3 . A non-transitory machine-readable medium comprising instructions that when executed by a processor executes a process comprising:
receiving a model trained using a plurality of images;
receiving a plurality of unlabeled images;
providing the plurality of unlabeled images to the model;
receiving uncertainty scores from the model for each of the plurality of unlabeled images and for each occurrence of a particular class in each of the plurality of unlabeled images;
summing the uncertainty scores for each occurrence of the particular class in each of the plurality of unlabeled images;
assigning to each unlabeled image a priority score, the priority score comprising a largest uncertainty score of the particular classes in each unlabeled image;
associating with each of the plurality of unlabeled images the particular class having a largest priority score;
grouping the plurality of unlabeled images into subsets, each subset associated with a same particular class;
ordering images of the subsets by the priority score;
computing a complement of recall for each subset, thereby generating a difficulty score for each subset;
selecting a number of images from each of the subsets based on the difficulty score for each of the subsets; and
retraining the model with the selected number of images from each of the subsets.
4 . The non-transitory machine-readable medium of claim 3 , wherein the uncertainty of each unlabeled image comprises an uncertainty of predictions for each of the particular classes of the unlabeled images.
5 . The non-transitory machine-readable medium of claim 3 , comprising annotating and labeling a portion of the plurality of unlabeled images from the subset with the lowest difficulty score.
6 . A process comprising:
providing a plurality of unlabeled images to a model, the model trained using a plurality of images;
receiving uncertainty scores from the model;
summing the uncertainty scores;
assigning to each unlabeled image a priority score;
associating with each of the plurality of unlabeled images a particular class having a largest priority score;
grouping the plurality of unlabeled images into subsets;
ordering the subsets by the priority score;
computing a complement of recall for each subset, thereby generating a difficulty score for each subset; and
selecting a number of images from each of the subsets based on the difficulty score for each of the subsets.
7 . The process of claim 6 , comprising retraining the model with the selected number of images from each of the subsets.
8 . The process of claim 6 , wherein the uncertainty scores received from the model are for each of the plurality of unlabeled images and for each occurrence of a particular class in each of the plurality of unlabeled images.
9 . The process of claim 6 , wherein the uncertainty scores are summed for each occurrence of the particular class in each of the plurality of unlabeled images.
10 . The process of claim 6 , wherein the priority score comprises a largest uncertainty score of the particular classes in each unlabeled image.
11 . The process of claim 6 , wherein each subset is associated with a same particular class.