Learning-based clean data selection
Systems/techniques that facilitate learning-based clean data selection are provided. In various embodiments, a system can access a raw dataset. In various aspects, the system can select, via execution of a data selection machine learning model, a clean dataset from the raw dataset. In various instances, the system can train a target machine learning model to perform a target task based on the clean dataset. In various aspects, the clean dataset can include candidate-annotation groupings that are in the raw dataset and that are determined by the data selection machine learning model to be suitable for training of the target machine learning model, and the clean dataset can exclude candidate-annotation groupings that are in the raw dataset and that are determined by the data selection machine learning model to not be suitable for training of the target machine learning model.
1. A system, comprising:
a memory that stores computer-executable components; and
a processor that executes at least one of the computer-executable components that:
trains a target machine learning model to perform a target task associated with images according to a group of qualitative criteria, wherein at least one qualitative criterion of the group of qualitative criteria does not have an associated defined quantitative metric for assessing the at least one qualitative criterion, and wherein the training comprises:
first training of a data selection machine learning model to identify clean training candidate-annotation groupings for training the target machine learning model according to the group of qualitative criteria, wherein the first training comprises:
accessing a manually-labeled training dataset comprising first training candidate-annotation groupings, wherein respective first training candidate-annotation groupings comprise a respective first training candidate image, a respective first training ground-truth annotation image, and a respective manually assigned label indicating whether the respective training candidate-annotation grouping is clean or unclean according to the group of qualitative criteria; and
training, using the manually-labeled training dataset, the data selection machine learning model to identify the clean training candidate-annotation groupings for training the target machine learning model according to the group of qualitative criteria; and
second training of the target machine learning model to perform the target task associated with images according to the group of qualitative criteria, wherein the second training comprises:
accessing a raw dataset comprising second training candidate- annotation groupings, wherein respective second training candidate- annotation groupings comprise a respective second training candidate image and a respective second training ground-truth annotation image generated by performing the target task on the respective second training candidate image;
selecting, via execution of the data selection machine learning model, a clean dataset comprising clean second training candidate- annotation groupings according to the group of qualitative criteria from the raw dataset; and
training, using the clean dataset, the target machine learning model to perform the target task associated with images according to the group of qualitative criteria.
2. The system of claim 1 , wherein the at least one of the computer-executable components further:
accesses a set of image candidates encountered during deployment of the target machine learning model; and
executes, after training on the clean dataset, the target machine learning model to perform the target task according to the group of qualitative criteria on the set of image candidates.
3. The system of claim 1 , wherein the data selection machine learning model is trained to label respective first training candidate-annotation groupings as clean or unclean according to the group of qualitative criteria.
4. The system of claim 1 , wherein the respective first training ground-truth annotation image comprises a classification label if the target task is data classification, and wherein the respective first training ground-truth annotation image comprises a segmentation mask if the target task is data segmentation.
5. The system of claim 1 , wherein respective first training candidate-annotation groupings that are manually-labeled as clean are used for training of the target machine learning model, and wherein respective first training candidate-annotation groupings that are manually-labeled as unclean are not suitable used for training of the target machine learning model.
6. The system of claim 1 , wherein the target task is to produce a high quality image from a low quality image according to the group of qualitative criteria, wherein the high quality image has higher quality than the low quality image according to the group of qualitative criteria.
7. The system of claim 6 , wherein the group of qualitative criteria comprise at least one of more contrast, less noise, the high quality image exhibits a substantially similar speckle pattern as the low quality image, or each subarea structure depicted in the high quality image exhibits is also depicted in the low quality image.
8. A computer-implemented method, comprising:
training, by a system comprising a processor, a target machine learning model to perform a target task associated with images according to a group of qualitative criteria, wherein at least one qualitative criterion of the group of qualitative criteria does not have an associated defined quantitative metric for assessing the at least one qualitative criterion, and wherein the training comprises:
first training of a data selection machine learning model to identify clean training candidate-annotation groupings for training the target machine learning model according to the group of qualitative criteria, wherein the first training comprises:
accessing a manually-labeled training dataset comprising first training candidate-annotation groupings, wherein respective first training candidate-annotation groupings comprise a respective first training candidate image, a respective first training ground-truth annotation image, and a respective manually assigned label indicating whether the respective training candidate-annotation grouping is clean or unclean according to the group of qualitative criteria; and
training, using the manually-labeled training dataset, the data selection machine learning model to identify the clean training candidate-annotation groupings for training the target machine learning model according to the group of qualitative criteria; and
second training of the target machine learning model to perform the target task associated with images according to the group of qualitative criteria, wherein the second training comprises:
accessing a raw dataset comprising second training candidate-annotation groupings, wherein respective second training candidate-annotation groupings comprise a respective second training candidate image and a respective second training ground-truth annotation image generated by performing the target task on the respective second training candidate image;
selecting, via execution of the data selection machine learning model, a clean dataset comprising clean second training candidate-annotation groupings according to the group of qualitative criteria from the raw dataset; and
training, using the clean dataset, the target machine learning model to perform the target task associated with images according to the group of qualitative criteria.
9. The computer-implemented method of claim 8 , further comprising:
accessing, by the system, a set of image candidates encountered during deployment of the target machine learning model; and
executing, by the system and after training on the clean dataset, the target machine learning model to perform the target task according to the group of qualitative criteria on the set of image candidates.
10. The computer-implemented method of claim 8 , wherein the data selection machine learning model is trained to label respective first training candidate-annotation groupings as clean or unclean according to the group of qualitative criteria.
11. The computer-implemented method of claim 8 , wherein respective first training ground-truth annotation image comprises a classification label if the target task is data classification, and wherein the respective first training ground-truth annotation image comprises a segmentation mask if the target task is data segmentation.
12. The computer-implemented method of claim 8 , wherein respective first training candidate-annotation groupings that are manually-labeled as clean are used for training of the target machine learning model, and wherein respective first training candidate-annotation groupings that are manually-labeled as unclean are not used for training of the target machine learning model.
13. The computer-implemented method of claim 8 , wherein the target task is to produce a high quality image from a low quality image according to the group of qualitative criteria, wherein the high quality image has higher quality than the low quality image according to the group of qualitative criteria.
14. The computer-implemented method of claim 13 , wherein the group of qualitative criteria comprise at least one of more contrast, less noise, the high quality image exhibits a substantially similar speckle pattern as the low quality image, or each subarea structure depicted in the high quality image exhibits is also depicted in the low quality image.
15. A computer program product for facilitating learning-based clean data selection, the computer program product comprising a computer-readable memory having program instructions embodied therewith, the program instructions executable by a processor to cause the processor to:
train a target machine learning model to perform a target task associated with images according to a group of qualitative criteria, wherein at least one qualitative criterion of the group of qualitative criteria does not have an associated defined quantitative metric for assessing the at least one qualitative criterion, and wherein the training comprises:
first training of a data selection machine learning model to identify clean training candidate-annotation groupings for training the target machine learning model according to the group of qualitative criteria, wherein the first training comprises:
accessing a manually-labeled training dataset comprising first training candidate-annotation groupings, wherein respective first training candidate-annotation groupings comprise a respective first training candidate image, a respective first training ground-truth annotation image, and a respective manually assigned label indicating whether the respective training candidate-annotation grouping is clean or unclean according to the group of qualitative criteria; and
training, using the manually-labeled training dataset, the data selection machine learning model to identify the clean training candidate-annotation groupings for training the target machine learning model according to the group of qualitative criteria; and
second training of the target machine learning model to perform the target task associated with images according to the group of qualitative criteria, wherein the second training comprises:
accessing a raw dataset comprising second training candidate-annotation groupings, wherein respective second training candidate-annotation groupings comprise a respective second training candidate image and a respective second training ground-truth annotation image generated by performing the target task on the respective second training candidate image;
selecting, via execution of the data selection machine learning model, a clean dataset comprising clean second training candidate-annotation groupings according to the group of qualitative criteria from the raw dataset; and
training, using the clean dataset, the target machine learning model to perform the target task associated with images according to the group of qualitative criteria.
16. The computer program product of claim 15 , wherein the program instructions are further executable to cause the processor to:
access a set of image candidates encountered during deployment of the target machine learning model; and
execute, after training on the clean dataset, the target machine learning model to perform the target task according to the group of qualitative criteria on the set of image candidates.
17. The computer program product of claim 15 , wherein the data selection machine learning model is trained to label respective first training candidate-annotation groupings as clean or unclean according to the group of qualitative criteria.
18. The computer program product of claim 15 , wherein the respective first training ground-truth annotation image comprises a classification label if the target task is data classification, and wherein the respective first training ground-truth annotation image comprises a segmentation mask if the target task is data segmentation.
19. The computer program product of claim 17 , wherein respective first training candidate-annotation groupings that are manually-labeled as clean are used for training of the target machine learning model, and wherein respective first training candidate-annotation groupings that are manually-labeled as unclean are not used for training of the target machine learning model.
20. The computer program product of claim 17 , wherein the target task is to produce a high quality image from a low quality image according to the group of qualitative criteria, wherein the high quality image has higher quality than the low quality image according to the group of qualitative criteria.