Generating strong labels for examples labelled with weak labels
A method for generating strong labels for examples labelled with weak labels leverages an artificial neural network, or ANN, which is assumed to have been trained on a training set of examples labelled according to weak labels (e.g., classes of structural defects in images of civil engineering structures). The method processes each example of a set of test examples by performing the following operations. The trained ANN is first executed on each example to infer a weak label. Then, the method extracts explanatory features from the ANN as executed on the example. The method generates a strong label (e.g., a region boundary of the structural defect), based on the extracted explanatory features. The method subsequently prompts a user to react to one or each of the inferred weak label and the generated strong label. The response obtained is then interpreted by the method to obtain a further weak label.
1. A computer-implemented method of generating strong labels for examples labelled with weak labels, the method comprising:
given a machine learning (ML) model trained on a training set of examples labelled according to one or more previously identified weak labels, processing a set of test examples by:
executing the trained ML model on an example of the set of test examples to infer a weak label,
extracting, from the executed ML model, explanatory features that have contributed to infer the weak label using an extraction process based on an explainability method, and
generating a strong label based on the extracted explanatory features;
prompting, via a graphical user interface, a user to react to one or each of the inferred weak label and the generated strong label, to obtain a human response; and
interpreting the human response to generate a further weak label for the example.
2. The method according to claim 1 , wherein:
the trained ML model is an artificial neural network (ANN); and
the explanatory features are extracted from outputs of an internal neural layer of the ANN.
3. The method according to claim 2 , wherein the explanatory features are extracted as an activation map.
4. The method according to claim 3 , wherein:
the ANN is configured as a classifier;
the weak labels correspond to classes assigned in accordance with example features that are present in at least some of the examples of each of the training set and the test set; and
the explainability method is based on a class activation mapping method, whereby said activation map is extracted as a class activation map, in accordance with the inferred weak label.
5. The method according to claim 4 , wherein
the examples of the training set and the test set are images, and
the example features are image features.
6. The method according to claim 5 , wherein the strong labels include region boundaries of said image features.
7. The method according to claim 6 , wherein the region boundaries of the image features are generated through an image segmentation process designed to obtain a segmentation mask based on the extracted activation map.
8. The method according to claim 6 , wherein:
the images are images of materials;
the image features consists of several types of structural defects;
the classes correspond to the several types of structural defects; and
the boundaries are region boundaries of the structural defects.
9. The method according to claim 8 , wherein:
the images are images of surfaces of civil engineering structures; and
the structural defects relate to two or more selected from the group consisting of:
cracking, spalling, delamination, and rust.
10. The method according to claim 1 , wherein at least some of the test examples are examples that are sampled from the training set.
11. The method according to claim 10 , wherein:
the examples of the training set and the test set are images;
the graphical user interface is a graphical user interface of a manual image annotation tool;
prompting the user to react to one or each of the inferred weak label and the generated strong label comprises prompting the user to interact with the graphical user interface to modify the generated strong label, whereby the obtained human response includes a modified strong label; and
the method further comprises storing the modified strong label.
12. The method according to claim 11 , wherein:
the human response includes user interactions with the graphical user interface, beside the modified strong label; and
interpreting the human response comprises parsing the user interactions to obtain the further weak label.
13. The method according to claim 12 , wherein prompting the user to interact with the graphical user interface comprises prompting the user to decide whether to:
accept the modified strong label;
reject the modified strong label; or
modify the generated strong label.
14. The method according to claim 13 , wherein the method further comprises refining the extraction process based on the modified strong label.
15. The method according to claim 1 , wherein the method further comprises re-training the trained ML model based on the test examples as labelled with the further weak labels.
16. The method according to claim 15 , wherein the method further comprises:
training the ML model based on the training set of examples prior to processing each example; and
re-training the ML model according to a supervised learning scheme.
17. A method of identifying structural defects in images of surfaces of materials, wherein the method comprises:
given an artificial neural network, or ANN, the ANN configured as a classifier trained on a training set of examples of images of materials, the examples labelled with weak labels corresponding to classes of structural defects that are present in at least some of the images, processing a set of test examples by:
executing the trained classifier on an example of the set of test examples to infer a weak label, which corresponds to a class of a structural defect,
extracting a class activation map from outputs of an internal neural layer of the executed classifier,
generating a strong label, which includes a region boundary of a structural defect in the example, from a segmentation mask obtained from the extracted class activation map,
prompting a user to interact with a graphical user interface of a manual image annotation tool in respect of one or each of the inferred weak label and the generated strong label to obtain user interactions; and
parsing the user interactions to obtain a further weak label for the example.
18. The method according to claim 17 , wherein the method further comprises:
re-training the ANN based on the test examples as labelled with the further weak labels, according to a weak supervision learning scheme.
19. The method according to claim 17 , wherein the method further comprises:
partitioning an initial image into several image portions to form the examples of the test set.
20. A computer program product for generating strong labels for examples labelled with weak labels, the computer program product comprising:
one or more computer readable storage media and program instructions stored on the one or more computer readable storage media, the program instructions comprising instructions to:
process, given a machine learning (ML) model trained on a training set of examples labelled according to said weak labels, a set of test examples by:
executing the trained ML model on an example of the set of test examples to infer a weak label,
extracting, from the executed ML model, explanatory features that have contributed to infer said weak label, thanks to an extraction process based on an explainability method, and
generating a strong label based on the extracted explanatory features;
prompt, via a graphical user interface, a user to react to one or each of the inferred weak label and the generated strong label, to obtain a human response; and
interpret the human response to generate a further weak label for said each example, so as to obtain further weak labels for all of the test examples.