Improving the usability of machine learning in industrial inspection systems.
The overall goal of the project is to enable the end-user to adequately deal with the complexity of automatic inspection systems during the set-up phase and during maintenance. The project aims at the following goals:
Active learning and in-line training:
- developing methodologies which reduce the requested number of class labels during off-line training and on-line adaptation of classifiers
- synthesizing artificial defect images that can be presented to acquire input in areas, where no real samples could yet be obtained
detecting a systematic change in the decision making (concept drift) and to efficiently update or correct earlier decision without the need of re-labelling a large part of the previous samples.
Adding new defect classes:
- acquiring high-level information about the defect class from the end-user in the presence of very little concrete data
data base search for similar samples and refining decision boundaries
Explaining and correcting false decisions:
- making progress in interpretability and representation of classifiers
- enabling the user to modify structural components and decision boundaries
- addressing reliability concepts (interpretability of classifiers outputs, reason finding)