I'm looking at ways in which "ambiguities" can be identified in labeled training data by a system undergoing some sort of inductive learning process. Do you know if there is any literature on this topic, and if so, where I could find that literature / what keywords to search for?
To better illustrate what I'm looking for, there is a classic parable of machine learning told by e.g. Dreyfus and Dreyfus (What Artificial Experts Can and Cannot Do, 1992) of an algorithm intended to classify whether or not pictures of woods contained a tank concealed between the trees. Pictures of empty woods were taken one day; pictures with concealed tanks were taken the next. The classifier identified the latter set with great accuracy, and tested extremely well on the portion of the data that had been withheld from training. However, the system performed poorly on new images. The first set of pictures were taken on a sunny day, whereas the latter were taken on a cloudy day. The classifier was not identifying tanks, it was identifying image brightness.
This leads to a question of how, in principle, one could take labeled training data and generate "ambiguities" that could be raised to user attention. The goal is not to generate "borderline" cases (e.g. images with medium brightness), the goal is to identify additional features in the data that could explain the training labels so that these can be turned into queries (e.g. "these labels can be explained both by brightness and by tanks").
I can think of a number of ways to approach this problem, but I'm sure it's already been tackled in various different ways. Can anyone point me towards the literature on this question?