Decting new classes (open world classification)

Question

In real world applications very often the entire set of classes is not known during the training phase (e.g. identifying objects, sounds, etc.). A system is needed that can classify observations into the predefined classes but also tells the user that there is an observation that probably is related to a completely new class which was not in the training data - sort of a combination of classification and anomaly detection.

I saw that this problem is also called "open world classification" or "open set classification" with a few papers addressing this issue.

A simpler approach would be just take the prediction probability of a neural network and interpret a low probability as an indication of a new class.

Do you see any problem with the latter approach? What is your opinion about that "open world classification" research?

Thanks

Answer 1

This approach may work well for points that are close to the training distribution. The prediction of the neural network for points outside the training distribution cannot be trusted a lot. It may happen that for points outside the training distribution the predictions of the neural network are somewhat strange and erratic. For more information on this topic have a look at this paper:

https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=1176122

Under the "IMPROPER BEHAVIOR OF MLP OUTSIDE THE BOUNDARY OF THE TRAINING SAMPLE" paragraph, you will find more information on this problem.