Looking for Definition of Open-World Learning Problem

The article discusses a machine learning project where the model had to handle more classes at inference than were present in the training data. The author used a metric learning approach to create an embedding space and cluster objects at inference.

Details

The author worked on a machine learning project where the initial training data had around 30 target classes. However, at inference, the model had to be able to handle a lot more classes than these 30 targets. This meant that a traditional classifier that predicts one of the 30 target classes would not work. Instead, the author adapted different flavors of metric learning approaches like arcface and cosface to create an embedding space that maximized inter-class distance and minimized intra-class distance. At inference, the model would cluster objects based on a similarity threshold, with the idea that objects in the same cluster would belong to the same target class. This approach worked surprisingly well, allowing the model to handle classes that it had never seen during training. The author is looking for the proper definition or term to describe this kind of open-world learning problem.