[D] Looking for definition of open-world ish learning problem

About this article

Hello! Recently I did a project where I initially had around 30 target classes. But at inference, the model had to be able to handle a lot more classes than these 30 targets i had in my training data. Therefore, I couldn’t just make a ”normal” classifier that predicts one of the 30 target classes. I instead went with a metric learning approach where i adapted different flavors of arcface/cosface etc. to create an embedding space that tried to maximize inter cosine distance, and minimize intra...