Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.
Sign up to join this community
What does the term "easy negatives" exactly mean in the context of machine learning for a classification problem or any problem in general?
From a quick google search, I think it means just negative examples in the training set.
Can someone please elaborate a bit more on why the term "easy" is brought into the picture?
Below, there is a screenshot taken from the paper where I found this term, which is underlined.
OK, I think I understood what this means.
Hard and easy negatives are the ones that have relatively large and small values for the loss function, respectively.
It refers to samples that are very easy for the model to classify. If you are interested in the positive class, having many easy negatives could produce misleading results as your model could really struggle to classify not-so-easy samples.
In a very hypothetical situation, imagine you are trying to classify brain scan images based on whether they show signs of a tumor or not. For the negative class, say you have a bunch of normal brain scans that have no tumor, but also a bunch of plain images, all black with nothing on them (you wouldn't have that, but let's imagine). For the positive class you got normal brain images with tumors. If you train a model, it might just learn that plain black pictures have no tumors, which is in fact true. Since half of your negative data have this kind of picture, the model could be virtually performing with a kinda-good accuracy but it would not be learning the actual problem.
