Is there any way to change the size of the training set during the learning process? For example, let's say we have four classes (with their distribution): [A (90%), B (5%), C(2%), D(3%)]. Can we first start with detecting if the test sample is from class A or Not, if Yes, quit! If Not, train the model on only [B, C, D] classes, and so on? I thought this technique may address the problem of highly imbalanced datasets, but I am not sure if it's a reasonable approach. Thanks.