given an unbalanced image classification task, I use CNN with undersampling of the majority class. In my context, undersampling gives quite stable results, even across different training sessions with different drawn majority instances. Nonetheless, I feel like drawing one undersampled data-set per epoch could improve the classification a bit;

Has anyone experience with this? Or is there some experimental or theoretical work on these undersampling schemes?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.