I tried to mix some negative samples when training my classifier under your framework. However, the validation accuracy drops from **94%** to **61%.** Should I never do that? Is it possible to modify the loss function to cope with negative samples?
I tried to mix some negative samples when training my classifier under your framework.
However, the validation accuracy drops from 94% to 61%.
Should I never do that?
Is it possible to modify the loss function to cope with negative samples?