OT-CAL uses optimal transport to adaptively select training samples and cut backprop cost
Researchers introduce OT-CAL, a data selection method that adjusts which samples receive backward passes as training progresses. Each sample is described by five informativeness features such as difficulty, learnability...