Open zhuyr97 opened 1 year ago
Hello! Thank for your attention on our work.
There is not a thresholding function to define which samples are hard or easy in our method. We consider the classification scores as confidences for samples. For example, samples with higher confidences are considered easier, and those with lower confidences are considered harder. In addition, the classification scores closer to 0 or 1 are more confident, whereas those closer to 0.5 means lower confidences. Therefore, the hardness of samples will continuously change during training.
I wonder about the gradient visualizations of the different samples. Is it rely on the human to pre-defined?