Is there a way to access the real label of a training sample for use during the backprop process? I’m implementing this paper:
Interpretable CNNs
and, to do the optimizations in the backpropagation section, I need to calculate c-hat, which is computed using the real labels, as described in section 3.2.