Instead of sorting or sampling our dataset based on the loss of the output neurons, back-propagate the loss up to the input layer, basically using how much did every input neuron contributed to the loss, and the samples with the largest mean loss on its input pixels should be given priority. (idk random thought but why not)
Instead of sorting or sampling our dataset based on the loss of the output neurons, back-propagate the loss up to the input layer, basically using how much did every input neuron contributed to the loss, and the samples with the largest mean loss on its input pixels should be given priority. (idk random thought but why not)