It means it can be trained by backpropagating the error gradient through the network. To train a neural network, you want to know how much each component contr

It means it can be trained by backpropagating the error gradient through the network. To train a neural network, you want to know how much each component contributed to an error. We do that by propagating the error through each component in reverse, using the partial derivatives of the corresponding function.