$1\mathrm{.}1$ One of the requirements for training a regular feedforward ANN is to normalize the desired outputs in the training data so that they lie within the output range of the activation functions $($either Ingistic or tanh$)$ of the output neurons. Let a$,$b be the output range of such an activation function. Then, in the normalization process, you will map the domain of the desired output into a region $[$a$+,$b$-],$a small positive number, because a$)$ mapping exactly into a$,$b may be unsafe due to the limitations of machine $($calculation$)$ accuracy which might put certain values below or above the actual range a$,$b $2.$ this is redundant and may be set to zero c$)$ When the desired outputs become close to either of the bounds the derivative of the activation function $-$ which plays a crucial role in transmitting the error upstream becomes vanishingly small leading to inefficient training the desired outputs serve as "guides" for the predicted outputs and pull the latter towards themselves during the training process $($by modulation of weights$).$ As the predicted outputs shift towards the desired ones, If the latter are very close to the range bounds, the derivatives of the output activation function $($that generates the predicted outputs$)$ tend towards zero, limiting the progress of training.