tensorflow - Best loss function for very small numbers -

what appropriate loss function neural network inwhich outputs include small , (large) numbers?

e.g. consider following inputs , outputs:

nn_inputs = [-1e-7, -2, 2e3] nn_outputs = [-2e-5, 200] 

just want show variation of inputs (which cause problem in normalization) , outputs.

when use mse loss function ( in tensorflow: loss = tf.reduce_mean(tf.squared_difference(nn_outputs, targets), loss have small value.