Removing decision surface skew using complementary inputs
Tim Andersen · 2003
Examines the tendency of backpropagation-based training algorithms to favor examples that have large input feature values, in terms of the ability of such examples to influence the weights of the network, and shows that this tendency can lead to sub-optimal decision surfaces. We propose a method for counteracting this tendency that modifies the original input feature vector through the addition of complementary inputs.