B.A. Golomb, D.T. Lawrence, T.J. Sejnowski
Sex identification in animals has biological importance. Humans are good at making this determination visually, but machines have not matched this ability. A neural network was trained to discriminate sex in human faces, and performed as well as humans on a set of 90 exemplars. Images sampled at 30x30 were compressed using a 900x40x900 fully-connected back-propagation network; activities of hidden units served as input to a back-propagation "SexNet" trained to produce values of 1 for male and o for female faces. The network's average error rate of 8.1% compared favorably to humans, who averaged 11.6%. Some SexNet errors mimicked those of humans.