Dual Space Learning with variational Autoencoders