Generation, Local Receptive Fields and Global Convergence Improve Perceptual Learning in Connectionist Networks