Learning Generative Models with Visual Attention