Learning Visual Reasoning Without Strong Priors