Normalization Matters in Zero-Shot Learning