Statistics for Fine-Grained Zero-Shot Learning with DNA as Side Information