Describir: Using drawings and deep neural networks to characterize the building blocks of human visual similarity.