Recognizing Variables from their Data via Deep Embeddings of Distributions