Lifting inference with kernel embeddings