Efficient pooling of predictions via kernel embeddings