Local Aggregation for Unsupervised Learning of Visual Embeddings