Automatically Tracking Metadata and Provenance of Machine Learning Experiments


We present a lightweight system to extract, store and manage metadata and provenance information of common artifacts in machine learning (ML) experiments: datasets, models, predictions, evaluations and training runs. Our system accelerates users in their ML workflow, and provides a basis for comparability and repeatability of ML experiments. We achieve this by tracking the lineage of produced artifacts and automatically extracting metadata such as hyperparameters of models, schemas of datasets or layouts of deep neural networks. Our system provides a general declarative representation of said ML artifacts, is integrated with popular frameworks such as MXNet, SparkML and scikit-learn, and meets the demands of various production use cases at Amazon.

Machine Learning Systems workshop at the conference on Neural Information Processing Systems (NIPS)