Documenting ML Experiments in HELIPORT

Abstract

Machine learning practitioners use a variety of tools to track their experiments. These tools have in common that they are only concerned with the machine learning aspect of the experiment: They may track model parameters or performance metrics of the model, but provenance of the training data or scientific outcomes produced with the trained model are largely overlooked. This is a drawback especially when it comes to experiments where machine learning meets scientific experiments and traditional simulations. In this contribution we present an initial evaluation on improving documentation of such machine learning experiments using our data management guidance system HELIPORT. We also explore existing experiment tracking tools and metadata schemas for ML experiments in the process and discuss their suitability for integration with HELIPORT.