Algorithm reproducibility means rerunning a deep learning algorithm on a specified dataset and obtaining the same or similar quality of results. Versioned training datasets make the exact files used in earlier runs retrievable. section 2.1.1
It supports trustworthy machine learning systems and makes performance regression troubleshooting practical: teams can investigate changes in training data and algorithm code instead of debugging an unrecoverable run. section 2.1.1
A reproducibility workflow should:
- Track training-dataset versions so previous training files can be recovered. section 2.1.1
- Distinguish algorithm reproducibility from model reproducibility, which additionally requires reproducing dataset outputs and training configuration. section 2.1.1
