Session

It Worked in My Notebook: Engineering Reproducible ML Pipelines

Machine learning pipelines often fail in a way conventional software rarely does: they continue to run but produce different results. A model retrained a month later may score differently although no one changed the code. Typical causes include an updated transitive dependency, a new accelerator driver, an undocumented preprocessing step, or an evaluation that calls a hosted model through a "latest" alias. Because nothing crashes, these discrepancies often remain unnoticed until results must be explained, audited, or reproduced.

This talk treats reproducibility as an engineering discipline. Drawing on practices from research software, where results must remain verifiable for years, it examines four sources of drift: dependencies and hardware, data and preprocessing, numerical nondeterminism, and hosted model versions. For each, it presents a concrete control: data and model versioning, deterministic training configuration, and continuous integration that validates results as well as execution. Examples use JAX, Keras, Colab, and the Gemini API, and the patterns apply to any stack. Attendees will leave with a tiered checklist for their own pipelines and clear criteria for deciding how much reproducibility a given project requires.

Keila Moral Figueroa

Research Software Engineer, FAIRmat at HU-Berlin | PhD in Particle Physics, ex-CERN

Berlin, Germany

Actions

Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.

Jump to top