Speaker
Description
Machine learning (ML) has rapidly emerged as one of the most dominant and cross-disciplinary scientific fields of our time. Driven by its immense predictive power, it has achieved widespread public integration. Historically, algorithmic decision-making tools specialized in singular tasks, performing them without significant error and thereby earning societal trust. Today, the public extrapolates this trustworthiness to modern AI and ML. However, unlike traditional algorithms, ML models are largely evaluated on their predictive accuracy on held-out data, frequently failing to report confidence intervals or quantifiable uncertainties.
This blind trust masks a profound epistemological shortcoming: current ML practices often fail to adhere to the scientific method. In the chase for high accuracy, the field frequently ignores the underlying causal and latent structures of its models. By largely abandoning hypothesis testing and uncertainty quantification, ML risks leaning toward pseudoscience, generating outputs that are highly correlated but scientifically uninterpretable.
Despite these epistemological flaws, machine learning possesses one distinct scientific advantage: high reproducibility. While many traditional sciences are currently hindered by a reproducibility crisis, ML experiments—given the same data and methods—can reliably be recreated by peers. However, simply reproducing an unscientific, black-box result does not magically render it scientific.
Furthermore, strict data-sharing is often bottlenecked by privacy concerns, proprietary restrictions, and poor data quality. In the absence of universally accessible data, the rigorous sharing of methodologies becomes paramount.
For machine learning to mature from a purely applied predictive tool into a scientifically rigorous discipline, it must fundamentally change its software practices. Replicable science in the AI era demands more than just shared weights; it requires the production and maintenance of reusable, modular, and open-source software. By building transparent codebases, Research Software Engineers empower the scientific community to peer into the black box, rigorously test hypotheses, and ensure that ML's undeniable predictive power is finally matched by genuine scientific trustworthiness.