Speaker
Description
Reproducibility is the cornerstone of scientific progress, yet in machine learning research, it remains an elusive goal. Many senior researchers lack formal training in computational science, leaving the responsibility of code development to junior researchers, often resulting in unstructured, undocumented "code dumps”. In some subfields, the situation is even worse - code is not published at all, and critical implementation details are buried in papers, creating inefficiencies, and barriers to progress.
This talk explores the current challenges in reproducible machine learning, from poor code practices to the limitations imposed by conference page limits. I will discuss what makes a "good" research code repository, the role of research software engineers in driving cultural change, and the emerging influence of coding agents in automating code generation. Finally, we’ll ask a provocative question: Should the quality check for a research code repository be whether a coding agent can reproduce the paper? In this talk, I hope to chart a path toward coding practices where reproducibility is the norm, not the exception, and where high-quality research code accelerates innovation across disciplines.