A Fever Dream of Machine Learning Framework Composability

This seminar was hosted by Microsoft Research Africa, Nairobi together with the Microsoft AI for Good team in December 2024.

In machine learning (ML), we’ve long dreamed of – and spearheaded – seamless interoperability between frameworks, datasets, and tools to build big and capable ML systems. Much progress has been made towards driving down the transaction costs of ML, in turn making more ML tasks feasible. In this talk I will take a look at different points in the ML systems stack, from interoperable data formats over data generation all the way to the community norms around it, to touch on successes and challenges towards the composable ML systems fever dream. I’ll explore how composable infrastructure can enable new possibilities for high-impact domains such as healthcare and geospatial, where cross-pollination between data-centric methods and domain expertise is critical. Drawing from our experiences building data-centric ML communities and infrastructure, the talk will conclude with a vision for the future where practitioners get the most cost-efficient utility out of their data without a fever.

Speaker Details

Luis works on composable systems for measuring, optimizing and exchanging data states across the entire data generating process in machine learning. In regular intervals, he shares his ideas through computer code and longer texts spanning topics such as data optimization [1, 2, 3, 4, 5], ML data formats [1, 2, 3] or measurement tools for ML systems [1, 2, 3, 4, 5]. He also enjoys promoting opportunities for community. He helped initiate machine learning venues such as Data-Centric Machine Learning Research (DMLR) and AI for Good and co-chaired conferences such as ML4H, ICLR or the DMLR workshop series. He is Head of Machine Learning at Swiss company Dotphoton.