Our paper, "Corruptions of Supervised Learning Problems: Typology and Mitigations," has been published in the Journal of Machine Learning Research (JMLR), one of the most selective venues in machine learning. The work is co-authored by Sentinel Assurance co-founder Nan Lu, together with Laura Iacovissi and Robert C. Williamson (University of Tübingen).
What the paper does
Almost all supervised learning theory assumes the training data is clean. In practice it rarely is: labels get mislabeled, features get corrupted or dropped, and the two can be corrupted together in ways that depend on each other. Prior work covered pieces of this picture, but no unified account of what "corruption" even means existed.
- A single, general theory of corruption, modeled as a Markov kernel acting on the data distribution, that covers label corruption, attribute corruption, and dependent combinations of the two.
- A provably exhaustive classification of corruption types, unifying models that were previously studied in isolation.
- A systematic analysis of how each corruption type distorts a learning problem's Bayes risk, the theoretical floor on how well any model can do.
- Mitigation methods that extend classical loss-correction techniques, built for simple label noise, to the harder dependent and joint corruption cases.
The label-only assumption is the one worth sitting with. Most correction techniques in the literature only patch the loss function, which works when corruption hits labels alone. Once corruption touches the input features too, the fix has to reach into the hypothesis class itself, not just the loss.
Our take
Every guarantee a machine learning system offers rests on an assumption about the data it was built on. That assumption quietly breaks more often than most teams check for, and it breaks in ways that a clean validation set won't surface. That's exactly the kind of gap we think matters for anyone deciding how much to trust a model, and it's the sort of foundational question we care about answering rigorously before the ML systems built on top get more autonomy over real decisions.