Our book, "Machine Learning from Weak Supervision: An Empirical Risk Minimization Approach," was published by MIT Press as part of the Adaptive Computation and Machine Learning series. It's co-authored by Sentinel Assurance co-founder Nan Lu, together with Masashi Sugiyama, Han Bao, Takashi Ishida, Tomoya Sakai, and Gang Niu.
What the book covers
- A unified empirical-risk-minimization framework for weakly supervised learning, covering settings like positive-unlabeled (PU) and positive-negative-unlabeled (PNU) classification.
- Theoretical foundations for learning when labels are missing or partial, not just noisy.
- Practical algorithms alongside the theory, written for both practitioners applying these methods and researchers building on them.
The throughline is that "not enough clean labels" isn't an edge case to work around with heuristics. It's common enough, and structured enough, to deserve the same rigor as fully supervised learning.
Our take
Most of the interesting failure modes we care about start upstream of the model, in what it was actually trained on and what it was told to optimize for. A rigorous account of learning under weak supervision is the kind of foundational work that makes it possible to reason precisely about what a model can and can't be expected to know, which is worth getting right before anything downstream depends on it.