A research method called CuBAS wants to fix how machine learning models choose their training data.
Most sampling strategies treat a dataset as a flat list and pull examples at random or by uncertainty scores. CuBAS takes a different approach: it models the labeled dataset as a statistical manifold — a geometric surface — and estimates local curvature at each data point using the ratio of second to first-order observed Fisher information, derived from a Potts Markov random field. It then builds a k-nearest-neighbor graph and assigns every node a curvature score. Low-curvature zones are smooth, redundant clusters; high-curvature zones hug decision boundaries, where examples do the most classification work. CuBAS samples from both.
The payoff, if the benchmarks hold up, is real: the researchers tested CuBAS across more than 60 datasets and reported consistent, statistically significant gains over random sampling and uncertainty-based baselines across a range of budgets and classifier types. That breadth matters because most sampling papers cherry-pick favorable evaluations. The method also scales linearly with the number of k-NN graph edges, so it does not blow up on large datasets.
Data-centric AI has been a talking point since Andrew Ng pushed it hard a few years ago, but most production pipelines still lean on brute-force data collection. CuBAS will not change that overnight — it is an arXiv preprint, not a shipping library — but it is a tidy demonstration that geometry-aware selection can squeeze more signal out of fewer labels.