As datasets grow, machine learning can become too computationally expensive to run efficiently. A dataset may even exceed the memory available on a single computer. Our research explores coreset-based conformal prediction: using a compact, weighted summary of a dataset to support learning at a larger scale.
Most of this work was carried out by my student Nery Riquelme-Granada.
A coreset is a small, weighted subset of a dataset that preserves properties relevant to a particular learning problem. It provides an approximation to the full dataset with theoretical guarantees on the approximation error, while ideally requiring much less storage and computation.
One approach to constructing a coreset is importance sampling. Instead of giving every data point an equal chance of selection, we assign sampling probabilities based on sensitivity scores that capture each point's potential influence on the learning objective. We then sample M points and weight them inversely to their sampling probabilities. The resulting weighted sample forms the coreset.