Conformal Compressors

Standard machine learning predictors become more effective as larger input datasets are made available. While this is desirable for enhancing predictive power, it often implies substantial computational costs. One feasible approach to mitigate this issue is to replace large datasets with smaller, carefully crafted representations that retain the essential properties of the original data. In this paper, we revisit the exploration of this approach through the interaction of coresets - small, provably correct summaries of data - and Conformal Prediction, a robust and general method for calibrating machine learning predictions. Specifically, we build on existing work to introduce Conformal Compressors, a method inspired by coresets that leverages Conformal Prediction for data compression. Initial results indicate that these compressors effectively capture meaningful information from the data while demonstrating significantly better stability and reliability compared to Uniform Random Sampling and state-of-the-art coreset constructions.