La IA y los problemas difíciles del mundo · Informática teórica · arXiv · 2026-10-20
Training with ML-labeled Data via Stochastic First- and Zeroth-Order Optimization
Training a machine learning model typically requires ground-truth labels, but human labeling is laborious, costly, and hard to scale to large numbers of data points. This work considers a setting where a subset of samples has ground-truth labels while the rest are labeled by an ML predictor that may assign incorrect labels, and proposes first-order and zeroth-order stochastic optimization methods for training on such mixed datasets. Empirical studies ranging from training a linear regression model to fine-tuning a large language model showcase the effectiveness of the proposed algorithms.