Collins, KM; Bhatt, U; Liu, W; Piratla, V; Sucholutsky, I; Love, B; Weller, A; (2023) Human-in-the-Loop Mixup. In: Proceedings of the Thirty-Ninth Conference on Uncertainty in Artificial Intelligence. (pp. pp. 454-464). PMLR (Proceedings of Machine Learning Research)

Preview

PDF collins23a.pdf - Published Version Download (2MB) | Preview

Abstract

Aligning model representations to humans has been found to improve robustness and generalization. However, such methods often focus on standard observational data. Synthetic data is proliferating and powering many advances in machine learning; yet, it is not always clear whether synthetic labels are perceptually aligned to humans - rendering it likely model representations are not human aligned. We focus on the synthetic data used in mixup: a powerful regularizer shown to improve model robustness, generalization, and calibration. We design a comprehensive series of elicitation interfaces, which we release as HILL MixE Suite, and recruit 159 participants to provide perceptual judgments along with their uncertainties, over mixup examples. We find that human perceptions do not consistently align with the labels traditionally used for synthetic points, and begin to demonstrate the applicability of these findings to potentially increase the reliability of downstream models, particularly when incorporating human uncertainty. We release all elicited judgments in a new data hub we call H-Mix.

Download activity - last month

Export as XMLJSONCSV

Download activity - last 12 months

Export as XMLJSONCSV

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item