Safe Repeated Data Use and Model Release for Exploratory Data Science

点击翻译按钮获取中文摘要

This project aims to develop new methods for repeated use of datasets and release of models trained on sensitive data. To achieve these aims, this project will develop efficient random samplers for estimating sensitivity of learning systems to data perturbation. This project expects to address the crisis of poor reproducibility and overfitting by repeated use of data sets in machine learning. Expected outcomes of this project include new methods and safety guarantees for repeated selection, training, evaluation, tuning and release of machine learners on fixed data sets. This should provide significant practical approaches for Australian industry to reuse valuable data and release privacy sensitive insights in data science pipelines.