Quantification, optimisation, and application of deep uncertainty

该项目旨在开发一个深度不确定性量化的框架。目前，深度学习研究与量化和管理不确定性所需的方法之间存在根本性的差距。该研究将提出一种新的无分布方法来生成深度预测不确定性估计，以避免现有方法的假设。估计数的质量将通过应用基于间隔的对抗性训练步骤来提高。该项目预计将帮助数据驱动的澳大利亚组织和行业更好地量化和管理预测的不确定性。该项目将通过更好的决策和更有力的规划为他们节省大量成本。

This project aims to develop a framework for deep uncertainty quantification. There is currently a fundamental gap between deep learning research and the methods required to quantify and manage uncertainties. The research will propose a novel distribution-free methodology to generate deep predictive uncertainty estimates to avoid the assumptions of existing methods. The quality of estimates will be enhanced by applying an interval-based adversarial training step. The project is expected to help data-driven Australian organisations and industries to better quantify and manage forecasting uncertainties. This project will provide them with significant cost savings through better decision making and more robust planning.