在本课程中,我们将学习机器学习、统计学和计算机科学中的现代技术,以估计黑匣子预测的不确定性。这包括共形预测、校准和多重校准、结果不可区分性,以及在没有任何分布假设的情况下产生最坏情况经验覆盖保证的最新技术。在此过程中,我们将探索这些技术在领域适应问题上的应用,以及它们在解决下游优化问题、经济和机制设计以及算法公平性方面的应用。
目录内容:
Conformal Prediction
Exchangeable Data: Full and Split Conformal Prediction
Distribution Shift and Time Series Data
Adaptive/Adversarial Conformal Prediction
Threshold Calibrated Multivalid Conformal Prediction
(Multi)Calibration
Proper Scoring Rules, Calibration, and Regret
Algorithms for offline (batch) multicalibration
Algorithms for online (adversarial) multicalibration
Moment Multicalibration
Applications of Multicalibration
Downstream Unconstrained Optimization (Omnipredictors)
Downstream Constrained Optimization
Proxies for Downstream Measurement
Distribution Shift
Tests and Predictors
Ignorantly Passing Tests: Possibility and Hardness
Outcome Indistinguishability
Other Topics
Smooth Calibration
Applications of Calibration in Mechanism Design
The Reference Class Problem
专知便捷查看
便捷下载,请关注专知公众号(点击上方蓝色专知关注)
后台回复“U134” 就可以获取《宾夕法尼亚大学最新《不确定性估计》课程笔记,134页pdf,附Slides》专知下载链接