Timezone: »

 
Poster
Calibrated Reliable Regression using Maximum Mean Discrepancy
Peng Cui · Wenbo Hu · Jun Zhu

Wed Dec 09 09:00 PM -- 11:00 PM (PST) @ Poster Session 4 #1259
in Poster Session 5 »

Accurate quantification of uncertainty is crucial for real-world applications of machine learning. However, modern deep neural networks still produce unreliable predictive uncertainty, often yielding over-confident predictions. In this paper, we are concerned with getting well-calibrated predictions in regression tasks. We propose the calibrated regression method using the maximum mean discrepancy by minimizing the kernel embedding measure. Theoretically, the calibration error of our method asymptotically converges to zero when the sample size is large enough. Experiments on non-trivial real datasets show that our method can produce well-calibrated and sharp prediction intervals, which outperforms the related state-of-the-art methods.

Author Information

Peng Cui (Tsinghua University)
Wenbo Hu (Tsinghua University / RealAI)

Homepage: https://wbhu.net/ Ph.D of AI from Tsinghua University. Currently Postdoc Researcher at Tsail Group of Tsinghua University. Also affiliated with RealAI.ai, an AI startup originated from Tsinghua Univ. Managing a R&D team of both machine leaning and data analysis fields. My research is machine learning methods, with extensions to predictive maintenance, relationship and AI interpretability. Published several top-lier conference/journal AI papers. Focusing on next-generation AI applications of industries and finances. [Highlight] We are developing RealSeries.github.io, a comprehensive out-of-the-box Python toolkit for various tasks, including Anomaly Detection, Granger causality and Forecast and Imputation, of dealing with Time Series Datasets. RealSeries has the following features: 1）Unified APIs, detailed documentation, easy-to-follow examples and straightforward visualizations. 2）All-levels of models, including simple thresholds, classification-based models, and deep (Bayesian) models. Thanks！ 个人主页: https://wbhu.net/ 清华大学的AI领域博士毕业。目前是清华大学AI研究院的博士后，同时在一家清华系的AI创业公司任职，负责管理AI算法研发和行业应用分析团队。 我的研究主要是机器学习方法，我主要关注机器学习在预测性维护、不确定性学习、可解释性学习的扩展。我在顶级的AI会议和期刊上发表过论文。目前还对AI在互联网、工业制造和金融领域的落地转化感兴趣。 【Highlight】我们研发了用于时间序列处理的算法库RealSeries.github.io。这是一个容易理解、开箱即用的算法库，包含了几乎设计到时间序列分析的工具，包括异常检测、序列预测和因果推理。RealSeries有统一的接口设计、详细的文档、易懂的样例，并且涵盖了几乎所有业界常用的算法。 谢谢！

Jun Zhu (Tsinghua University)

More from the Same Authors