Model calibration usually requires optimizing some parameters (e.g., temperature) w.r.t an objective function (e.g., negative log-likelihood). In this paper, we report a plain, important but often neglected fact that the objective function is influenced by calibration set difficulty, i.e., the ratio of the number of incorrectly classified samples to that of correctly classified samples. If a test set has a drastically different difficulty level from the calibration set, the optimal calibration parameters of the two datasets would be different. In other words, a calibrator optimal on the calibration set would be suboptimal on the OOD test set and thus has degraded performance. With this knowledge, we propose a simple and effective method named adaptive calibrator ensemble (ACE) to calibrate OOD datasets whose difficulty is usually higher than the calibration set. Specifically, two calibration functions are trained, one for in-distribution data (low difficulty), and the other for severely OOD data (high difficulty). To achieve desirable calibration on a new OOD dataset, ACE uses an adaptive weighting method that strikes a balance between the two extreme functions. When plugged in, ACE generally improves the performance of a few state-of-the-art calibration schemes on a series of OOD benchmarks. Importantly, such improvement does not come at the cost of the in-distribution calibration accuracy.


翻译:模型校准通常要求优化某些参数(例如温度) w.r.r.t.t 一种客观功能(例如负日志相似性)。在本文中,我们报告一个简单、重要但经常被忽视的事实,即目标功能受校准设置困难的影响,即误分类样本数量与正确分类样本数量之比。如果一个测试组的难度与校准组大不相同,两个数据集的最佳校准参数就会大不相同。换句话说,校准器在OOOD测试组上是不最佳的,因而降低了性能。有了这一知识,我们建议了一个简单而有效的方法,称为适应性校准标数,用以校准通常比校准组高的OODD数据集。具体地说,培训了两个校准功能,一个是分配数据(低难度),另一个是严重OOD数据组的优化。为了在新的OOD数据集上实现适当的校准,ACE使用一种适应性加权方法,而不是在OD-ral 精确性调整两个功能之间达到一种高度的精确性调整。当ACE-ralizal-ral a climate salbalislation supalislation sal sal supleck suplation supleck the sal sal sleck sleck sal sal sal sal sal sal sal sal sal sal slupal sal sal be sal s bal sal s bal sal sal sal sal sal sal sal sal sal sal sal sal sal salse salse sal sal sal salse sal sal sal salse salse sal sal sal sal s gs sal s gs s gs sal sal sal sal sal sal sal s bal sal sal s bal s bal s bal sal sal sal sald sal sal sal sal sal sal sal sal pral sal salse sal sal se s bal se s bals se se sal sal sal sal sal sal s balse sal s s g se sal sal </s>

0
下载
关闭预览

相关内容

专知会员服务
42+阅读 · 2020年12月18日
专知会员服务
50+阅读 · 2020年12月14日
因果图,Causal Graphs,52页ppt
专知会员服务
246+阅读 · 2020年4月19日
机器学习入门的经验与建议
专知会员服务
92+阅读 · 2019年10月10日
【哈佛大学商学院课程Fall 2019】机器学习可解释性
专知会员服务
103+阅读 · 2019年10月9日
VCIP 2022 Call for Demos
CCF多媒体专委会
1+阅读 · 2022年6月6日
Hierarchically Structured Meta-learning
CreateAMind
26+阅读 · 2019年5月22日
Transferring Knowledge across Learning Processes
CreateAMind
27+阅读 · 2019年5月18日
逆强化学习-学习人先验的动机
CreateAMind
15+阅读 · 2019年1月18日
无监督元学习表示学习
CreateAMind
27+阅读 · 2019年1月4日
Unsupervised Learning via Meta-Learning
CreateAMind
42+阅读 · 2019年1月3日
A Technical Overview of AI & ML in 2018 & Trends for 2019
待字闺中
16+阅读 · 2018年12月24日
【SIGIR2018】五篇对抗训练文章
专知
12+阅读 · 2018年7月9日
ResNet, AlexNet, VGG, Inception:各种卷积网络架构的理解
全球人工智能
19+阅读 · 2017年12月17日
【推荐】ResNet, AlexNet, VGG, Inception:各种卷积网络架构的理解
机器学习研究会
20+阅读 · 2017年12月17日
国家自然科学基金
1+阅读 · 2015年12月31日
国家自然科学基金
0+阅读 · 2015年12月31日
国家自然科学基金
1+阅读 · 2013年12月31日
国家自然科学基金
0+阅读 · 2013年12月31日
国家自然科学基金
0+阅读 · 2013年12月31日
国家自然科学基金
1+阅读 · 2012年12月31日
国家自然科学基金
0+阅读 · 2008年12月31日
Arxiv
14+阅读 · 2022年10月15日
VIP会员
相关VIP内容
专知会员服务
42+阅读 · 2020年12月18日
专知会员服务
50+阅读 · 2020年12月14日
因果图,Causal Graphs,52页ppt
专知会员服务
246+阅读 · 2020年4月19日
机器学习入门的经验与建议
专知会员服务
92+阅读 · 2019年10月10日
【哈佛大学商学院课程Fall 2019】机器学习可解释性
专知会员服务
103+阅读 · 2019年10月9日
相关资讯
VCIP 2022 Call for Demos
CCF多媒体专委会
1+阅读 · 2022年6月6日
Hierarchically Structured Meta-learning
CreateAMind
26+阅读 · 2019年5月22日
Transferring Knowledge across Learning Processes
CreateAMind
27+阅读 · 2019年5月18日
逆强化学习-学习人先验的动机
CreateAMind
15+阅读 · 2019年1月18日
无监督元学习表示学习
CreateAMind
27+阅读 · 2019年1月4日
Unsupervised Learning via Meta-Learning
CreateAMind
42+阅读 · 2019年1月3日
A Technical Overview of AI & ML in 2018 & Trends for 2019
待字闺中
16+阅读 · 2018年12月24日
【SIGIR2018】五篇对抗训练文章
专知
12+阅读 · 2018年7月9日
ResNet, AlexNet, VGG, Inception:各种卷积网络架构的理解
全球人工智能
19+阅读 · 2017年12月17日
【推荐】ResNet, AlexNet, VGG, Inception:各种卷积网络架构的理解
机器学习研究会
20+阅读 · 2017年12月17日
相关基金
国家自然科学基金
1+阅读 · 2015年12月31日
国家自然科学基金
0+阅读 · 2015年12月31日
国家自然科学基金
1+阅读 · 2013年12月31日
国家自然科学基金
0+阅读 · 2013年12月31日
国家自然科学基金
0+阅读 · 2013年12月31日
国家自然科学基金
1+阅读 · 2012年12月31日
国家自然科学基金
0+阅读 · 2008年12月31日
Top
微信扫码咨询专知VIP会员