校准Calibration

模型报出的自信程度，经常高于它的实际正确率：自信不等于对。A model's stated confidence often exceeds its actual accuracy: confident is not the same as correct.

Guo 等人 2017 年的发现：现代神经网络普遍校准不佳。

表现是输出的置信度与实际正确率不符，常表现为过度自信。

大模型一本正经地自信说错，正是这个问题的直观呈现。

论文提出了温度缩放等事后校正手段。

Guo et al. 2017: modern neural networks are poorly calibrated.

The symptom: output confidence does not match real accuracy, usually overconfidence.

A large model confidently stating something wrong is this problem in plain view.

The paper proposed post-hoc fixes such as temperature scaling.

模型说「我确定」不是统计保证。知道置信度普遍虚高，你对「自信地错」的答案会多一分警惕。'I'm sure' from a model is not a statistical guarantee. Knowing confidence is generally inflated makes you more wary of confidently wrong answers.

去导航：找现成的工具、平台、评测与学习资源Directory: ready-made tools, platforms, benchmarks and courses ↗

回到名词地图看第2站的其他 35 个词 →Back to the map: the other 35 terms in station 2 →

• Guo 等《On Calibration of Modern Neural Networks》（ICML 2017）Guo et al., On Calibration of Modern Neural Networks (ICML 2017) ↗