{
  "checked_on": "2026-10-20",
  "scope": "目标导向的小范围检索；不是系统综述，不据此宣称新颖性",
  "queries": [
    "\"temperature scaling\" \"small\" \"calibration set\" calibration",
    "\"temperature scaling\" \"regularization\" calibration sample size",
    "site.proceedings.mlr.press \"Verified Uncertainty Calibration\""
  ],
  "papers": [
    {
      "title": "On Calibration of Modern Neural Networks",
      "authors": "Chuan Guo; Geoff Pleiss; Yu Sun; Kilian Q. Weinberger",
      "year": 2017,
      "version": "ICML / PMLR 70:1321–1330，会议版正文与补充材料",
      "url": "https://proceedings.mlr.press/v70/guo17a.html",
      "read": "第 44 章精读：温度公式、正文表 1、补充表 S2，核对固定提交的作者代码",
      "role": "本项目复现的方法来源",
      "difference": "本项目改用 digits 和小型 MLP，不能称为原论文基准复制"
    },
    {
      "title": "Evaluating model calibration in classification",
      "authors": "Juozas Vaicenavicius; David Widmann; Carl Andersson; Fredrik Lindsten; Jacob Roll; Thomas Schön",
      "year": 2019,
      "version": "AISTATS / PMLR 89:3459–3467，会议版",
      "url": "https://proceedings.mlr.press/v89/vaicenavicius19a.html",
      "read": "引言、校准对象、4.3 节分箱估计及其解释",
      "role": "帮助理解 ECE 与有限样本评价的限制",
      "difference": "本项目未实现该文提出的完整校准检验"
    },
    {
      "title": "Confidence Calibration of Classifiers with Many Classes",
      "authors": "Adrien Le Coz; Stéphane Herbin; Faouzi Adjed",
      "year": 2024,
      "version": "NeurIPS 37，会议版",
      "url": "https://proceedings.neurips.cc/paper_files/paper/2024/hash/8df80d7115a55eb81010c967a247b1ae-Abstract-Conference.html",
      "read": "摘要、引言与 5.4 节校准样本量实验",
      "role": "说明校准样本量与表达能力已有相关研究，扩展问题不是凭空出现",
      "difference": "该文关注大量类别、TvA 和多种校准方法；本项目仅研究十类 MLP 的共同温度，没有运行 TvA"
    }
  ]
}