{
  "meta": {
    "id": "CALYR-LIT-001",
    "title": "Calibration literature and data graph",
    "route": "/research/methods/scientific-ai-numerics/literature/",
    "status": "Living evidence index",
    "scope": "Neural-network calibration, uncertainty, evaluation data, and method lineage extracted from Guo et al. (2017).",
    "seed_paper": "guo2017-calibration",
    "source_of_truth": "content/literature.yaml",
    "generated_index": "/generated/literature.index.json",
    "extraction_contract": [
      "Preserve the distinction between a paper's claim and calyr.aí's interpretation.",
      "Link every extracted idea to a source and a location in that source.",
      "Represent datasets separately from the papers that introduce or use them.",
      "Mark inferred, incomplete, or unverified metadata explicitly.",
      "Treat dataset availability and licensing as fields to verify, never assumptions."
    ]
  },
  "counts": {
    "papers": 25,
    "datasets": 9,
    "dataset_source_citations": 6,
    "ideas": 8,
    "relations": 12
  },
  "papers": [
    {
      "id": "guo2017-calibration",
      "title": "On Calibration of Modern Neural Networks",
      "authors": [
        "Chuan Guo",
        "Geoff Pleiss",
        "Yu Sun",
        "Kilian Q. Weinberger"
      ],
      "year": 2017,
      "venue": "ICML / PMLR 70",
      "url": "https://proceedings.mlr.press/v70/guo17a.html",
      "pdf": "https://proceedings.mlr.press/v70/guo17a/guo17a.pdf",
      "role": "seed",
      "topics": [
        "calibration",
        "temperature scaling",
        "ECE",
        "reliability diagrams"
      ],
      "extractable_idea": "Modern neural networks can become more accurate yet less calibrated; a validation-set scalar temperature often repairs confidence without changing class predictions.",
      "verification": "primary-source-verified"
    },
    {
      "id": "degroot1983-forecasters",
      "title": "The Comparison and Evaluation of Forecasters",
      "authors": [
        "Morris H. DeGroot",
        "Stephen E. Fienberg"
      ],
      "year": 1983,
      "role": "metric-lineage",
      "topics": [
        "reliability",
        "calibration",
        "forecasting"
      ],
      "extractable_idea": "Separate calibration from refinement when evaluating probabilistic forecasts; reliability diagrams operationalize this distinction.",
      "verification": "bibliography-verified"
    },
    {
      "id": "niculescu2005-probabilities",
      "title": "Predicting Good Probabilities with Supervised Learning",
      "authors": [
        "Alexandru Niculescu-Mizil",
        "Rich Caruana"
      ],
      "year": 2005,
      "venue": "ICML",
      "url": "https://www.cs.cornell.edu/~caruana/niculescu.icml05.pdf",
      "role": "calibration-baseline",
      "topics": [
        "probability calibration",
        "Platt scaling",
        "isotonic regression"
      ],
      "extractable_idea": "Compare classifiers by the quality of their probability estimates and calibrate distorted scores with held-out data.",
      "verification": "primary-link-present"
    },
    {
      "id": "naeini2015-bbq",
      "title": "Obtaining Well Calibrated Probabilities Using Bayesian Binning",
      "authors": [
        "Mahdi Pakdaman Naeini",
        "Gregory F. Cooper",
        "Milos Hauskrecht"
      ],
      "year": 2015,
      "venue": "AAAI",
      "url": "https://ojs.aaai.org/index.php/AAAI/article/view/9602",
      "role": "calibration-method",
      "topics": [
        "Bayesian binning",
        "ECE",
        "MCE"
      ],
      "extractable_idea": "Average over plausible binning schemes instead of committing to one arbitrary discretization; ECE and MCE summarize calibration gaps.",
      "verification": "primary-link-present"
    },
    {
      "id": "platt1999-probabilistic",
      "title": "Probabilistic Outputs for Support Vector Machines and Comparisons to Regularized Likelihood Methods",
      "authors": [
        "John Platt"
      ],
      "year": 1999,
      "role": "calibration-method",
      "topics": [
        "Platt scaling",
        "sigmoid calibration"
      ],
      "extractable_idea": "Fit a low-dimensional post-hoc mapping from decision scores to probabilities on held-out data.",
      "verification": "bibliography-verified"
    },
    {
      "id": "zadrozny2001-calibrated",
      "title": "Obtaining Calibrated Probability Estimates from Decision Trees and Naive Bayesian Classifiers",
      "authors": [
        "Bianca Zadrozny",
        "Charles Elkan"
      ],
      "year": 2001,
      "venue": "ICML",
      "role": "calibration-method",
      "topics": [
        "histogram binning",
        "isotonic regression"
      ],
      "extractable_idea": "Correct systematic probability distortion through a learned monotone or binned mapping.",
      "verification": "bibliography-verified"
    },
    {
      "id": "zadrozny2002-multiclass",
      "title": "Transforming Classifier Scores into Accurate Multiclass Probability Estimates",
      "authors": [
        "Bianca Zadrozny",
        "Charles Elkan"
      ],
      "year": 2002,
      "venue": "KDD",
      "role": "multiclass-extension",
      "topics": [
        "multiclass calibration",
        "one-vs-all"
      ],
      "extractable_idea": "Extend binary calibration to multiclass outputs through classwise calibration and probability normalization.",
      "verification": "bibliography-verified"
    },
    {
      "id": "denker1990-output-probabilities",
      "title": "Transforming Neural-Net Output Levels to Probability Distributions",
      "authors": [
        "John S. Denker",
        "Yann LeCun"
      ],
      "year": 1990,
      "venue": "NeurIPS",
      "role": "neural-calibration-lineage",
      "topics": [
        "neural networks",
        "probability estimation"
      ],
      "extractable_idea": "Neural outputs require an explicit transformation and validation before they can be interpreted as probabilities.",
      "verification": "bibliography-verified"
    },
    {
      "id": "mackay1992-bayesian",
      "title": "A Practical Bayesian Framework for Backpropagation Networks",
      "authors": [
        "David J. C. MacKay"
      ],
      "year": 1992,
      "venue": "Neural Computation",
      "role": "uncertainty-lineage",
      "topics": [
        "Bayesian neural networks",
        "predictive uncertainty"
      ],
      "extractable_idea": "Integrate parameter uncertainty rather than treating one fitted network as the complete predictive distribution.",
      "verification": "bibliography-verified"
    },
    {
      "id": "gal2016-dropout",
      "title": "Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning",
      "authors": [
        "Yarin Gal",
        "Zoubin Ghahramani"
      ],
      "year": 2016,
      "venue": "ICML",
      "url": "https://proceedings.mlr.press/v48/gal16.html",
      "role": "uncertainty-method",
      "topics": [
        "MC dropout",
        "epistemic uncertainty"
      ],
      "extractable_idea": "Repeated stochastic dropout passes approximate Bayesian model averaging and expose predictive uncertainty.",
      "verification": "primary-link-present"
    },
    {
      "id": "lakshminarayanan2017-ensembles",
      "title": "Simple and Scalable Predictive Uncertainty Estimation Using Deep Ensembles",
      "authors": [
        "Balaji Lakshminarayanan",
        "Alexander Pritzel",
        "Charles Blundell"
      ],
      "year": 2017,
      "venue": "NeurIPS",
      "url": "https://arxiv.org/abs/1612.01474",
      "role": "uncertainty-method",
      "topics": [
        "deep ensembles",
        "predictive uncertainty",
        "calibration"
      ],
      "extractable_idea": "Independently trained probabilistic networks provide a strong, scalable uncertainty and calibration baseline.",
      "verification": "primary-link-present"
    },
    {
      "id": "pereyra2017-confidence-penalty",
      "title": "Regularizing Neural Networks by Penalizing Confident Output Distributions",
      "authors": [
        "Gabriel Pereyra",
        "George Tucker",
        "Jan Chorowski",
        "Łukasz Kaiser",
        "Geoffrey Hinton"
      ],
      "year": 2017,
      "url": "https://arxiv.org/abs/1701.06548",
      "role": "training-method",
      "topics": [
        "confidence penalty",
        "entropy regularization"
      ],
      "extractable_idea": "Penalize unjustified certainty during training instead of correcting confidence only after training.",
      "verification": "primary-link-present"
    },
    {
      "id": "hendrycks2017-ood",
      "title": "A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks",
      "authors": [
        "Dan Hendrycks",
        "Kevin Gimpel"
      ],
      "year": 2017,
      "venue": "ICLR",
      "url": "https://arxiv.org/abs/1610.02136",
      "role": "applicability-domain",
      "topics": [
        "out-of-distribution detection",
        "confidence"
      ],
      "extractable_idea": "Confidence signals can help separate ordinary predictions from likely errors and distribution shift, but require explicit evaluation.",
      "verification": "primary-link-present"
    },
    {
      "id": "kuleshov2016-online",
      "title": "Reliable Confidence Estimation via Online Learning",
      "authors": [
        "Volodymyr Kuleshov",
        "Stefano Ermon"
      ],
      "year": 2016,
      "url": "https://arxiv.org/abs/1607.03594",
      "role": "calibration-setting",
      "topics": [
        "online calibration",
        "adversarial sequences"
      ],
      "extractable_idea": "Calibration can be formulated as an online guarantee when observations arrive sequentially or adversarially.",
      "verification": "primary-link-present"
    },
    {
      "id": "kuleshov2015-structured",
      "title": "Calibrated Structured Prediction",
      "authors": [
        "Volodymyr Kuleshov",
        "Percy Liang"
      ],
      "year": 2015,
      "venue": "NeurIPS",
      "role": "structured-output",
      "topics": [
        "structured prediction",
        "calibration"
      ],
      "extractable_idea": "Calibration must be defined against meaningful events when the output is a structure rather than a single class.",
      "verification": "bibliography-verified"
    },
    {
      "id": "kendall2017-uncertainties",
      "title": "What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision?",
      "authors": [
        "Alex Kendall",
        "Yarin Gal"
      ],
      "year": 2017,
      "url": "https://arxiv.org/abs/1703.04977",
      "role": "uncertainty-taxonomy",
      "topics": [
        "aleatoric uncertainty",
        "epistemic uncertainty",
        "computer vision"
      ],
      "extractable_idea": "Keep data uncertainty and model uncertainty distinct because they imply different interventions and failure responses.",
      "verification": "primary-link-present"
    },
    {
      "id": "friedman2001-elements",
      "title": "The Elements of Statistical Learning",
      "authors": [
        "Jerome Friedman",
        "Trevor Hastie",
        "Robert Tibshirani"
      ],
      "year": 2001,
      "role": "loss-lineage",
      "topics": [
        "negative log likelihood",
        "statistical learning"
      ],
      "extractable_idea": "Proper probabilistic losses evaluate the full predictive distribution, not just the winning class.",
      "verification": "bibliography-verified"
    },
    {
      "id": "jaynes1957-information",
      "title": "Information Theory and Statistical Mechanics",
      "authors": [
        "Edwin T. Jaynes"
      ],
      "year": 1957,
      "venue": "Physical Review",
      "role": "temperature-lineage",
      "topics": [
        "entropy",
        "temperature",
        "constrained distributions"
      ],
      "extractable_idea": "Temperature controls concentration in exponential-family distributions and connects calibration to constrained maximum entropy.",
      "verification": "bibliography-verified"
    },
    {
      "id": "hinton2015-distilling",
      "title": "Distilling the Knowledge in a Neural Network",
      "authors": [
        "Geoffrey Hinton",
        "Oriol Vinyals",
        "Jeff Dean"
      ],
      "year": 2015,
      "url": "https://arxiv.org/abs/1503.02531",
      "role": "temperature-lineage",
      "topics": [
        "knowledge distillation",
        "soft targets",
        "temperature"
      ],
      "extractable_idea": "Temperature exposes relative class evidence through softened probability targets; calibration uses the same control for a different objective.",
      "verification": "primary-link-present"
    },
    {
      "id": "he2016-resnet",
      "title": "Deep Residual Learning for Image Recognition",
      "authors": [
        "Kaiming He",
        "Xiangyu Zhang",
        "Shaoqing Ren",
        "Jian Sun"
      ],
      "year": 2016,
      "venue": "CVPR",
      "url": "https://arxiv.org/abs/1512.03385",
      "role": "evaluated-architecture",
      "topics": [
        "ResNet",
        "depth"
      ],
      "extractable_idea": "Residual connections enable much deeper classifiers; Guo et al. use this capacity trend to study calibration degradation.",
      "verification": "primary-link-present"
    },
    {
      "id": "huang2016-stochastic-depth",
      "title": "Deep Networks with Stochastic Depth",
      "authors": [
        "Gao Huang",
        "Yu Sun",
        "Zhuang Liu",
        "Daniel Sedra",
        "Kilian Q. Weinberger"
      ],
      "year": 2016,
      "venue": "ECCV",
      "url": "https://arxiv.org/abs/1603.09382",
      "role": "evaluated-architecture",
      "topics": [
        "stochastic depth",
        "ResNet"
      ],
      "extractable_idea": "Randomly shortening networks during training regularizes very deep residual models; calibration still needs separate measurement.",
      "verification": "primary-link-present"
    },
    {
      "id": "zagoruyko2016-wide-resnet",
      "title": "Wide Residual Networks",
      "authors": [
        "Sergey Zagoruyko",
        "Nikos Komodakis"
      ],
      "year": 2016,
      "venue": "BMVC",
      "url": "https://arxiv.org/abs/1605.07146",
      "role": "evaluated-architecture",
      "topics": [
        "model width",
        "ResNet"
      ],
      "extractable_idea": "Width is an independent capacity axis whose accuracy benefit does not guarantee calibrated confidence.",
      "verification": "primary-link-present"
    },
    {
      "id": "huang2017-densenet",
      "title": "Densely Connected Convolutional Networks",
      "authors": [
        "Gao Huang",
        "Zhuang Liu",
        "Laurens van der Maaten",
        "Kilian Q. Weinberger"
      ],
      "year": 2017,
      "venue": "CVPR",
      "url": "https://arxiv.org/abs/1608.06993",
      "role": "evaluated-architecture",
      "topics": [
        "DenseNet",
        "connectivity"
      ],
      "extractable_idea": "Dense feature reuse changes optimization and capacity, while calibration remains a separate empirical property.",
      "verification": "primary-link-present"
    },
    {
      "id": "iyyer2015-dan",
      "title": "Deep Unordered Composition Rivals Syntactic Methods for Text Classification",
      "authors": [
        "Mohit Iyyer",
        "Varun Manjunatha",
        "Jordan Boyd-Graber",
        "Hal Daumé III"
      ],
      "year": 2015,
      "venue": "ACL",
      "url": "https://aclanthology.org/P15-1162/",
      "role": "evaluated-architecture",
      "topics": [
        "Deep Averaging Networks",
        "document classification"
      ],
      "extractable_idea": "Simple averaged text representations provide a useful architecture contrast for calibration experiments.",
      "verification": "primary-link-present"
    },
    {
      "id": "tai2015-treelstm",
      "title": "Improved Semantic Representations From Tree-Structured Long Short-Term Memory Networks",
      "authors": [
        "Kai Sheng Tai",
        "Richard Socher",
        "Christopher D. Manning"
      ],
      "year": 2015,
      "venue": "ACL",
      "url": "https://aclanthology.org/P15-1150/",
      "role": "evaluated-architecture",
      "topics": [
        "TreeLSTM",
        "sentiment classification"
      ],
      "extractable_idea": "Structured recurrent models make calibration testable on hierarchical NLP outputs, not only images.",
      "verification": "primary-link-present"
    }
  ],
  "datasets": [
    {
      "id": "cub-200-2011",
      "name": "Caltech-UCSD Birds 200-2011",
      "domain": "image classification",
      "source_paper": "welinder2010-cub",
      "cited_by": "guo2017-calibration",
      "url": "https://www.vision.caltech.edu/datasets/cub_200_2011/",
      "classes": 200,
      "split_reported_by_guo": {
        "train": 5994,
        "validation": 2897,
        "test": 2897
      },
      "ideas": [
        "fine-grained classification",
        "high-class-count calibration"
      ],
      "availability": "public-download",
      "license_status": "verify-before-reuse",
      "provenance_note": "Guo et al. cite Welinder et al. (2010); the current official CUB-200-2011 page requests citation of Wah et al. (2011) and restricts images to non-commercial research and education."
    },
    {
      "id": "stanford-cars",
      "name": "Stanford Cars",
      "domain": "image classification",
      "source_paper": "krause2013-cars",
      "cited_by": "guo2017-calibration",
      "url": "https://ai.stanford.edu/~jkrause/",
      "classes": 196,
      "split_reported_by_guo": {
        "train": 8041,
        "validation": 4020,
        "test": 4020
      },
      "ideas": [
        "fine-grained classification",
        "matrix-scaling stress test"
      ],
      "availability": "original-download-unavailable",
      "license_status": "verify-before-reuse",
      "provenance_note": "The original Stanford dataset URL cited by downstream libraries currently returns 404; use the author archive to verify provenance before selecting a maintained mirror."
    },
    {
      "id": "imagenet-2012",
      "name": "ImageNet Large Scale Visual Recognition Challenge 2012",
      "domain": "image classification",
      "source_paper": "deng2009-imagenet",
      "cited_by": "guo2017-calibration",
      "url": "https://www.image-net.org/challenges/LSVRC/2012/",
      "classes": 1000,
      "split_reported_by_guo": {
        "train": 1300000,
        "validation": 25000,
        "test": 25000
      },
      "ideas": [
        "large-scale calibration",
        "transfer learning",
        "parameter-scaling limit"
      ],
      "availability": "registration-required",
      "license_status": "restricted-terms-check-required"
    },
    {
      "id": "cifar-10",
      "name": "CIFAR-10",
      "domain": "image classification",
      "source_paper": "krizhevsky2009-cifar",
      "cited_by": "guo2017-calibration",
      "url": "https://www.cs.toronto.edu/~kriz/cifar.html",
      "classes": 10,
      "split_reported_by_guo": {
        "train": 45000,
        "validation": 5000,
        "test": 10000
      },
      "ideas": [
        "controlled architecture comparison",
        "calibration benchmark"
      ],
      "availability": "public-download",
      "license_status": "verify-before-reuse"
    },
    {
      "id": "cifar-100",
      "name": "CIFAR-100",
      "domain": "image classification",
      "source_paper": "krizhevsky2009-cifar",
      "cited_by": "guo2017-calibration",
      "url": "https://www.cs.toronto.edu/~kriz/cifar.html",
      "classes": 100,
      "split_reported_by_guo": {
        "train": 45000,
        "validation": 5000,
        "test": 10000
      },
      "ideas": [
        "reliability diagrams",
        "capacity-vs-calibration",
        "temperature scaling"
      ],
      "availability": "public-download",
      "license_status": "verify-before-reuse"
    },
    {
      "id": "svhn",
      "name": "Street View House Numbers",
      "domain": "image classification",
      "source_paper": "netzer2011-svhn",
      "cited_by": "guo2017-calibration",
      "url": "http://ufldl.stanford.edu/housenumbers/",
      "classes": 10,
      "split_reported_by_guo": {
        "train": 604388,
        "validation": 6000,
        "test": 26032
      },
      "ideas": [
        "low-error calibration regime",
        "cropped digit recognition"
      ],
      "availability": "public-download",
      "license_status": "verify-before-reuse"
    },
    {
      "id": "twenty-newsgroups",
      "name": "20 Newsgroups",
      "domain": "document classification",
      "source_paper": "uncited-dataset-origin",
      "cited_by": "guo2017-calibration",
      "url": "http://qwone.com/~jason/20Newsgroups/",
      "classes": 20,
      "split_reported_by_guo": {
        "train": 9034,
        "validation": 2259,
        "test": 7528
      },
      "ideas": [
        "document calibration",
        "Deep Averaging Network evaluation"
      ],
      "availability": "public-download",
      "license_status": "verify-before-reuse"
    },
    {
      "id": "reuters",
      "name": "Reuters document classification corpus",
      "domain": "document classification",
      "source_paper": "uncited-dataset-origin",
      "cited_by": "guo2017-calibration",
      "url": "https://kdd.ics.uci.edu/databases/reuters21578/reuters21578.html",
      "classes": 8,
      "split_reported_by_guo": {
        "train": 4388,
        "validation": 1097,
        "test": 2189
      },
      "ideas": [
        "already-calibrated baseline",
        "post-processing may be unnecessary"
      ],
      "availability": "public-download",
      "license_status": "terms-check-required"
    },
    {
      "id": "sst",
      "name": "Stanford Sentiment Treebank",
      "domain": "sentiment classification",
      "source_paper": "socher2013-sst",
      "cited_by": "guo2017-calibration",
      "url": "https://nlp.stanford.edu/sentiment/",
      "classes": [
        2,
        5
      ],
      "split_reported_by_guo": {
        "binary": {
          "train": 6920,
          "validation": 872,
          "test": 1821
        },
        "fine_grained": {
          "train": 544,
          "validation": 1101,
          "test": 2210
        }
      },
      "ideas": [
        "structured sentiment",
        "binary-vs-multiclass calibration"
      ],
      "availability": "public-download",
      "license_status": "verify-before-reuse"
    }
  ],
  "dataset_source_citations": [
    {
      "id": "welinder2010-cub",
      "title": "Caltech-UCSD Birds 200",
      "year": 2010,
      "authors": [
        "P. Welinder",
        "S. Branson",
        "T. Mita",
        "C. Wah",
        "F. Schroff",
        "S. Belongie",
        "P. Perona"
      ]
    },
    {
      "id": "krause2013-cars",
      "title": "3D Object Representations for Fine-Grained Categorization",
      "year": 2013,
      "authors": [
        "Jonathan Krause",
        "Michael Stark",
        "Jia Deng",
        "Li Fei-Fei"
      ]
    },
    {
      "id": "deng2009-imagenet",
      "title": "ImageNet: A Large-Scale Hierarchical Image Database",
      "year": 2009,
      "authors": [
        "Jia Deng",
        "Wei Dong",
        "Richard Socher",
        "Li-Jia Li",
        "Kai Li",
        "Li Fei-Fei"
      ]
    },
    {
      "id": "krizhevsky2009-cifar",
      "title": "Learning Multiple Layers of Features from Tiny Images",
      "year": 2009,
      "authors": [
        "Alex Krizhevsky",
        "Geoffrey Hinton"
      ]
    },
    {
      "id": "netzer2011-svhn",
      "title": "Reading Digits in Natural Images with Unsupervised Feature Learning",
      "year": 2011,
      "authors": [
        "Yuval Netzer",
        "Tao Wang",
        "Adam Coates",
        "Alessandro Bissacco",
        "Bo Wu",
        "Andrew Y. Ng"
      ]
    },
    {
      "id": "socher2013-sst",
      "title": "Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank",
      "year": 2013,
      "authors": [
        "Richard Socher",
        "Alex Perelygin",
        "Jean Wu",
        "Jason Chuang",
        "Christopher D. Manning",
        "Andrew Ng",
        "Christopher Potts"
      ]
    }
  ],
  "ideas": [
    {
      "id": "idea-calibration-is-separate",
      "title": "Accuracy and calibration are separate model properties",
      "source": "guo2017-calibration",
      "location": "Sections 1 and 3; Figures 1-3",
      "kind": "finding",
      "statement": "Lower classification error can coexist with worse probabilistic calibration.",
      "operationalization": "Evaluate a probabilistic predictor with discrimination metrics and calibration metrics; never substitute one for the other.",
      "tags": [
        "accuracy",
        "calibration",
        "validation"
      ]
    },
    {
      "id": "idea-ece-is-an-approximation",
      "title": "ECE is a finite-bin approximation, not ground truth",
      "source": "guo2017-calibration",
      "location": "Section 2, Equations 2-3",
      "kind": "measurement",
      "statement": "ECE compresses confidence-accuracy gaps using a chosen binning scheme and can change with that scheme.",
      "operationalization": "Store binning policy, sample counts, uncertainty, and at least one complementary calibration diagnostic.",
      "tags": [
        "ECE",
        "binning",
        "measurement uncertainty"
      ]
    },
    {
      "id": "idea-temperature-scaling",
      "title": "Temperature scaling is a one-parameter calibration layer",
      "source": "guo2017-calibration",
      "location": "Section 4.2, Equation 9",
      "kind": "method",
      "statement": "Optimize one positive scalar on validation NLL and divide logits by it before softmax.",
      "operationalization": "Fit only on held-out validation data; freeze the predictive model; report pre/post NLL, ECE, reliability, and unchanged accuracy.",
      "tags": [
        "temperature scaling",
        "logits",
        "post-hoc calibration"
      ]
    },
    {
      "id": "idea-low-dimensional-miscalibration",
      "title": "Miscalibration can be intrinsically low dimensional",
      "source": "guo2017-calibration",
      "location": "Section 5",
      "kind": "hypothesis",
      "statement": "A scalar correction can outperform more expressive vector or matrix transformations on the tested networks.",
      "operationalization": "Start with the lowest-capacity calibrator and increase complexity only under held-out evidence.",
      "tags": [
        "model selection",
        "regularization",
        "simplicity"
      ]
    },
    {
      "id": "idea-capacity-regularization",
      "title": "Capacity and regularization affect confidence quality",
      "source": "guo2017-calibration",
      "location": "Section 3, Figure 2",
      "kind": "finding",
      "statement": "Depth, width, Batch Normalization, and weight decay correlate with calibration changes in the reported experiments.",
      "operationalization": "Treat calibration as an architecture- and training-dependent validation target, while preserving the paper's warning that the observations are not causal proof.",
      "tags": [
        "capacity",
        "Batch Normalization",
        "weight decay",
        "causality"
      ]
    },
    {
      "id": "idea-applicability-gate",
      "title": "Calibration does not establish applicability",
      "source": "hendrycks2017-ood",
      "location": "method lineage cited by Guo et al., Section 4.3",
      "kind": "boundary",
      "statement": "In-distribution calibration and out-of-distribution detection answer different questions.",
      "operationalization": "Gate predictions with a separate applicability-domain or shift detector before interpreting calibrated probabilities.",
      "tags": [
        "OOD",
        "applicability domain",
        "safety"
      ]
    },
    {
      "id": "idea-uncertainty-components",
      "title": "Separate aleatoric and epistemic uncertainty",
      "source": "kendall2017-uncertainties",
      "location": "related-work lineage cited by Guo et al., Section 4.3",
      "kind": "taxonomy",
      "statement": "Data noise and model uncertainty require different measurements and next actions.",
      "operationalization": "Record uncertainty component, estimator, calibration evidence, and the action triggered by high uncertainty.",
      "tags": [
        "aleatoric",
        "epistemic",
        "decision"
      ]
    },
    {
      "id": "idea-dataset-specific-calibration",
      "title": "Calibration conclusions are dataset-conditional",
      "source": "guo2017-calibration",
      "location": "Section 5, Table 1",
      "kind": "limitation",
      "statement": "Temperature scaling was broadly effective in the study but Reuters was already well calibrated and showed little benefit.",
      "operationalization": "Refit and revalidate calibration for each model, dataset, split, and deployment regime; never transfer a temperature blindly.",
      "tags": [
        "dataset shift",
        "validation",
        "Reuters"
      ]
    }
  ],
  "relations": [
    {
      "from": "guo2017-calibration",
      "type": "builds_on",
      "to": "platt1999-probabilistic"
    },
    {
      "from": "guo2017-calibration",
      "type": "evaluates",
      "to": "naeini2015-bbq"
    },
    {
      "from": "guo2017-calibration",
      "type": "compares_with",
      "to": "zadrozny2001-calibrated"
    },
    {
      "from": "guo2017-calibration",
      "type": "extends_to_multiclass",
      "to": "zadrozny2002-multiclass"
    },
    {
      "from": "guo2017-calibration",
      "type": "uses_metric_lineage",
      "to": "degroot1983-forecasters"
    },
    {
      "from": "guo2017-calibration",
      "type": "uses_metric_lineage",
      "to": "niculescu2005-probabilities"
    },
    {
      "from": "guo2017-calibration",
      "type": "contrasts_with",
      "to": "gal2016-dropout"
    },
    {
      "from": "guo2017-calibration",
      "type": "contrasts_with",
      "to": "lakshminarayanan2017-ensembles"
    },
    {
      "from": "guo2017-calibration",
      "type": "evaluates_architecture",
      "to": "he2016-resnet"
    },
    {
      "from": "guo2017-calibration",
      "type": "evaluates_architecture",
      "to": "huang2016-stochastic-depth"
    },
    {
      "from": "guo2017-calibration",
      "type": "evaluates_architecture",
      "to": "zagoruyko2016-wide-resnet"
    },
    {
      "from": "guo2017-calibration",
      "type": "evaluates_architecture",
      "to": "huang2017-densenet"
    }
  ]
}