{
  "id": 277001,
  "title": "Model calibration",
  "url": "/competitions/rsna-miccai-brain-tumor-radiogenomic-classification/discussion/277001",
  "author_name": "Michał Choiński",
  "post_date": "2021-10-07T13:13:02.463000",
  "votes": 7,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Deep Learning classification models tend to return overconfident predictions. We say then that they are miscalibrated.<br>\nCalibrated models, on the other hand, take into account the accuracy of the estimator so that the obtained probabilities reflect the real probability of the class given the uncertainty of the model.<br>\nFurthermore, reducing calibration error improves overall performance of the model which is the ultimate goal that we all want to achieve here.</p>\n<p>I uploaded a <a href=\"https://www.kaggle.com/mikecho/calibration-framework-for-deep-learning\" target=\"_blank\">framework</a> that implements several methods to tackle the mentioned problem. Feel free to play with it and to calibrate your models smoothly.</p>",
  "messages": [
    {
      "id": 1537420,
      "postDate": "2021-10-07T13:13:02.463Z",
      "content": "<p>Deep Learning classification models tend to return overconfident predictions. We say then that they are miscalibrated.<br>\nCalibrated models, on the other hand, take into account the accuracy of the estimator so that the obtained probabilities reflect the real probability of the class given the uncertainty of the model.<br>\nFurthermore, reducing calibration error improves overall performance of the model which is the ultimate goal that we all want to achieve here.</p>\n<p>I uploaded a <a href=\"https://www.kaggle.com/mikecho/calibration-framework-for-deep-learning\" target=\"_blank\">framework</a> that implements several methods to tackle the mentioned problem. Feel free to play with it and to calibrate your models smoothly.</p>",
      "rawMarkdown": "Deep Learning classification models tend to return overconfident predictions. We say then that they are miscalibrated.\nCalibrated models, on the other hand, take into account the accuracy of the estimator so that the obtained probabilities reflect the real probability of the class given the uncertainty of the model.\nFurthermore, reducing calibration error improves overall performance of the model which is the ultimate goal that we all want to achieve here.\n\nI uploaded a [framework](https://www.kaggle.com/mikecho/calibration-framework-for-deep-learning) that implements several methods to tackle the mentioned problem. Feel free to play with it and to calibrate your models smoothly.",
      "votes": 7
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1537420": "Deep Learning classification models tend to return overconfident predictions. We say then that they are miscalibrated.\nCalibrated models, on the other hand, take into account the accuracy of the estimator so that the obtained probabilities reflect the real probability of the class given the uncertainty of the model.\nFurthermore, reducing calibration error improves overall performance of the model which is the ultimate goal that we all want to achieve here.\n\nI uploaded a [framework](https://www.kaggle.com/mikecho/calibration-framework-for-deep-learning) that implements several methods to tackle the mentioned problem. Feel free to play with it and to calibrate your models smoothly."
  }
}