{
  "id": 168469,
  "title": "How good are we doing so far?",
  "url": "/competitions/osic-pulmonary-fibrosis-progression/discussion/168469",
  "author_name": "",
  "post_date": "2020-07-20T18:52:31.606101200Z",
  "votes": 45,
  "comment_count": 8,
  "views": 0,
  "content": "<p>Given the loss of the competition ,  we are trying to minimize:</p>\n\n<p>$$ \\ell = \\frac{\\sqrt 2\\Delta}{\\sigma} + \\log(\\sqrt 2 \\sigma) \\text{ with a gap } \\Delta \n \\text{ and a standard deviation } \\sigma \\geq 70  $$</p>\n\n<p>The theoritical minimum is achieved for $$\\Delta=0 \\text{ and } \\quad  \\sigma = 70 \\text{ and is equal to } \\hat{\\ell} = \\log(70 \\sqrt 2) \\approx 4.59 $$.</p>\n\n<p>Overall the scores are grouped around 6.83, so we have a true gap of <strong>2.24</strong> to dampen. I hope you guys will reduce it with your magic. I hope that a great breakthrough can be done by  a wise adding of CT scans since most of the best kernels only use tabular data.</p>",
  "messages": [
    {
      "id": "937142",
      "postDate": "07/20/2020 18:52:31",
      "content": "<p>Given the loss of the competition ,  we are trying to minimize:</p>\n\n<p>$$ \\ell = \\frac{\\sqrt 2\\Delta}{\\sigma} + \\log(\\sqrt 2 \\sigma) \\text{ with a gap } \\Delta \n \\text{ and a standard deviation } \\sigma \\geq 70  $$</p>\n\n<p>The theoritical minimum is achieved for $$\\Delta=0 \\text{ and } \\quad  \\sigma = 70 \\text{ and is equal to } \\hat{\\ell} = \\log(70 \\sqrt 2) \\approx 4.59 $$.</p>\n\n<p>Overall the scores are grouped around 6.83, so we have a true gap of <strong>2.24</strong> to dampen. I hope you guys will reduce it with your magic. I hope that a great breakthrough can be done by  a wise adding of CT scans since most of the best kernels only use tabular data.</p>",
      "rawMarkdown": "Given the loss of the competition ,  we are trying to minimize:\n\n$$ \\ell = \\frac{\\sqrt 2\\Delta}{\\sigma} + \\log(\\sqrt 2 \\sigma) \\text{ with a gap } \\Delta \n \\text{ and a standard deviation } \\sigma \\geq 70  $$\n\nThe theoritical minimum is achieved for $$\\Delta=0 \\text{ and } \\quad  \\sigma = 70 \\text{ and is equal to } \\hat{\\ell} = \\log(70 \\sqrt 2) \\approx 4.59 $$.\n\nOverall the scores are grouped around 6.83, so we have a true gap of **2.24** to dampen. I hope you guys will reduce it with your magic. I hope that a great breakthrough can be done by  a wise adding of CT scans since most of the best kernels only use tabular data.",
      "votes": null
    },
    {
      "id": "938603",
      "postDate": "07/21/2020 16:16:06",
      "content": "<p>I've tried tuning hyperparameters and changing models, but I don't think that would help me much in the long term. We would need to change the techniques being used or wait for the <strong>\"magic\"</strong>.</p>\n<p>I hope we eventually get to see the gap close :)</p>",
      "rawMarkdown": "I've tried tuning hyperparameters and changing models, but I don't think that would help me much in the long term. We would need to change the techniques being used or wait for the **\"magic\"**.\n\nI hope we eventually get to see the gap close :)",
      "votes": null
    },
    {
      "id": "938662",
      "postDate": "07/21/2020 16:55:20",
      "content": "<p>Agree. Maybe i can repeat myself but i have high expectations on CT scan based features.</p>",
      "rawMarkdown": "Agree. Maybe i can repeat myself but i have high expectations on CT scan based features.",
      "votes": null
    },
    {
      "id": "938680",
      "postDate": "07/21/2020 17:14:26",
      "content": "<p>I'm working on CT scan based features, it's a bit hard to integrate with the model, but I'll try my best to extract something from the scans!</p>",
      "rawMarkdown": "I'm working on CT scan based features, it's a bit hard to integrate with the model, but I'll try my best to extract something from the scans!",
      "votes": null
    },
    {
      "id": "938739",
      "postDate": "07/21/2020 18:05:18",
      "content": "<p>good luck, <a href=\"https://www.kaggle.com/aadhavvignesh\" target=\"_blank\">@aadhavvignesh</a> </p>",
      "rawMarkdown": "good luck, @aadhavvignesh",
      "votes": null
    },
    {
      "id": "938962",
      "postDate": "07/21/2020 22:38:35",
      "content": "<p>Thanks for pointing out the best score achievable - always good to have a target!!</p>\n\n<p>Agreed, my next step is to do something with the image data. I'm hoping the \"Domain Expert's insights\" are going to be valuable.</p>",
      "rawMarkdown": "Thanks for pointing out the best score achievable - always good to have a target!!\n\nAgreed, my next step is to do something with the image data. I'm hoping the \"Domain Expert's insights\" are going to be valuable.",
      "votes": null
    },
    {
      "id": "990976",
      "postDate": "08/30/2020 02:44:18",
      "content": "<p><a href=\"https://www.kaggle.com/ulrich07\" target=\"_blank\">@ulrich07</a>  in your nn kernel ,can we assume that if delta is narrow so would be confidence prediction ?</p>",
      "rawMarkdown": "ulrich07  in your nn kernel ,can we assume that if delta is narrow so would be confidence prediction ?",
      "votes": null
    },
    {
      "id": "996001",
      "postDate": "09/03/2020 03:15:27",
      "content": "<p>Thanks for sharing this. i will try out the same as well to increase my score too</p>",
      "rawMarkdown": "Thanks for sharing this. i will try out the same as well to increase my score too",
      "votes": null
    },
    {
      "id": "996042",
      "postDate": "09/03/2020 04:03:31",
      "content": "<p>Hi! <a href=\"https://www.kaggle.com/ulrich07\" target=\"_blank\">@ulrich07</a>, I learned quite a lot from your Notebook trying to incorporate this weekend some CT image into the model to improve the score, thanks for pointing out the best score achievable </p>",
      "rawMarkdown": "Hi! @ulrich07, I learned quite a lot from your Notebook trying to incorporate this weekend some CT image into the model to improve the score, thanks for pointing out the best score achievable",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 938603,
      "author_name": "aadhavvignesh",
      "author_url": "",
      "post_date": "07/21/2020 16:16:06",
      "content": "<p>I've tried tuning hyperparameters and changing models, but I don't think that would help me much in the long term. We would need to change the techniques being used or wait for the <strong>\"magic\"</strong>.</p>\n<p>I hope we eventually get to see the gap close :)</p>",
      "votes": null,
      "replies": [
        {
          "id": 938662,
          "author_name": "ulrich07",
          "author_url": "",
          "post_date": "07/21/2020 16:55:20",
          "content": "<p>Agree. Maybe i can repeat myself but i have high expectations on CT scan based features.</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 938680,
          "author_name": "aadhavvignesh",
          "author_url": "",
          "post_date": "07/21/2020 17:14:26",
          "content": "<p>I'm working on CT scan based features, it's a bit hard to integrate with the model, but I'll try my best to extract something from the scans!</p>",
          "votes": null,
          "replies": []
        },
        {
          "id": 938739,
          "author_name": "ulrich07",
          "author_url": "",
          "post_date": "07/21/2020 18:05:18",
          "content": "<p>good luck, <a href=\"https://www.kaggle.com/aadhavvignesh\" target=\"_blank\">@aadhavvignesh</a> </p>",
          "votes": null,
          "replies": []
        }
      ]
    },
    {
      "id": 990976,
      "author_name": "jaideepvalani",
      "author_url": "",
      "post_date": "08/30/2020 02:44:18",
      "content": "<p><a href=\"https://www.kaggle.com/ulrich07\" target=\"_blank\">@ulrich07</a>  in your nn kernel ,can we assume that if delta is narrow so would be confidence prediction ?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 996042,
      "author_name": "cv13j0",
      "author_url": "",
      "post_date": "09/03/2020 04:03:31",
      "content": "<p>Hi! <a href=\"https://www.kaggle.com/ulrich07\" target=\"_blank\">@ulrich07</a>, I learned quite a lot from your Notebook trying to incorporate this weekend some CT image into the model to improve the score, thanks for pointing out the best score achievable </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 938962,
      "author_name": "andypenrose",
      "author_url": "",
      "post_date": "07/21/2020 22:38:35",
      "content": "<p>Thanks for pointing out the best score achievable - always good to have a target!!</p>\n\n<p>Agreed, my next step is to do something with the image data. I'm hoping the \"Domain Expert's insights\" are going to be valuable.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 996001,
      "author_name": "blessondensil294",
      "author_url": "",
      "post_date": "09/03/2020 03:15:27",
      "content": "<p>Thanks for sharing this. i will try out the same as well to increase my score too</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "937142": "Given the loss of the competition ,  we are trying to minimize:\n\n$$ \\ell = \\frac{\\sqrt 2\\Delta}{\\sigma} + \\log(\\sqrt 2 \\sigma) \\text{ with a gap } \\Delta \n \\text{ and a standard deviation } \\sigma \\geq 70  $$\n\nThe theoritical minimum is achieved for $$\\Delta=0 \\text{ and } \\quad  \\sigma = 70 \\text{ and is equal to } \\hat{\\ell} = \\log(70 \\sqrt 2) \\approx 4.59 $$.\n\nOverall the scores are grouped around 6.83, so we have a true gap of **2.24** to dampen. I hope you guys will reduce it with your magic. I hope that a great breakthrough can be done by  a wise adding of CT scans since most of the best kernels only use tabular data.",
    "938603": "I've tried tuning hyperparameters and changing models, but I don't think that would help me much in the long term. We would need to change the techniques being used or wait for the **\"magic\"**.\n\nI hope we eventually get to see the gap close :)",
    "938662": "Agree. Maybe i can repeat myself but i have high expectations on CT scan based features.",
    "938680": "I'm working on CT scan based features, it's a bit hard to integrate with the model, but I'll try my best to extract something from the scans!",
    "938739": "good luck, @aadhavvignesh",
    "938962": "Thanks for pointing out the best score achievable - always good to have a target!!\n\nAgreed, my next step is to do something with the image data. I'm hoping the \"Domain Expert's insights\" are going to be valuable.",
    "990976": "ulrich07  in your nn kernel ,can we assume that if delta is narrow so would be confidence prediction ?",
    "996001": "Thanks for sharing this. i will try out the same as well to increase my score too",
    "996042": "Hi! @ulrich07, I learned quite a lot from your Notebook trying to incorporate this weekend some CT image into the model to improve the score, thanks for pointing out the best score achievable"
  },
  "source": "meta"
}