{
  "id": 68358,
  "title": "Is the image with recognition=False useful?",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/68358",
  "author_name": "lulala",
  "post_date": "2018-10-11T22:19:08.765000",
  "votes": 3,
  "comment_count": 6,
  "views": 0,
  "content": "<p>Hi everyone, </p>\n\n<p>I'm wondering if the image instances with recognition=False are useful. In other words, will the performance/accuracy be improved by removing the unrecognized images from the training set?</p>\n\n<p>Thanks,</p>",
  "messages": [
    {
      "id": 402561,
      "postDate": "2018-10-11T22:19:08.767Z",
      "content": "<p>Hi everyone, </p>\n\n<p>I'm wondering if the image instances with recognition=False are useful. In other words, will the performance/accuracy be improved by removing the unrecognized images from the training set?</p>\n\n<p>Thanks,</p>",
      "rawMarkdown": "Hi everyone, \n\nI'm wondering if the image instances with recognition=False are useful. In other words, will the performance/accuracy be improved by removing the unrecognized images from the training set?\n\nThanks,\n ",
      "votes": 3
    },
    {
      "id": 405077,
      "postDate": "2018-10-16T20:38:35.093Z",
      "content": "<p>I tried to remove them and the LB score got a bit worse. It is up to you what you feed your neural nets though...</p>",
      "rawMarkdown": "I tried to remove them and the LB score got a bit worse. It is up to you what you feed your neural nets though...",
      "votes": 1
    },
    {
      "id": 427594,
      "postDate": "2018-11-25T19:58:20.090Z",
      "content": "<p>recognized=False  samples are definitely useful, I use all samples including recognized=False samples,  and give them a lower sample weight, like 0.5.</p>",
      "rawMarkdown": "recognized=False  samples are definitely useful, I use all samples including recognized=False samples,  and give them a lower sample weight, like 0.5."
    },
    {
      "id": 407482,
      "postDate": "2018-10-21T09:37:50.810Z",
      "content": "<p>My idea is to reduce their sample weights, but I have not compared different reductions yet.</p>",
      "rawMarkdown": "My idea is to reduce their sample weights, but I have not compared different reductions yet."
    },
    {
      "id": 403294,
      "postDate": "2018-10-13T09:02:26.957Z",
      "content": "<p>My guess is that removing them from the training set will decrease your accuracy. The best way to find out is to compare with and without performance leaving everything else the same.</p>",
      "rawMarkdown": "My guess is that removing them from the training set will decrease your accuracy. The best way to find out is to compare with and without performance leaving everything else the same."
    },
    {
      "id": 403107,
      "postDate": "2018-10-12T22:07:52.253Z",
      "content": "<p>As for me, given I dont know if the images from the test set  are recognized or not , I didn't filter them out from the training set. </p>\n\n<p>Anyway I use only features present in both the training and test set (drawing, countrycode) ....</p>\n\n<p>But why not trying it and see if it improves ? </p>",
      "rawMarkdown": "As for me, given I dont know if the images from the test set  are recognized or not , I didn't filter them out from the training set. \n\nAnyway I use only features present in both the training and test set (drawing, countrycode) ....\n\nBut why not trying it and see if it improves ? "
    },
    {
      "id": 402879,
      "postDate": "2018-10-12T13:52:49.387Z",
      "content": "<p>I have the same question, waiting for someone's reply:D</p>",
      "rawMarkdown": "I have the same question, waiting for someone's reply:D"
    }
  ],
  "comments": [
    {
      "id": 405077,
      "author_name": "beluga",
      "author_url": "",
      "post_date": "2018-10-16T20:38:35.093000",
      "content": "<p>I tried to remove them and the LB score got a bit worse. It is up to you what you feed your neural nets though...</p>",
      "votes": 1,
      "replies": []
    },
    {
      "id": 427594,
      "author_name": "[he.ai]soulmachine",
      "author_url": "",
      "post_date": "2018-11-25T19:58:20.090000",
      "content": "<p>recognized=False  samples are definitely useful, I use all samples including recognized=False samples,  and give them a lower sample weight, like 0.5.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 407482,
      "author_name": "Philipp Sodmann",
      "author_url": "",
      "post_date": "2018-10-21T09:37:50.810000",
      "content": "<p>My idea is to reduce their sample weights, but I have not compared different reductions yet.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 403294,
      "author_name": "Paul Jurczak",
      "author_url": "",
      "post_date": "2018-10-13T09:02:26.957000",
      "content": "<p>My guess is that removing them from the training set will decrease your accuracy. The best way to find out is to compare with and without performance leaving everything else the same.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 403107,
      "author_name": "Serigne ",
      "author_url": "",
      "post_date": "2018-10-12T22:07:52.253000",
      "content": "<p>As for me, given I dont know if the images from the test set  are recognized or not , I didn't filter them out from the training set. </p>\n\n<p>Anyway I use only features present in both the training and test set (drawing, countrycode) ....</p>\n\n<p>But why not trying it and see if it improves ? </p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 402879,
      "author_name": "upup",
      "author_url": "",
      "post_date": "2018-10-12T13:52:49.387000",
      "content": "<p>I have the same question, waiting for someone's reply:D</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "402561": "Hi everyone, \n\nI'm wondering if the image instances with recognition=False are useful. In other words, will the performance/accuracy be improved by removing the unrecognized images from the training set?\n\nThanks,\n ",
    "405077": "I tried to remove them and the LB score got a bit worse. It is up to you what you feed your neural nets though...",
    "427594": "recognized=False  samples are definitely useful, I use all samples including recognized=False samples,  and give them a lower sample weight, like 0.5.",
    "407482": "My idea is to reduce their sample weights, but I have not compared different reductions yet.",
    "403294": "My guess is that removing them from the training set will decrease your accuracy. The best way to find out is to compare with and without performance leaving everything else the same.",
    "403107": "As for me, given I dont know if the images from the test set  are recognized or not , I didn't filter them out from the training set. \n\nAnyway I use only features present in both the training and test set (drawing, countrycode) ....\n\nBut why not trying it and see if it improves ? ",
    "402879": "I have the same question, waiting for someone's reply:D"
  }
}