{
  "id": 73473,
  "title": "online easy example mining for dealing with noise",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/73473",
  "author_name": "DavidGbodiOdaibo",
  "post_date": "2018-12-03T13:25:18.192000",
  "votes": 0,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I have not implemented this but I think this approach can be an efficient and natural way to deal with a noisy data-set. Online hard example mining identifies hard examples by excluding the most correct samples in backprop and taking a certain percentage of the most difficult predictions  in the gradient batch update. I think we can flip this and take a certain percentage of the easiest samples during batch gradient update which will naturally exclude the noisy/wrong samples because the noisy samples are more likely to be predicted wrong by the model.  We simply need the estimate the percentage of noise we think is in the data-set as a hyper-parameter. Correct samples that are unfortunately initially excluded from the gradient update should naturally work their way into the easy sample pool.</p>",
  "messages": [
    {
      "id": 432289,
      "postDate": "2018-12-03T17:14:13.657Z",
      "content": "<p>I believe a very similar idea has been investigated in this paper: <a href=\"https://arxiv.org/abs/1809.11008\">https://arxiv.org/abs/1809.11008</a></p>",
      "rawMarkdown": "I believe a very similar idea has been investigated in this paper: https://arxiv.org/abs/1809.11008",
      "votes": 1,
      "replies": [
        {
          "id": 432380,
          "postDate": "2018-12-03T19:37:44.940Z",
          "content": "<p>interesting!</p>",
          "rawMarkdown": "interesting!"
        }
      ]
    },
    {
      "id": 434290,
      "postDate": "2018-12-06T06:52:53.087Z",
      "content": "<p>emm, this is exactly what i am thinking, although i have no time to train ....</p>",
      "rawMarkdown": "emm, this is exactly what i am thinking, although i have no time to train ...."
    },
    {
      "id": 432153,
      "postDate": "2018-12-03T13:25:18.193Z",
      "content": "<p>I have not implemented this but I think this approach can be an efficient and natural way to deal with a noisy data-set. Online hard example mining identifies hard examples by excluding the most correct samples in backprop and taking a certain percentage of the most difficult predictions  in the gradient batch update. I think we can flip this and take a certain percentage of the easiest samples during batch gradient update which will naturally exclude the noisy/wrong samples because the noisy samples are more likely to be predicted wrong by the model.  We simply need the estimate the percentage of noise we think is in the data-set as a hyper-parameter. Correct samples that are unfortunately initially excluded from the gradient update should naturally work their way into the easy sample pool.</p>",
      "rawMarkdown": "I have not implemented this but I think this approach can be an efficient and natural way to deal with a noisy data-set. Online hard example mining identifies hard examples by excluding the most correct samples in backprop and taking a certain percentage of the most difficult predictions  in the gradient batch update. I think we can flip this and take a certain percentage of the easiest samples during batch gradient update which will naturally exclude the noisy/wrong samples because the noisy samples are more likely to be predicted wrong by the model.  We simply need the estimate the percentage of noise we think is in the data-set as a hyper-parameter. Correct samples that are unfortunately initially excluded from the gradient update should naturally work their way into the easy sample pool."
    }
  ],
  "comments": [
    {
      "id": 432289,
      "author_name": "Dmytro Danevskyi",
      "author_url": "",
      "post_date": "2018-12-03T17:14:13.657000",
      "content": "<p>I believe a very similar idea has been investigated in this paper: <a href=\"https://arxiv.org/abs/1809.11008\">https://arxiv.org/abs/1809.11008</a></p>",
      "votes": 1,
      "replies": [
        {
          "id": 432380,
          "author_name": "DavidGbodiOdaibo",
          "author_url": "",
          "post_date": "2018-12-03T19:37:44.940000",
          "content": "<p>interesting!</p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 434290,
      "author_name": "good good study",
      "author_url": "",
      "post_date": "2018-12-06T06:52:53.087000",
      "content": "<p>emm, this is exactly what i am thinking, although i have no time to train ....</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "432289": "I believe a very similar idea has been investigated in this paper: https://arxiv.org/abs/1809.11008",
    "434290": "emm, this is exactly what i am thinking, although i have no time to train ....",
    "432153": "I have not implemented this but I think this approach can be an efficient and natural way to deal with a noisy data-set. Online hard example mining identifies hard examples by excluding the most correct samples in backprop and taking a certain percentage of the most difficult predictions  in the gradient batch update. I think we can flip this and take a certain percentage of the easiest samples during batch gradient update which will naturally exclude the noisy/wrong samples because the noisy samples are more likely to be predicted wrong by the model.  We simply need the estimate the percentage of noise we think is in the data-set as a hyper-parameter. Correct samples that are unfortunately initially excluded from the gradient update should naturally work their way into the easy sample pool."
  }
}