{
  "id": 172103,
  "title": "What is better? CNN with no meta INFO + Meta ensemble VS CNN with Meta",
  "url": "/competitions/siim-isic-melanoma-classification/discussion/172103",
  "author_name": "",
  "post_date": "2020-08-03T17:37:30.897588300Z",
  "votes": 4,
  "comment_count": 3,
  "views": 0,
  "content": "<p>i have experimented CNN with no meta INFO + Meta Ensemble vs CNN with Meta</p>\n\n<p>I have always seen CNN with no meta INFO + Meta Ensemble is better!\n(My current state is using CNN  + meta blending ensemble(LB:9600))</p>\n\n<p>I have 384 to 784 all noisy-student 5 fold CV</p>\n\n<p>It contradicts my intuition and confuses me</p>\n\n<p>ISIC 2019 competition 1st solution also deployed META+CNN concat strategy</p>",
  "messages": [
    {
      "id": "956693",
      "postDate": "08/03/2020 17:37:30",
      "content": "<p>i have experimented CNN with no meta INFO + Meta Ensemble vs CNN with Meta</p>\n\n<p>I have always seen CNN with no meta INFO + Meta Ensemble is better!\n(My current state is using CNN  + meta blending ensemble(LB:9600))</p>\n\n<p>I have 384 to 784 all noisy-student 5 fold CV</p>\n\n<p>It contradicts my intuition and confuses me</p>\n\n<p>ISIC 2019 competition 1st solution also deployed META+CNN concat strategy</p>",
      "rawMarkdown": "i have experimented CNN with no meta INFO + Meta Ensemble vs CNN with Meta\n\nI have always seen CNN with no meta INFO + Meta Ensemble is better!\n(My current state is using CNN  + meta blending ensemble(LB:9600))\n\nI have 384 to 784 all noisy-student 5 fold CV\n\nIt contradicts my intuition and confuses me\n\n\nISIC 2019 competition 1st solution also deployed META+CNN concat strategy",
      "votes": null
    },
    {
      "id": "956721",
      "postDate": "08/03/2020 18:08:28",
      "content": "<p>I think that Meta Info in Neural Network with CNN lead us to overfitting, consider that it is better to use simple models for Meta Info and then blend them with CNN. \nDo not see the reason to intersect these two sources in a big Neural Network. Also it seems that Meta Info in Train and in Test behave very different, it will be convenient just to exclude Meta Info from blending if you find some bad feature among Meta</p>",
      "rawMarkdown": "I think that Meta Info in Neural Network with CNN lead us to overfitting, consider that it is better to use simple models for Meta Info and then blend them with CNN. \nDo not see the reason to intersect these two sources in a big Neural Network. Also it seems that Meta Info in Train and in Test behave very different, it will be convenient just to exclude Meta Info from blending if you find some bad feature among Meta",
      "votes": null
    },
    {
      "id": "968606",
      "postDate": "08/13/2020 06:19:51",
      "content": "<p>As I had tried out many experiments, I found out that image and meta dual input CNN leads to overfitting to CV and lowers LB<br>\nIt makes CV go higher, but lowers LB.</p>",
      "rawMarkdown": "As I had tried out many experiments, I found out that image and meta dual input CNN leads to overfitting to CV and lowers LB\nIt makes CV go higher, but lowers LB.",
      "votes": null
    },
    {
      "id": "972529",
      "postDate": "08/16/2020 16:06:06",
      "content": "<p>for me, CNN without meta but with additional 2018 data is better than CNN with meta using 2019 data alone. </p>",
      "rawMarkdown": "for me, CNN without meta but with additional 2018 data is better than CNN with meta using 2019 data alone.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 972529,
      "author_name": "yimacs",
      "author_url": "",
      "post_date": "08/16/2020 16:06:06",
      "content": "<p>for me, CNN without meta but with additional 2018 data is better than CNN with meta using 2019 data alone. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 956721,
      "author_name": "aybatov",
      "author_url": "",
      "post_date": "08/03/2020 18:08:28",
      "content": "<p>I think that Meta Info in Neural Network with CNN lead us to overfitting, consider that it is better to use simple models for Meta Info and then blend them with CNN. \nDo not see the reason to intersect these two sources in a big Neural Network. Also it seems that Meta Info in Train and in Test behave very different, it will be convenient just to exclude Meta Info from blending if you find some bad feature among Meta</p>",
      "votes": null,
      "replies": [
        {
          "id": 968606,
          "author_name": "deepkim",
          "author_url": "",
          "post_date": "08/13/2020 06:19:51",
          "content": "<p>As I had tried out many experiments, I found out that image and meta dual input CNN leads to overfitting to CV and lowers LB<br>\nIt makes CV go higher, but lowers LB.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "956693": "i have experimented CNN with no meta INFO + Meta Ensemble vs CNN with Meta\n\nI have always seen CNN with no meta INFO + Meta Ensemble is better!\n(My current state is using CNN  + meta blending ensemble(LB:9600))\n\nI have 384 to 784 all noisy-student 5 fold CV\n\nIt contradicts my intuition and confuses me\n\n\nISIC 2019 competition 1st solution also deployed META+CNN concat strategy",
    "956721": "I think that Meta Info in Neural Network with CNN lead us to overfitting, consider that it is better to use simple models for Meta Info and then blend them with CNN. \nDo not see the reason to intersect these two sources in a big Neural Network. Also it seems that Meta Info in Train and in Test behave very different, it will be convenient just to exclude Meta Info from blending if you find some bad feature among Meta",
    "968606": "As I had tried out many experiments, I found out that image and meta dual input CNN leads to overfitting to CV and lowers LB\nIt makes CV go higher, but lowers LB.",
    "972529": "for me, CNN without meta but with additional 2018 data is better than CNN with meta using 2019 data alone."
  },
  "source": "meta"
}