{
  "id": 13019,
  "title": "Plankton classification as standard ML benchmark",
  "url": "/competitions/datasciencebowl/discussion/13019",
  "author_name": "",
  "post_date": "2015-03-24T08:18:15.357Z",
  "votes": 5,
  "comment_count": 2,
  "views": 1293,
  "content": "<p>Dear @Jessica Luo and&nbsp;Hatfield Marine Science Center team,</p>\n<p>What do you think about maintaining a some small separate server for permanent evaluation of the new-coming algorithms as it is done for Pascal VOC&nbsp;http://host.robots.ox.ac.uk:8080/ or ImageNet competition&nbsp;http://www.image-net.org/challenges/LSVRC/2014/ ?&nbsp;</p>\n<p>The benefits:&nbsp;</p>\n<p>For&nbsp;Hatfield Marine Science Center: with time you will get better and better algorithms just because some ML &amp; CV researcher would like to test their new ideas.&nbsp;</p>\n<p>Additional bonus - each one who publish results, will cite some paper describing dataset.</p>\n<p>For CV &amp; ML researchers: new nice medium-size&nbsp;dataset where you can improve a lot (CIFAR and MNIST are &quot;solved&quot; long time ago and ImageNet still too time consuming). And eventually some of them could join your or other biological team.</p>",
  "messages": [
    {
      "id": "67920",
      "postDate": "03/24/2015 08:18:15",
      "content": "<p>Dear @Jessica Luo and&nbsp;Hatfield Marine Science Center team,</p>\n<p>What do you think about maintaining a some small separate server for permanent evaluation of the new-coming algorithms as it is done for Pascal VOC&nbsp;http://host.robots.ox.ac.uk:8080/ or ImageNet competition&nbsp;http://www.image-net.org/challenges/LSVRC/2014/ ?&nbsp;</p>\n<p>The benefits:&nbsp;</p>\n<p>For&nbsp;Hatfield Marine Science Center: with time you will get better and better algorithms just because some ML &amp; CV researcher would like to test their new ideas.&nbsp;</p>\n<p>Additional bonus - each one who publish results, will cite some paper describing dataset.</p>\n<p>For CV &amp; ML researchers: new nice medium-size&nbsp;dataset where you can improve a lot (CIFAR and MNIST are &quot;solved&quot; long time ago and ImageNet still too time consuming). And eventually some of them could join your or other biological team.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "68021",
      "postDate": "03/24/2015 17:10:09",
      "content": "<p>Hi @old-ufo,&nbsp;</p>\n\n<p>Thanks for your suggestion. We have thought about putting the data online somewhere for the sake of having a citation source&nbsp;for any papers that would come out of this effort, but kaggle is also going to maintain the data (and keep submissions open) for this competition, so we didn't feel the immediate need to do so. However, if there truly is interest in keeping this data public and available as a ML benchmark dataset, we can certainly explore the possibility of making the data available on our own servers. For &quot;Planktonset 1.1,&quot; or whatever we call it, we will probably need to clean up a few labels and classes that were problematic. Also possibly explore a slightly different train/test ratio than 50:50. I see that MNIST is 90:10, so we could go with that....</p>\n\n<p>Thanks,</p>\n<p>Jessica</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "68444",
      "postDate": "03/26/2015 17:03:23",
      "content": "",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 68021,
      "author_name": "jessicaluo",
      "author_url": "",
      "post_date": "03/24/2015 17:10:09",
      "content": "<p>Hi @old-ufo,&nbsp;</p>\n\n<p>Thanks for your suggestion. We have thought about putting the data online somewhere for the sake of having a citation source&nbsp;for any papers that would come out of this effort, but kaggle is also going to maintain the data (and keep submissions open) for this competition, so we didn't feel the immediate need to do so. However, if there truly is interest in keeping this data public and available as a ML benchmark dataset, we can certainly explore the possibility of making the data available on our own servers. For &quot;Planktonset 1.1,&quot; or whatever we call it, we will probably need to clean up a few labels and classes that were problematic. Also possibly explore a slightly different train/test ratio than 50:50. I see that MNIST is 90:10, so we could go with that....</p>\n\n<p>Thanks,</p>\n<p>Jessica</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 68444,
      "author_name": "kulish",
      "author_url": "",
      "post_date": "03/26/2015 17:03:23",
      "content": "",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "67920": "",
    "68021": "",
    "68444": ""
  },
  "source": "meta"
}