{
  "id": 6825,
  "title": "Best result without any external data?",
  "url": "/competitions/dogs-vs-cats/discussion/6825",
  "author_name": "",
  "post_date": "2014-01-13T14:04:06.340Z",
  "votes": null,
  "comment_count": 4,
  "views": 4153,
  "content": "<p>Given this is a research comp, I thought I would ask: what is your best result WITHOUT external data? It seems many topic performers are utilizing/incorporating pre-trained models with ImageNet data.</p>\n<p>&nbsp;</p>\n<p>&nbsp;</p>",
  "messages": [
    {
      "id": "37348",
      "postDate": "01/13/2014 14:04:06",
      "content": "<p>Given this is a research comp, I thought I would ask: what is your best result WITHOUT external data? It seems many topic performers are utilizing/incorporating pre-trained models with ImageNet data.</p>\n<p>&nbsp;</p>\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "37399",
      "postDate": "01/15/2014 12:22:13",
      "content": "<p>I implemented the approach which have been described in the paper &quot;Machine Learning Attacks Against the Asirra CAPTCHA&quot;, which was also mentioned in the description of the competition. I only implemented the features based on colors, not on textures. That gave me a score of 77,4 %, which is pretty close to the figures reported in that paper.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "37400",
      "postDate": "01/15/2014 13:11:04",
      "content": "<p>My best leaderboard&nbsp;score without the use of external data was&nbsp;0.89787. &nbsp;I used a convnet with three convolutional layers and two fully connected layers with dropout.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "37402",
      "postDate": "01/15/2014 13:30:17",
      "content": "<p>My best submission comes from averaging two convolutional neural networks trained on 24'000 samples out of the 25'000 provided (LB score:&nbsp;0.91387). 1000 samples are the hold-out set to prevent overfitting. No other preprocessing applied other than rescaling. But I think there is room for improvement (rotate images, flip images, train on different subsets and average results etc.)</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "37407",
      "postDate": "01/15/2014 15:59:20",
      "content": "<p>0.969</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 37399,
      "author_name": "safadurimo",
      "author_url": "",
      "post_date": "01/15/2014 12:22:13",
      "content": "<p>I implemented the approach which have been described in the paper &quot;Machine Learning Attacks Against the Asirra CAPTCHA&quot;, which was also mentioned in the description of the competition. I only implemented the features based on colors, not on textures. That gave me a score of 77,4 %, which is pretty close to the figures reported in that paper.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 37400,
      "author_name": "dnouri",
      "author_url": "",
      "post_date": "01/15/2014 13:11:04",
      "content": "<p>My best leaderboard&nbsp;score without the use of external data was&nbsp;0.89787. &nbsp;I used a convnet with three convolutional layers and two fully connected layers with dropout.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 37402,
      "author_name": "mattvonrohr",
      "author_url": "",
      "post_date": "01/15/2014 13:30:17",
      "content": "<p>My best submission comes from averaging two convolutional neural networks trained on 24'000 samples out of the 25'000 provided (LB score:&nbsp;0.91387). 1000 samples are the hold-out set to prevent overfitting. No other preprocessing applied other than rescaling. But I think there is room for improvement (rotate images, flip images, train on different subsets and average results etc.)</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 37407,
      "author_name": "neocortex",
      "author_url": "",
      "post_date": "01/15/2014 15:59:20",
      "content": "<p>0.969</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "37348": "",
    "37399": "",
    "37400": "",
    "37402": "",
    "37407": ""
  },
  "source": "meta"
}