{
  "id": 6981,
  "title": "Reproducing the paper \"Machine Learning Attacks Against the Asirra CAPTCHA\"",
  "url": "/competitions/dogs-vs-cats/discussion/6981",
  "author_name": "",
  "post_date": "2014-02-01T12:52:15.280Z",
  "votes": 4,
  "comment_count": 4,
  "views": 4118,
  "content": "<p>Did anyone of your read this paper, which was also mentioned in the Description of the competition? I won't help to go on the top of the leaderboard, but I found it nevertheless interesting. I tried to reproduce everything with python and scikit leanrn und posted my result on github and as an <a href=\"http://nbviewer.ipython.org/github/Safadurimo/cats-and-dogs/blob/master/catsdogs.ipynb\">ipython-notebook</a>. While I suceeded reproducing parts very exactly achieving an accurancy of 77%, I failed on other parts.</p>\n<p>I would love to hear your feedback on this!</p>",
  "messages": [
    {
      "id": "38223",
      "postDate": "02/01/2014 12:52:15",
      "content": "<p>Did anyone of your read this paper, which was also mentioned in the Description of the competition? I won't help to go on the top of the leaderboard, but I found it nevertheless interesting. I tried to reproduce everything with python and scikit leanrn und posted my result on github and as an <a href=\"http://nbviewer.ipython.org/github/Safadurimo/cats-and-dogs/blob/master/catsdogs.ipynb\">ipython-notebook</a>. While I suceeded reproducing parts very exactly achieving an accurancy of 77%, I failed on other parts.</p>\n<p>I would love to hear your feedback on this!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "38230",
      "postDate": "02/01/2014 17:29:50",
      "content": "<p>Thanks for the write-up, it was interesting.</p>\n<p>I noticed you switched to Logistic Regression which seemed to work better on your simplified textures. Did you try any Grid Search to tune hyperparameters, or try other algorithms (e.g. Stochastic Gradient Descent)?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "38256",
      "postDate": "02/02/2014 01:58:46",
      "content": "<p>Thanks for sharing the notebook.</p>\n<p>I have used only the color features as described in the paper and achieved an accuracy of 79.4%.</p>\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "38431",
      "postDate": "02/06/2014 16:35:43",
      "content": "<p>I did the same thing...tried out the color features approach...achieved about 77% as well. &nbsp;I tried to do the tile approach but decided not to pursue. &nbsp;Initial tests indicated that it would take about 20 seconds per image on my laptop to extract the features. &nbsp;Initially I was doing about 40 seconds per image but with some vectorization of R code and in-lining functions I reduced the time but could not get it faster than the 20 seconds.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "38674",
      "postDate": "02/12/2014 12:34:09",
      "content": "<p>I contacted the author about the texture features, he was very helpful and confirmed that the texture features really take some time (in the order of tens of seconds per image). So it was just patience, which was missing :-)</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 38230,
      "author_name": "justinsheek",
      "author_url": "",
      "post_date": "02/01/2014 17:29:50",
      "content": "<p>Thanks for the write-up, it was interesting.</p>\n<p>I noticed you switched to Logistic Regression which seemed to work better on your simplified textures. Did you try any Grid Search to tune hyperparameters, or try other algorithms (e.g. Stochastic Gradient Descent)?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 38256,
      "author_name": "kirubakumaresh",
      "author_url": "",
      "post_date": "02/02/2014 01:58:46",
      "content": "<p>Thanks for sharing the notebook.</p>\n<p>I have used only the color features as described in the paper and achieved an accuracy of 79.4%.</p>\n<p>&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 38431,
      "author_name": "jimthompson",
      "author_url": "",
      "post_date": "02/06/2014 16:35:43",
      "content": "<p>I did the same thing...tried out the color features approach...achieved about 77% as well. &nbsp;I tried to do the tile approach but decided not to pursue. &nbsp;Initial tests indicated that it would take about 20 seconds per image on my laptop to extract the features. &nbsp;Initially I was doing about 40 seconds per image but with some vectorization of R code and in-lining functions I reduced the time but could not get it faster than the 20 seconds.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 38674,
      "author_name": "safadurimo",
      "author_url": "",
      "post_date": "02/12/2014 12:34:09",
      "content": "<p>I contacted the author about the texture features, he was very helpful and confirmed that the texture features really take some time (in the order of tens of seconds per image). So it was just patience, which was missing :-)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "38223": "",
    "38230": "",
    "38256": "",
    "38431": "",
    "38674": ""
  },
  "source": "meta"
}