{
  "id": 11481,
  "title": "How complex is it? How many features/predictors/computing power does it take?",
  "url": "/competitions/datasciencebowl/discussion/11481",
  "author_name": "",
  "post_date": "2015-01-10T12:27:58.523Z",
  "votes": 1,
  "comment_count": 6,
  "views": 3864,
  "content": "<p>This is my first project on Kaggle and I try to get an idea of the complexity of the task at hand!</p>\n<p>Would the leaders accept to share a bit of their methodology, without giving away too much details such that new comers can get up to speed?</p>\n<p>In particular,</p>\n<p>- How many features/predictors do you need to distinguish between the 121 classes?</p>\n<p>- How much computing power/computation time does it take to classify the test sample?</p>\n<p>- What kind of machine learning algorithms are you using?</p>\n<p>Thanks</p>",
  "messages": [
    {
      "id": "62556",
      "postDate": "01/10/2015 12:27:58",
      "content": "<p>This is my first project on Kaggle and I try to get an idea of the complexity of the task at hand!</p>\n<p>Would the leaders accept to share a bit of their methodology, without giving away too much details such that new comers can get up to speed?</p>\n<p>In particular,</p>\n<p>- How many features/predictors do you need to distinguish between the 121 classes?</p>\n<p>- How much computing power/computation time does it take to classify the test sample?</p>\n<p>- What kind of machine learning algorithms are you using?</p>\n<p>Thanks</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62590",
      "postDate": "01/10/2015 22:29:19",
      "content": "<p>See the starting guides for <a href=\"https://www.kaggle.com/c/datasciencebowl/forums/t/11242/getting-started-a-tutorial\">python</a>, <a href=\"https://www.kaggle.com/c/datasciencebowl/forums/t/11279/public-start-guide-of-deep-network-with-a-score-of-1-382285\">neural net</a>, and&nbsp;<a href=\"https://www.kaggle.com/c/datasciencebowl/forums/t/11408/starter-code-for-r-classification\">R</a></p>\n\n<p>The Python and R starting guides are very simple: They read the images and calculate various dimensions. It is a good start, and&nbsp;you can easily&nbsp;improve the code to get a better model. The neural net code is much more complex, is uses python to prepare the images and then launches another program (CXXNET) that you have to install. Most modern image classifiers use this kind of deep neural net so it will&nbsp;give you a good result if you get it working.</p>\n\n<p>Good luck,</p>\n<p>Jeff</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62594",
      "postDate": "01/10/2015 23:19:33",
      "content": "<p>Since the data is coming in image form, I would bet that some variant of a convolutional neural network will almost certainly win. That being said, you could certainly try other methods (trees, SVMs, etc.) on some extracted feature set to see how they perform.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62638",
      "postDate": "01/12/2015 09:48:45",
      "content": "<p>Thanks for your replies. I started with a simple model to get my feet wet..</p>\n<p>A logistic model with up to 15 predictors. The features are purely based on some image metrics:</p>\n<ul>\n<li>filesize in bytes</li>\n</ul>\n<ul>\n<li>width in pixels</li>\n</ul>\n<ul>\n<li>height in pixels</li>\n<li>area in pixel^2</li>\n<li>aspect ratio</li>\n<li>compression ratio (=compressed filesize/uncompressed filesize)</li>\n<li>image mean</li>\n<li>image standard deviation</li>\n<li>histogram</li>\n</ul>\n<p>This gives a score of 2.842449...</p>\n<p>Some of these predictors seem to be able to distinguish some classes from the others</p>\n<p>I will try to add more predictors to the model and see how it goes...</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62720",
      "postDate": "01/14/2015 15:29:20",
      "content": "<p>I know neural nets and all, what is a &quot;convolutional&quot; neural net? Could you please post some references from where one can learn about them?&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63026",
      "postDate": "01/22/2015 19:33:53",
      "content": "<p>This should give you something to start on</p>\n<p>http://deeplearning.net/tutorial/lenet.html</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "63112",
      "postDate": "01/24/2015 20:15:11",
      "content": "<p>Not sure why Udaya's post got so many downvotes. Are we penalizing people for asking questions now?</p>\n<p>But Udaya, check out Geoffery Hinton's course on Neural Networks if you want to go in depth:</p>\n<p>https://class.coursera.org/neuralnets-2012-001</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 62590,
      "author_name": "jeffhebert",
      "author_url": "",
      "post_date": "01/10/2015 22:29:19",
      "content": "<p>See the starting guides for <a href=\"https://www.kaggle.com/c/datasciencebowl/forums/t/11242/getting-started-a-tutorial\">python</a>, <a href=\"https://www.kaggle.com/c/datasciencebowl/forums/t/11279/public-start-guide-of-deep-network-with-a-score-of-1-382285\">neural net</a>, and&nbsp;<a href=\"https://www.kaggle.com/c/datasciencebowl/forums/t/11408/starter-code-for-r-classification\">R</a></p>\n\n<p>The Python and R starting guides are very simple: They read the images and calculate various dimensions. It is a good start, and&nbsp;you can easily&nbsp;improve the code to get a better model. The neural net code is much more complex, is uses python to prepare the images and then launches another program (CXXNET) that you have to install. Most modern image classifiers use this kind of deep neural net so it will&nbsp;give you a good result if you get it working.</p>\n\n<p>Good luck,</p>\n<p>Jeff</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62594,
      "author_name": "jesselivezey",
      "author_url": "",
      "post_date": "01/10/2015 23:19:33",
      "content": "<p>Since the data is coming in image form, I would bet that some variant of a convolutional neural network will almost certainly win. That being said, you could certainly try other methods (trees, SVMs, etc.) on some extracted feature set to see how they perform.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62638,
      "author_name": "kappaanalytics",
      "author_url": "",
      "post_date": "01/12/2015 09:48:45",
      "content": "<p>Thanks for your replies. I started with a simple model to get my feet wet..</p>\n<p>A logistic model with up to 15 predictors. The features are purely based on some image metrics:</p>\n<ul>\n<li>filesize in bytes</li>\n</ul>\n<ul>\n<li>width in pixels</li>\n</ul>\n<ul>\n<li>height in pixels</li>\n<li>area in pixel^2</li>\n<li>aspect ratio</li>\n<li>compression ratio (=compressed filesize/uncompressed filesize)</li>\n<li>image mean</li>\n<li>image standard deviation</li>\n<li>histogram</li>\n</ul>\n<p>This gives a score of 2.842449...</p>\n<p>Some of these predictors seem to be able to distinguish some classes from the others</p>\n<p>I will try to add more predictors to the model and see how it goes...</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62720,
      "author_name": "udayabhanu",
      "author_url": "",
      "post_date": "01/14/2015 15:29:20",
      "content": "<p>I know neural nets and all, what is a &quot;convolutional&quot; neural net? Could you please post some references from where one can learn about them?&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63026,
      "author_name": "jaythodenvanvelzen",
      "author_url": "",
      "post_date": "01/22/2015 19:33:53",
      "content": "<p>This should give you something to start on</p>\n<p>http://deeplearning.net/tutorial/lenet.html</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 63112,
      "author_name": "miroslaw",
      "author_url": "",
      "post_date": "01/24/2015 20:15:11",
      "content": "<p>Not sure why Udaya's post got so many downvotes. Are we penalizing people for asking questions now?</p>\n<p>But Udaya, check out Geoffery Hinton's course on Neural Networks if you want to go in depth:</p>\n<p>https://class.coursera.org/neuralnets-2012-001</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "62556": "",
    "62590": "",
    "62594": "",
    "62638": "",
    "62720": "",
    "63026": "",
    "63112": ""
  },
  "source": "meta"
}