{
  "id": 11495,
  "title": "Feature based approaches",
  "url": "/competitions/datasciencebowl/discussion/11495",
  "author_name": "",
  "post_date": "2015-01-12T09:41:49.570Z",
  "votes": 1,
  "comment_count": 5,
  "views": 2165,
  "content": "<p>I was wondering if someone using only feature based + shallow learning wants to share public ldb score or cv to see how they compete to deep learning.</p>\n<p>I use morphological features+RF: 1.61 on public ldb</p>",
  "messages": [
    {
      "id": "62637",
      "postDate": "01/12/2015 09:41:49",
      "content": "<p>I was wondering if someone using only feature based + shallow learning wants to share public ldb score or cv to see how they compete to deep learning.</p>\n<p>I use morphological features+RF: 1.61 on public ldb</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62661",
      "postDate": "01/12/2015 17:04:21",
      "content": "<p>My approach is feature-based with&nbsp;a linear svm I am at 1.09 by 10-fold cross validation on the training data.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62673",
      "postDate": "01/13/2015 03:59:31",
      "content": "<p>I've just started with a feature-based approach and got 1.85 so far with 11 features (aspect ratio and various areas) and RF.</p>\n<p>Not having a CUDA GPU, I plan to see how far I can get with just image features and maybe a simpler net.</p>\n<p>@kartal, 1.09 is pretty good, any features you suggest?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62674",
      "postDate": "01/13/2015 04:09:13",
      "content": "<p>I have close to 10K features. These are mostly patch-based and includes some convolution features as well. So all I am doing is obtaining &nbsp;patches of different scales and then pass them through k-means to obtain visual words. Then use SVM to classify images. I am not sure how far I will go by this technique but I have a feeling this technique can be as good as a coarsely-tuned CNN.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62693",
      "postDate": "01/13/2015 14:36:40",
      "content": "<p>@kartal: Can you give some references to the features you are using. I am using deep belief networks and want to experiment with some feature engineering as well.&nbsp;Thanks!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62694",
      "postDate": "01/13/2015 14:43:41",
      "content": "<p>Hi Mike,</p>\n<p>This paper should help.&nbsp;</p>\n<p>http://web.eecs.umich.edu/~honglak/aistats11-AnalysisSingleLayerUnsupervisedFeatureLearning.pdf</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 62661,
      "author_name": "blackeagle",
      "author_url": "",
      "post_date": "01/12/2015 17:04:21",
      "content": "<p>My approach is feature-based with&nbsp;a linear svm I am at 1.09 by 10-fold cross validation on the training data.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62673,
      "author_name": "photosynth",
      "author_url": "",
      "post_date": "01/13/2015 03:59:31",
      "content": "<p>I've just started with a feature-based approach and got 1.85 so far with 11 features (aspect ratio and various areas) and RF.</p>\n<p>Not having a CUDA GPU, I plan to see how far I can get with just image features and maybe a simpler net.</p>\n<p>@kartal, 1.09 is pretty good, any features you suggest?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62674,
      "author_name": "blackeagle",
      "author_url": "",
      "post_date": "01/13/2015 04:09:13",
      "content": "<p>I have close to 10K features. These are mostly patch-based and includes some convolution features as well. So all I am doing is obtaining &nbsp;patches of different scales and then pass them through k-means to obtain visual words. Then use SVM to classify images. I am not sure how far I will go by this technique but I have a feeling this technique can be as good as a coarsely-tuned CNN.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62693,
      "author_name": "msegala",
      "author_url": "",
      "post_date": "01/13/2015 14:36:40",
      "content": "<p>@kartal: Can you give some references to the features you are using. I am using deep belief networks and want to experiment with some feature engineering as well.&nbsp;Thanks!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 62694,
      "author_name": "blackeagle",
      "author_url": "",
      "post_date": "01/13/2015 14:43:41",
      "content": "<p>Hi Mike,</p>\n<p>This paper should help.&nbsp;</p>\n<p>http://web.eecs.umich.edu/~honglak/aistats11-AnalysisSingleLayerUnsupervisedFeatureLearning.pdf</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "62637": "",
    "62661": "",
    "62673": "",
    "62674": "",
    "62693": "",
    "62694": ""
  },
  "source": "meta"
}