{
  "id": 11524,
  "title": "Any luck with ZCA+Whitening?",
  "url": "/competitions/datasciencebowl/discussion/11524",
  "author_name": "",
  "post_date": "2015-01-15T14:04:24.423Z",
  "votes": 1,
  "comment_count": 1,
  "views": 1526,
  "content": "<p>Hi everyone,</p>\n<p>I was playing around with different preprocessing steps and I noticed that if I first normalize all my features to [0-1] and then apply ZCA + whitening I get worse results by about 10%. Has anyone had any luck with this or can suggest a better approach. My current setup is to just scale the features to the range of [0-1] and this combined with my learning algo lands me a score of 1.3.</p>\n<p>Thanks!</p>",
  "messages": [
    {
      "id": "62750",
      "postDate": "01/15/2015 14:04:24",
      "content": "<p>Hi everyone,</p>\n<p>I was playing around with different preprocessing steps and I noticed that if I first normalize all my features to [0-1] and then apply ZCA + whitening I get worse results by about 10%. Has anyone had any luck with this or can suggest a better approach. My current setup is to just scale the features to the range of [0-1] and this combined with my learning algo lands me a score of 1.3.</p>\n<p>Thanks!</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "62789",
      "postDate": "01/16/2015 12:48:57",
      "content": "<p>Actually I was trying to replicate ZCA whitening + k-means for patches (as described in <a href=\"http://web.eecs.umich.edu/~honglak/aistats11-AnalysisSingleLayerUnsupervisedFeatureLearning.pdf\">pdf</a>) but I also got worse results then by simple scaling + k-means. Still checking it ...</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 62789,
      "author_name": "toma78",
      "author_url": "",
      "post_date": "01/16/2015 12:48:57",
      "content": "<p>Actually I was trying to replicate ZCA whitening + k-means for patches (as described in <a href=\"http://web.eecs.umich.edu/~honglak/aistats11-AnalysisSingleLayerUnsupervisedFeatureLearning.pdf\">pdf</a>) but I also got worse results then by simple scaling + k-means. Still checking it ...</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "62750": "",
    "62789": ""
  },
  "source": "meta"
}