{
  "id": 6574,
  "title": "benchmark with no preprocessing",
  "url": "/competitions/cifar-10/discussion/6574",
  "author_name": "",
  "post_date": "2013-12-11T07:57:49.720Z",
  "votes": null,
  "comment_count": 5,
  "views": 4155,
  "content": "<p>Hi,</p>\n<p>Can someone share cross validation accuracy for any algorithm with no preprocessing applied?</p>\n<p>I see that it never exceeds 12%</p>\n<p>&nbsp;</p>\n<p>&nbsp;</p>",
  "messages": [
    {
      "id": "36051",
      "postDate": "12/11/2013 07:57:49",
      "content": "<p>Hi,</p>\n<p>Can someone share cross validation accuracy for any algorithm with no preprocessing applied?</p>\n<p>I see that it never exceeds 12%</p>\n<p>&nbsp;</p>\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "36062",
      "postDate": "12/11/2013 14:02:35",
      "content": "<p>[quote=Vignesh;36051]</p>\n<p>Hi,</p>\n<p>Can someone share cross validation accuracy for any algorithm with no preprocessing applied?</p>\n<p>I see that it never exceeds 12%</p>\n<p>[/quote]</p>\n<p>That's the kind of value I'd expect if you treated the raw pixel values as the features, and ran any kind of classifier that ignored location of the pixels.</p>\n<p>&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "36080",
      "postDate": "12/11/2013 17:55:29",
      "content": "<p>Even if you just unroll the pixels into a single dimensional array, isn't some location information still maintained?&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "36082",
      "postDate": "12/11/2013 18:06:06",
      "content": "<p>[quote=Vignesh;36080]</p>\n<p>Even if you just unroll the pixels into a single dimensional array, isn't some location information still maintained?&nbsp;</p>\n<p>[/quote]</p>\n<p>Only if the classification method you then feed it into takes any notice of the index position. Many will not.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "36360",
      "postDate": "12/19/2013 08:28:31",
      "content": "<p>The other day, I was glancing over the paper linked in the competitions page and if I remember correctly, somewhere its mentioned that raw pixels with logistic regressions gives around 37% accuracy.&nbsp;</p>\n<p>To add more in the same context, I just converted the rgb to gray pixels and logistic regression is giving me around 39% accuracy.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "502379",
      "postDate": "03/28/2019 14:15:59",
      "content": "<p>I trained a DenseNet with SGD optimizer (12 filters, 3X3 kernel size, 0.2 dropout), without any image pre-processing and it gave me around ~70% validation accuracy</p>",
      "rawMarkdown": "I trained a DenseNet with SGD optimizer (12 filters, 3X3 kernel size, 0.2 dropout), without any image pre-processing and it gave me around ~70% validation accuracy",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 36062,
      "author_name": "slobo777",
      "author_url": "",
      "post_date": "12/11/2013 14:02:35",
      "content": "<p>[quote=Vignesh;36051]</p>\n<p>Hi,</p>\n<p>Can someone share cross validation accuracy for any algorithm with no preprocessing applied?</p>\n<p>I see that it never exceeds 12%</p>\n<p>[/quote]</p>\n<p>That's the kind of value I'd expect if you treated the raw pixel values as the features, and ran any kind of classifier that ignored location of the pixels.</p>\n<p>&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 36080,
      "author_name": "vignesh0",
      "author_url": "",
      "post_date": "12/11/2013 17:55:29",
      "content": "<p>Even if you just unroll the pixels into a single dimensional array, isn't some location information still maintained?&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 36082,
      "author_name": "slobo777",
      "author_url": "",
      "post_date": "12/11/2013 18:06:06",
      "content": "<p>[quote=Vignesh;36080]</p>\n<p>Even if you just unroll the pixels into a single dimensional array, isn't some location information still maintained?&nbsp;</p>\n<p>[/quote]</p>\n<p>Only if the classification method you then feed it into takes any notice of the index position. Many will not.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 36360,
      "author_name": "pythonomic",
      "author_url": "",
      "post_date": "12/19/2013 08:28:31",
      "content": "<p>The other day, I was glancing over the paper linked in the competitions page and if I remember correctly, somewhere its mentioned that raw pixels with logistic regressions gives around 37% accuracy.&nbsp;</p>\n<p>To add more in the same context, I just converted the rgb to gray pixels and logistic regression is giving me around 39% accuracy.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 502379,
      "author_name": "techytushar",
      "author_url": "",
      "post_date": "03/28/2019 14:15:59",
      "content": "<p>I trained a DenseNet with SGD optimizer (12 filters, 3X3 kernel size, 0.2 dropout), without any image pre-processing and it gave me around ~70% validation accuracy</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "36051": "",
    "36062": "",
    "36080": "",
    "36082": "",
    "36360": "",
    "502379": "I trained a DenseNet with SGD optimizer (12 filters, 3X3 kernel size, 0.2 dropout), without any image pre-processing and it gave me around ~70% validation accuracy"
  },
  "source": "meta"
}