{
  "id": 32185,
  "title": "Accuracy is dropping after image normalization",
  "url": "/competitions/digit-recognizer/discussion/32185",
  "author_name": "Rifayat Samee",
  "post_date": "2017-04-27T15:14:49.021000",
  "votes": 0,
  "comment_count": 3,
  "views": null,
  "content": "<p>Hi, i am using a convolutional neural network for \"digit-recognize\" problem. I have a score of 0.98614.But when i normalize the data from 0:255 to 0:1.0, my score is reduced to 0.95643. Any idea why is it happening? what am i doing wrong here?</p>\n\n<p>My kernel link: <a href=\"https://www.kaggle.com/sanzee/digit-recognizer/convolutional-neural-network/run/1111387\">link</a></p>",
  "messages": [
    {
      "id": 184810,
      "postDate": "2017-05-23T07:07:09.187Z",
      "content": "<p>Your network responds to you based on several parameters, including the activation functions. Some activation functions want normalized data (eg sigmoid function) while other activation functions perform better with unregulated data; I think your performance drop depends on this.</p>",
      "rawMarkdown": "Your network responds to you based on several parameters, including the activation functions. Some activation functions want normalized data (eg sigmoid function) while other activation functions perform better with unregulated data; I think your performance drop depends on this."
    },
    {
      "id": 178552,
      "postDate": "2017-04-28T06:57:56.170Z",
      "content": "<p>Thanks for the reply.\ni have normalized the test data at line 85\n    test_x = test_x/255.0</p>",
      "rawMarkdown": "Thanks for the reply.\ni have normalized the test data at line 85\n    test_x = test_x/255.0\n\n"
    },
    {
      "id": 178538,
      "postDate": "2017-04-28T06:00:38.547Z",
      "content": "<p>It does not look like the test set is normalized. This makes the network see much higher activations than it is used to.</p>",
      "rawMarkdown": "It does not look like the test set is normalized. This makes the network see much higher activations than it is used to."
    },
    {
      "id": 178389,
      "postDate": "2017-04-27T15:14:49.023Z",
      "content": "<p>Hi, i am using a convolutional neural network for \"digit-recognize\" problem. I have a score of 0.98614.But when i normalize the data from 0:255 to 0:1.0, my score is reduced to 0.95643. Any idea why is it happening? what am i doing wrong here?</p>\n\n<p>My kernel link: <a href=\"https://www.kaggle.com/sanzee/digit-recognizer/convolutional-neural-network/run/1111387\">link</a></p>",
      "rawMarkdown": "Hi, i am using a convolutional neural network for \"digit-recognize\" problem. I have a score of 0.98614.But when i normalize the data from 0:255 to 0:1.0, my score is reduced to 0.95643. Any idea why is it happening? what am i doing wrong here?\n\nMy kernel link: [link][1]\n\n\n  [1]: https://www.kaggle.com/sanzee/digit-recognizer/convolutional-neural-network/run/1111387"
    }
  ],
  "comments": [
    {
      "id": 184810,
      "author_name": "EnzoMaritati",
      "author_url": "",
      "post_date": "2017-05-23T07:07:09.187000",
      "content": "<p>Your network responds to you based on several parameters, including the activation functions. Some activation functions want normalized data (eg sigmoid function) while other activation functions perform better with unregulated data; I think your performance drop depends on this.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 178552,
      "author_name": "Rifayat Samee",
      "author_url": "",
      "post_date": "2017-04-28T06:57:56.170000",
      "content": "<p>Thanks for the reply.\ni have normalized the test data at line 85\n    test_x = test_x/255.0</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 178538,
      "author_name": "Kenneth Kihara",
      "author_url": "",
      "post_date": "2017-04-28T06:00:38.547000",
      "content": "<p>It does not look like the test set is normalized. This makes the network see much higher activations than it is used to.</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "184810": "Your network responds to you based on several parameters, including the activation functions. Some activation functions want normalized data (eg sigmoid function) while other activation functions perform better with unregulated data; I think your performance drop depends on this.",
    "178552": "Thanks for the reply.\ni have normalized the test data at line 85\n    test_x = test_x/255.0\n\n",
    "178538": "It does not look like the test set is normalized. This makes the network see much higher activations than it is used to.",
    "178389": "Hi, i am using a convolutional neural network for \"digit-recognize\" problem. I have a score of 0.98614.But when i normalize the data from 0:255 to 0:1.0, my score is reduced to 0.95643. Any idea why is it happening? what am i doing wrong here?\n\nMy kernel link: [link][1]\n\n\n  [1]: https://www.kaggle.com/sanzee/digit-recognizer/convolutional-neural-network/run/1111387"
  }
}