{
  "id": 111549,
  "title": "difference between 1 and 0",
  "url": "/competitions/Kannada-MNIST/discussion/111549",
  "author_name": "",
  "post_date": "2019-10-06T20:06:53.904223600Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>As I see from my confusion matrix, most of errors comes from mixing 1 and 0, which is not surprising  - they differs only by short line break at the bottom! What is the best technic to train model again this problem?</p>",
  "messages": [
    {
      "id": "642917",
      "postDate": "10/06/2019 20:06:53",
      "content": "<p>As I see from my confusion matrix, most of errors comes from mixing 1 and 0, which is not surprising  - they differs only by short line break at the bottom! What is the best technic to train model again this problem?</p>",
      "rawMarkdown": "As I see from my confusion matrix, most of errors comes from mixing 1 and 0, which is not surprising  - they differs only by short line break at the bottom! What is the best technic to train model again this problem?",
      "votes": null
    },
    {
      "id": "650340",
      "postDate": "10/16/2019 09:35:30",
      "content": "<p>Hi Serge, we currently have the same score on the leaderboard. \nI have not yet looked at my confussion matrix and I will do that next week when I get new GPU time. </p>\n\n<p>If I have this issue I would probably create a separate classifier to reclassify(1 and 0), and have a pretty shallow network that is more or less a circle detector. </p>\n\n<p>So for everything that is classified as either 1 or 0 I would reclassify them again and update my prediction with the new values!</p>\n\n<p>Thanks for this insight!</p>",
      "rawMarkdown": "Hi Serge, we currently have the same score on the leaderboard. \nI have not yet looked at my confussion matrix and I will do that next week when I get new GPU time. \n\nIf I have this issue I would probably create a separate classifier to reclassify(1 and 0), and have a pretty shallow network that is more or less a circle detector. \n\nSo for everything that is classified as either 1 or 0 I would reclassify them again and update my prediction with the new values!\n\nThanks for this insight!",
      "votes": null
    },
    {
      "id": "660681",
      "postDate": "10/29/2019 13:18:41",
      "content": "<p>Hi, i checked my confussion matrix and I don't have this effect at all, What type of network are you using, I am using a Resnet inspired architecture.</p>",
      "rawMarkdown": "Hi, i checked my confussion matrix and I don't have this effect at all, What type of network are you using, I am using a Resnet inspired architecture.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 650340,
      "author_name": "rincewind007",
      "author_url": "",
      "post_date": "10/16/2019 09:35:30",
      "content": "<p>Hi Serge, we currently have the same score on the leaderboard. \nI have not yet looked at my confussion matrix and I will do that next week when I get new GPU time. </p>\n\n<p>If I have this issue I would probably create a separate classifier to reclassify(1 and 0), and have a pretty shallow network that is more or less a circle detector. </p>\n\n<p>So for everything that is classified as either 1 or 0 I would reclassify them again and update my prediction with the new values!</p>\n\n<p>Thanks for this insight!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 660681,
      "author_name": "rincewind007",
      "author_url": "",
      "post_date": "10/29/2019 13:18:41",
      "content": "<p>Hi, i checked my confussion matrix and I don't have this effect at all, What type of network are you using, I am using a Resnet inspired architecture.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "642917": "As I see from my confusion matrix, most of errors comes from mixing 1 and 0, which is not surprising  - they differs only by short line break at the bottom! What is the best technic to train model again this problem?",
    "650340": "Hi Serge, we currently have the same score on the leaderboard. \nI have not yet looked at my confussion matrix and I will do that next week when I get new GPU time. \n\nIf I have this issue I would probably create a separate classifier to reclassify(1 and 0), and have a pretty shallow network that is more or less a circle detector. \n\nSo for everything that is classified as either 1 or 0 I would reclassify them again and update my prediction with the new values!\n\nThanks for this insight!",
    "660681": "Hi, i checked my confussion matrix and I don't have this effect at all, What type of network are you using, I am using a Resnet inspired architecture."
  },
  "source": "meta"
}