{
  "id": 47947,
  "title": "Why the best score < 0.92?",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/47947",
  "author_name": "",
  "post_date": "2018-01-21T07:17:38.702430300Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Why do you think the best score was lower than 0.92? noisy ground truth label? Under optimal solutions?</p>",
  "messages": [
    {
      "id": "271702",
      "postDate": "01/21/2018 07:17:38",
      "content": "<p>Why do you think the best score was lower than 0.92? noisy ground truth label? Under optimal solutions?</p>",
      "rawMarkdown": "Why do you think the best score was lower than 0.92? noisy ground truth label? Under optimal solutions?",
      "votes": null
    },
    {
      "id": "273676",
      "postDate": "01/25/2018 01:23:16",
      "content": "<p>I thought about this a lot during the competition. Many people focused on the unknown unknowns, but there simply weren't enough of those as a percent of the LB data to justify the low score. </p>\n\n<p>My working hypothesis is that the competition sponsors wanted to make sure the competition was challenging (fair enough) and perhaps only included samples in the LB that were misclassified by an internal model. </p>\n\n<p>Or perhaps different regional dialects were intentionally placed in the test set? </p>\n\n<p>Or different microphone conditions?</p>",
      "rawMarkdown": "I thought about this a lot during the competition. Many people focused on the unknown unknowns, but there simply weren't enough of those as a percent of the LB data to justify the low score. \n\nMy working hypothesis is that the competition sponsors wanted to make sure the competition was challenging (fair enough) and perhaps only included samples in the LB that were misclassified by an internal model. \n\nOr perhaps different regional dialects were intentionally placed in the test set? \n\nOr different microphone conditions?",
      "votes": null
    },
    {
      "id": "274867",
      "postDate": "01/27/2018 15:12:57",
      "content": "<p>Would be interesting to know human level performance. I checked some of the test data, and was not sure myself of the labels</p>",
      "rawMarkdown": "Would be interesting to know human level performance. I checked some of the test data, and was not sure myself of the labels",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 273676,
      "author_name": "omalleyt",
      "author_url": "",
      "post_date": "01/25/2018 01:23:16",
      "content": "<p>I thought about this a lot during the competition. Many people focused on the unknown unknowns, but there simply weren't enough of those as a percent of the LB data to justify the low score. </p>\n\n<p>My working hypothesis is that the competition sponsors wanted to make sure the competition was challenging (fair enough) and perhaps only included samples in the LB that were misclassified by an internal model. </p>\n\n<p>Or perhaps different regional dialects were intentionally placed in the test set? </p>\n\n<p>Or different microphone conditions?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 274867,
      "author_name": "christofhenkel",
      "author_url": "",
      "post_date": "01/27/2018 15:12:57",
      "content": "<p>Would be interesting to know human level performance. I checked some of the test data, and was not sure myself of the labels</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "271702": "Why do you think the best score was lower than 0.92? noisy ground truth label? Under optimal solutions?",
    "273676": "I thought about this a lot during the competition. Many people focused on the unknown unknowns, but there simply weren't enough of those as a percent of the LB data to justify the low score. \n\nMy working hypothesis is that the competition sponsors wanted to make sure the competition was challenging (fair enough) and perhaps only included samples in the LB that were misclassified by an internal model. \n\nOr perhaps different regional dialects were intentionally placed in the test set? \n\nOr different microphone conditions?",
    "274867": "Would be interesting to know human level performance. I checked some of the test data, and was not sure myself of the labels"
  },
  "source": "meta"
}