{
  "id": 122258,
  "title": "Estimates for best possible solutions error",
  "url": "/competitions/deepfake-detection-challenge/discussion/122258",
  "author_name": "Eugene Dorokhin",
  "post_date": "2019-12-18T21:23:30.118000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>For a contest I participated in a while back, I had some fun and tried to reason a lower bound on how well the best possible solution could do (<a href=\"https://www.kaggle.com/eeegnu/analysis-of-hypothetical-optimal-score\">https://www.kaggle.com/eeegnu/analysis-of-hypothetical-optimal-score</a> ). Log Loss was used there as well, but there was room for error in the actual data there, whereas here I imagine there isn't (i.e. they generate the fakes themselves.) I'm not so familiar with the literature on deep fakes, so I'm wondering what the threshold score would be between current best known methods, and something unexpected, with this score metric. Any thoughts?</p>",
  "messages": [
    {
      "id": 698148,
      "postDate": "2019-12-18T21:23:30.120Z",
      "content": "<p>For a contest I participated in a while back, I had some fun and tried to reason a lower bound on how well the best possible solution could do (<a href=\"https://www.kaggle.com/eeegnu/analysis-of-hypothetical-optimal-score\">https://www.kaggle.com/eeegnu/analysis-of-hypothetical-optimal-score</a> ). Log Loss was used there as well, but there was room for error in the actual data there, whereas here I imagine there isn't (i.e. they generate the fakes themselves.) I'm not so familiar with the literature on deep fakes, so I'm wondering what the threshold score would be between current best known methods, and something unexpected, with this score metric. Any thoughts?</p>",
      "rawMarkdown": "For a contest I participated in a while back, I had some fun and tried to reason a lower bound on how well the best possible solution could do (https://www.kaggle.com/eeegnu/analysis-of-hypothetical-optimal-score ). Log Loss was used there as well, but there was room for error in the actual data there, whereas here I imagine there isn't (i.e. they generate the fakes themselves.) I'm not so familiar with the literature on deep fakes, so I'm wondering what the threshold score would be between current best known methods, and something unexpected, with this score metric. Any thoughts?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "698148": "For a contest I participated in a while back, I had some fun and tried to reason a lower bound on how well the best possible solution could do (https://www.kaggle.com/eeegnu/analysis-of-hypothetical-optimal-score ). Log Loss was used there as well, but there was room for error in the actual data there, whereas here I imagine there isn't (i.e. they generate the fakes themselves.) I'm not so familiar with the literature on deep fakes, so I'm wondering what the threshold score would be between current best known methods, and something unexpected, with this score metric. Any thoughts?"
  }
}