{
  "id": 47408,
  "title": "A though experiment ...",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/47408",
  "author_name": "hengck23",
  "post_date": "2018-01-13T14:11:05.462000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<ol>\n<li><p>Say, one estimates that he is worse than the top rank results by K LB test samples.</p></li>\n<li><p>Using current predicted score of the LB test samples, one can select a*K LB test samples that are now most likely to be wrong and most easily corrected. (e.g. LB samples with confidence score  e.g. 0.3 to 0.4 and is of class \"xx\",\"yy\" ... or \"zz\").</p></li>\n<li><p>From the predicted probability, we can assign 3 possible labels to each of  the  a*K LB test samples.</p></li>\n<li><p>We can come up with 3^(a*K) possible label combinations and our aim to the guess which is the most correct one.</p></li>\n<li><p>You can do it using some optimization to maximize the label probability  or using brute force search. Here is the idea using  brute force search:</p>\n\n<ul><li><p>for each 3^(a*K) possible label combinations:</p>\n\n<ul><li><p>set train set = a*K LB test samples with the current combination labels and validation set = train samples with real ground truth labels</p></li>\n<li><p>train a model and measure validation error</p></li></ul></li>\n<li><p>the combination label with the lowest validation error is the predicted true label of the a*K LB test samples.</p></li></ul></li>\n</ol>\n\n<p>Does this work, assume that we have infinite computation power for brute force search ?</p>",
  "messages": [
    {
      "id": 268140,
      "postDate": "2018-01-13T14:11:05.463Z",
      "content": "<ol>\n<li><p>Say, one estimates that he is worse than the top rank results by K LB test samples.</p></li>\n<li><p>Using current predicted score of the LB test samples, one can select a*K LB test samples that are now most likely to be wrong and most easily corrected. (e.g. LB samples with confidence score  e.g. 0.3 to 0.4 and is of class \"xx\",\"yy\" ... or \"zz\").</p></li>\n<li><p>From the predicted probability, we can assign 3 possible labels to each of  the  a*K LB test samples.</p></li>\n<li><p>We can come up with 3^(a*K) possible label combinations and our aim to the guess which is the most correct one.</p></li>\n<li><p>You can do it using some optimization to maximize the label probability  or using brute force search. Here is the idea using  brute force search:</p>\n\n<ul><li><p>for each 3^(a*K) possible label combinations:</p>\n\n<ul><li><p>set train set = a*K LB test samples with the current combination labels and validation set = train samples with real ground truth labels</p></li>\n<li><p>train a model and measure validation error</p></li></ul></li>\n<li><p>the combination label with the lowest validation error is the predicted true label of the a*K LB test samples.</p></li></ul></li>\n</ol>\n\n<p>Does this work, assume that we have infinite computation power for brute force search ?</p>",
      "rawMarkdown": " 1.  Say, one estimates that he is worse than the top rank results by K LB test samples.\n\n 2. Using current predicted score of the LB test samples, one can select a*K LB test samples that are now most likely to be wrong and most easily corrected. (e.g. LB samples with confidence score  e.g. 0.3 to 0.4 and is of class \"xx\",\"yy\" ... or \"zz\").\n\n 3. From the predicted probability, we can assign 3 possible labels to each of  the  a*K LB test samples.\n\n 4. We can come up with 3^(a*K) possible label combinations and our aim to the guess which is the most correct one.\n\n 5. You can do it using some optimization to maximize the label probability  or using brute force search. Here is the idea using  brute force search:\n\n  - for each 3^(a*K) possible label combinations:\n\n        - set train set = a*K LB test samples with the current combination labels and validation set = train samples with real ground truth labels\n\n        - train a model and measure validation error\n\n - the combination label with the lowest validation error is the predicted true label of the a*K LB test samples.\n\n\nDoes this work, assume that we have infinite computation power for brute force search ?\n  \n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "268140": " 1.  Say, one estimates that he is worse than the top rank results by K LB test samples.\n\n 2. Using current predicted score of the LB test samples, one can select a*K LB test samples that are now most likely to be wrong and most easily corrected. (e.g. LB samples with confidence score  e.g. 0.3 to 0.4 and is of class \"xx\",\"yy\" ... or \"zz\").\n\n 3. From the predicted probability, we can assign 3 possible labels to each of  the  a*K LB test samples.\n\n 4. We can come up with 3^(a*K) possible label combinations and our aim to the guess which is the most correct one.\n\n 5. You can do it using some optimization to maximize the label probability  or using brute force search. Here is the idea using  brute force search:\n\n  - for each 3^(a*K) possible label combinations:\n\n        - set train set = a*K LB test samples with the current combination labels and validation set = train samples with real ground truth labels\n\n        - train a model and measure validation error\n\n - the combination label with the lowest validation error is the predicted true label of the a*K LB test samples.\n\n\nDoes this work, assume that we have infinite computation power for brute force search ?\n  \n"
  }
}