{
  "id": 48046,
  "title": "Anyone else use entropy of predictions to detect silence?",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/48046",
  "author_name": "Liam",
  "post_date": "2018-01-22T21:42:10.748000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I found that calculating the entropy H(p) = -\\sum_i (p_i log p_i) of all the test set predictions, and setting those with H(p) &gt; 2.3 (after some tweaking) to be labeled as silence was quite effective at predicting silence (based on LB movements and validation set). Curious if anyone else used this trick?</p>",
  "messages": [
    {
      "id": 272356,
      "postDate": "2018-01-22T21:42:10.750Z",
      "content": "<p>I found that calculating the entropy H(p) = -\\sum_i (p_i log p_i) of all the test set predictions, and setting those with H(p) &gt; 2.3 (after some tweaking) to be labeled as silence was quite effective at predicting silence (based on LB movements and validation set). Curious if anyone else used this trick?</p>",
      "rawMarkdown": "I found that calculating the entropy H(p) = -\\sum_i (p_i log p_i) of all the test set predictions, and setting those with H(p) &gt; 2.3 (after some tweaking) to be labeled as silence was quite effective at predicting silence (based on LB movements and validation set). Curious if anyone else used this trick?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "272356": "I found that calculating the entropy H(p) = -\\sum_i (p_i log p_i) of all the test set predictions, and setting those with H(p) &gt; 2.3 (after some tweaking) to be labeled as silence was quite effective at predicting silence (based on LB movements and validation set). Curious if anyone else used this trick?"
  }
}