{
  "id": 18540,
  "title": "Cumulating Predictions - which is better? (Softmax or Hard Target)",
  "url": "/competitions/second-annual-data-science-bowl/discussion/18540",
  "author_name": "DavidGbodiOdaibo",
  "post_date": "2016-01-23T17:56:27.380000",
  "votes": 0,
  "comment_count": 0,
  "views": 538,
  "content": "<p>I achieve high classification accuracy during training with a Softmax output on my neural net, but this is not translating into a low CRPS score once all the samples are averaged. My intuition is that Softmax distributes the probability across all the output volumes and the confidence in the correct volume prediction is not high; I would like to know what others with low CRPS scores are doing :) . Averaging Softmax outputs or putting all the mass on a single prediction?</p>",
  "messages": [
    {
      "id": 105486,
      "postDate": "2016-01-23T17:56:27.380Z",
      "content": "<p>I achieve high classification accuracy during training with a Softmax output on my neural net, but this is not translating into a low CRPS score once all the samples are averaged. My intuition is that Softmax distributes the probability across all the output volumes and the confidence in the correct volume prediction is not high; I would like to know what others with low CRPS scores are doing :) . Averaging Softmax outputs or putting all the mass on a single prediction?</p>",
      "rawMarkdown": "I achieve high classification accuracy during training with a Softmax output on my neural net, but this is not translating into a low CRPS score once all the samples are averaged. My intuition is that Softmax distributes the probability across all the output volumes and the confidence in the correct volume prediction is not high; I would like to know what others with low CRPS scores are doing :) . Averaging Softmax outputs or putting all the mass on a single prediction?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "105486": "I achieve high classification accuracy during training with a Softmax output on my neural net, but this is not translating into a low CRPS score once all the samples are averaged. My intuition is that Softmax distributes the probability across all the output volumes and the confidence in the correct volume prediction is not high; I would like to know what others with low CRPS scores are doing :) . Averaging Softmax outputs or putting all the mass on a single prediction?"
  }
}