{
  "id": 15043,
  "title": "Deconv net to produce image for rebalancing dataset",
  "url": "/competitions/diabetic-retinopathy-detection/discussion/15043",
  "author_name": "",
  "post_date": "2015-07-04T15:03:34.417Z",
  "votes": null,
  "comment_count": 1,
  "views": 794,
  "content": "<p>Whether does anyone attempt to use deconv net (by&nbsp;matthew zeiler) in high-level conv features' KNN space to produce new images for rebalancing given dataset or online update trained model?</p>",
  "messages": [
    {
      "id": "83416",
      "postDate": "07/04/2015 15:03:34",
      "content": "<p>Whether does anyone attempt to use deconv net (by&nbsp;matthew zeiler) in high-level conv features' KNN space to produce new images for rebalancing given dataset or online update trained model?</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "83418",
      "postDate": "07/04/2015 15:37:27",
      "content": "<p>I think it will be very effective, but it shouldn't produce much more.</p>\n<p>When some data are misclassified, these data's high-level features' K nearest neighbors have a more important role to change model's decision boundary. Whatever images deconvolutional net produces from high-level features' KNN space are all positive and complementary&nbsp;to model update. But when it reaches some point, we may move to some wrong decision boundary if we produce much more deconv images. If a bit more images, test data can regularize it.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 83418,
      "author_name": "huyaoquan",
      "author_url": "",
      "post_date": "07/04/2015 15:37:27",
      "content": "<p>I think it will be very effective, but it shouldn't produce much more.</p>\n<p>When some data are misclassified, these data's high-level features' K nearest neighbors have a more important role to change model's decision boundary. Whatever images deconvolutional net produces from high-level features' KNN space are all positive and complementary&nbsp;to model update. But when it reaches some point, we may move to some wrong decision boundary if we produce much more deconv images. If a bit more images, test data can regularize it.&nbsp;</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "83416": "",
    "83418": ""
  },
  "source": "meta"
}