{
  "id": 72769,
  "title": "Smooth Loss Functions for Deep Top-k Classification",
  "url": "/competitions/quickdraw-doodle-recognition/discussion/72769",
  "author_name": "William Horton",
  "post_date": "2018-11-27T03:44:40.842000",
  "votes": 3,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Came across this paper today and thought it could be useful for this competition. It formulates smooth loss functions that let you optimize top-k classification more directly, as well as computationally efficient implementations of these functions.</p>\n\n<p>Arxiv: <a href=\"https://arxiv.org/abs/1802.07595\">https://arxiv.org/abs/1802.07595</a>\nCode (in Pytorch): <a href=\"https://github.com/oval-group/smooth-topk\">https://github.com/oval-group/smooth-topk</a></p>\n\n<p>Thanks to @mpekalski for posting it in the KaggleNoobs Slack!</p>",
  "messages": [
    {
      "id": 428301,
      "postDate": "2018-11-27T03:44:40.843Z",
      "content": "<p>Came across this paper today and thought it could be useful for this competition. It formulates smooth loss functions that let you optimize top-k classification more directly, as well as computationally efficient implementations of these functions.</p>\n\n<p>Arxiv: <a href=\"https://arxiv.org/abs/1802.07595\">https://arxiv.org/abs/1802.07595</a>\nCode (in Pytorch): <a href=\"https://github.com/oval-group/smooth-topk\">https://github.com/oval-group/smooth-topk</a></p>\n\n<p>Thanks to @mpekalski for posting it in the KaggleNoobs Slack!</p>",
      "rawMarkdown": "Came across this paper today and thought it could be useful for this competition. It formulates smooth loss functions that let you optimize top-k classification more directly, as well as computationally efficient implementations of these functions.\n\nArxiv: https://arxiv.org/abs/1802.07595\nCode (in Pytorch): https://github.com/oval-group/smooth-topk\n\nThanks to @mpekalski for posting it in the KaggleNoobs Slack!",
      "votes": 3
    },
    {
      "id": 428351,
      "postDate": "2018-11-27T06:04:08.023Z",
      "content": "<p>Thanks William! It is always nice to learn a new thing.</p>",
      "rawMarkdown": "Thanks William! It is always nice to learn a new thing."
    },
    {
      "id": 428326,
      "postDate": "2018-11-27T04:57:29.887Z",
      "content": "<p>In the paper it shows good results for limited datasets, so this could help especially if you’re resource constrained and can only use  1% or 5% of the training data</p>",
      "rawMarkdown": "In the paper it shows good results for limited datasets, so this could help especially if you’re resource constrained and can only use  1% or 5% of the training data"
    }
  ],
  "comments": [
    {
      "id": 428351,
      "author_name": "Neuron Engineer",
      "author_url": "",
      "post_date": "2018-11-27T06:04:08.023000",
      "content": "<p>Thanks William! It is always nice to learn a new thing.</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 428326,
      "author_name": "William Horton",
      "author_url": "",
      "post_date": "2018-11-27T04:57:29.887000",
      "content": "<p>In the paper it shows good results for limited datasets, so this could help especially if you’re resource constrained and can only use  1% or 5% of the training data</p>",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "428301": "Came across this paper today and thought it could be useful for this competition. It formulates smooth loss functions that let you optimize top-k classification more directly, as well as computationally efficient implementations of these functions.\n\nArxiv: https://arxiv.org/abs/1802.07595\nCode (in Pytorch): https://github.com/oval-group/smooth-topk\n\nThanks to @mpekalski for posting it in the KaggleNoobs Slack!",
    "428351": "Thanks William! It is always nice to learn a new thing.",
    "428326": "In the paper it shows good results for limited datasets, so this could help especially if you’re resource constrained and can only use  1% or 5% of the training data"
  }
}