{
  "id": 533328,
  "title": "What is special about LB=0.54?",
  "url": "/competitions/rsna-2024-lumbar-spine-degenerative-classification/discussion/533328",
  "author_name": "",
  "post_date": "2024-09-10T17:07:18.943159200Z",
  "votes": 2,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I keep seeing all sorts of different 1-stage approaches landing around this mark.</p>\n<p>My guess is the imbalance for some labels like L1/L2 and L5/S1 mean one can only do so much when there is not enough data to demarcate moderate and severe classes properly. And off-the-shelf models are powerful enough to do as much as possible with the data before converging at the same ceiling.</p>\n<p>But I am curious about the mathematics behind this. If it is due to the data constraints like I am guessing, and the models are all hitting this same ceiling, then there should be a way to calculate it somehow.</p>",
  "messages": [
    {
      "id": "2985479",
      "postDate": "09/10/2024 17:07:18",
      "content": "<p>I keep seeing all sorts of different 1-stage approaches landing around this mark.</p>\n<p>My guess is the imbalance for some labels like L1/L2 and L5/S1 mean one can only do so much when there is not enough data to demarcate moderate and severe classes properly. And off-the-shelf models are powerful enough to do as much as possible with the data before converging at the same ceiling.</p>\n<p>But I am curious about the mathematics behind this. If it is due to the data constraints like I am guessing, and the models are all hitting this same ceiling, then there should be a way to calculate it somehow.</p>",
      "rawMarkdown": "I keep seeing all sorts of different 1-stage approaches landing around this mark.\n\nMy guess is the imbalance for some labels like L1/L2 and L5/S1 mean one can only do so much when there is not enough data to demarcate moderate and severe classes properly. And off-the-shelf models are powerful enough to do as much as possible with the data before converging at the same ceiling.\n\nBut I am curious about the mathematics behind this. If it is due to the data constraints like I am guessing, and the models are all hitting this same ceiling, then there should be a way to calculate it somehow.",
      "votes": null
    },
    {
      "id": "2986947",
      "postDate": "09/12/2024 07:54:28",
      "content": "<p>Have you ever used data whose labels contained NaN?</p>",
      "rawMarkdown": "Have you ever used data whose labels contained NaN?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2986947,
      "author_name": "xukongji",
      "author_url": "",
      "post_date": "09/12/2024 07:54:28",
      "content": "<p>Have you ever used data whose labels contained NaN?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2985479": "I keep seeing all sorts of different 1-stage approaches landing around this mark.\n\nMy guess is the imbalance for some labels like L1/L2 and L5/S1 mean one can only do so much when there is not enough data to demarcate moderate and severe classes properly. And off-the-shelf models are powerful enough to do as much as possible with the data before converging at the same ceiling.\n\nBut I am curious about the mathematics behind this. If it is due to the data constraints like I am guessing, and the models are all hitting this same ceiling, then there should be a way to calculate it somehow.",
    "2986947": "Have you ever used data whose labels contained NaN?"
  },
  "source": "meta"
}