{
  "id": 48312,
  "title": "Use of  comprehensive   approaches not black  box",
  "url": "/competitions/data-science-bowl-2018/discussion/48312",
  "author_name": "",
  "post_date": "2018-01-26T03:14:12.074429500Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Hello ,</p>\n\n<p>I am currious   if it is better to use  somekind of comprehensive approaches  regarding  or somekind of  domain knowledege then just  run models  like a  black box.</p>\n\n<p>If  you    think better , what is  the purpose to  do a  good prediction if  you dont know  why it   was good ?\nIt  is  like  you go  to a    ,, vodo '' expert assuming it exists  ,    if  it will tell you are goin to   win lotery   or   going to   be dead in   the next month ,   you end  up  home  getting drunk  because you just  have a prediction without   any explanation</p>",
  "messages": [
    {
      "id": "274200",
      "postDate": "01/26/2018 03:14:12",
      "content": "<p>Hello ,</p>\n\n<p>I am currious   if it is better to use  somekind of comprehensive approaches  regarding  or somekind of  domain knowledege then just  run models  like a  black box.</p>\n\n<p>If  you    think better , what is  the purpose to  do a  good prediction if  you dont know  why it   was good ?\nIt  is  like  you go  to a    ,, vodo '' expert assuming it exists  ,    if  it will tell you are goin to   win lotery   or   going to   be dead in   the next month ,   you end  up  home  getting drunk  because you just  have a prediction without   any explanation</p>",
      "rawMarkdown": "Hello ,\n\nI am currious   if it is better to use  somekind of comprehensive approaches  regarding  or somekind of  domain knowledege then just  run models  like a  black box.\n\nIf  you    think better , what is  the purpose to  do a  good prediction if  you dont know  why it   was good ?\nIt  is  like  you go  to a    ,, vodo '' expert assuming it exists  ,    if  it will tell you are goin to   win lotery   or   going to   be dead in   the next month ,   you end  up  home  getting drunk  because you just  have a prediction without   any explanation",
      "votes": null
    },
    {
      "id": "274976",
      "postDate": "01/27/2018 23:30:46",
      "content": "<p>Interpretation would be useful. In fact, I think many people believe that the features created for this competition would be useful outside of this problem and could drive some useful biological discovery. However, the problem itself doesn’t necessitate interpretation since image segmentation and object detection isn’t directly responsible for any discovery (i.e. “black boxes” are totally fine). The output of a model (or algorithm) for this problem only helps biologists identify a nucleus. The biologist would be expected to perform some work to extract useful features from the detected pixels. </p>",
      "rawMarkdown": "Interpretation would be useful. In fact, I think many people believe that the features created for this competition would be useful outside of this problem and could drive some useful biological discovery. However, the problem itself doesn’t necessitate interpretation since image segmentation and object detection isn’t directly responsible for any discovery (i.e. “black boxes” are totally fine). The output of a model (or algorithm) for this problem only helps biologists identify a nucleus. The biologist would be expected to perform some work to extract useful features from the detected pixels.",
      "votes": null
    },
    {
      "id": "275168",
      "postDate": "01/28/2018 12:27:19",
      "content": "<p>I'd add that solving the segmentation problem that is presented is by far the main goal. I personally would be surprised if any biological insight came from solving the challenge even if models were interpretable. After all, the challenge is to match human accuracy so we already know what we are looking for (even if experts disagree a bit!)  If we had some alternate source of ground truth other than human judgment, THEN interpretability would be a lot more interesting.</p>\n\n<p>So I would advise to solve the problem as best you can, don't worry at all about whether the solution offers interpretability.</p>",
      "rawMarkdown": "I'd add that solving the segmentation problem that is presented is by far the main goal. I personally would be surprised if any biological insight came from solving the challenge even if models were interpretable. After all, the challenge is to match human accuracy so we already know what we are looking for (even if experts disagree a bit!)  If we had some alternate source of ground truth other than human judgment, THEN interpretability would be a lot more interesting.\n\nSo I would advise to solve the problem as best you can, don't worry at all about whether the solution offers interpretability.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 274976,
      "author_name": "agoodman",
      "author_url": "",
      "post_date": "01/27/2018 23:30:46",
      "content": "<p>Interpretation would be useful. In fact, I think many people believe that the features created for this competition would be useful outside of this problem and could drive some useful biological discovery. However, the problem itself doesn’t necessitate interpretation since image segmentation and object detection isn’t directly responsible for any discovery (i.e. “black boxes” are totally fine). The output of a model (or algorithm) for this problem only helps biologists identify a nucleus. The biologist would be expected to perform some work to extract useful features from the detected pixels. </p>",
      "votes": null,
      "replies": [
        {
          "id": 275168,
          "author_name": "drannecarpenter",
          "author_url": "",
          "post_date": "01/28/2018 12:27:19",
          "content": "<p>I'd add that solving the segmentation problem that is presented is by far the main goal. I personally would be surprised if any biological insight came from solving the challenge even if models were interpretable. After all, the challenge is to match human accuracy so we already know what we are looking for (even if experts disagree a bit!)  If we had some alternate source of ground truth other than human judgment, THEN interpretability would be a lot more interesting.</p>\n\n<p>So I would advise to solve the problem as best you can, don't worry at all about whether the solution offers interpretability.</p>",
          "votes": null,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "274200": "Hello ,\n\nI am currious   if it is better to use  somekind of comprehensive approaches  regarding  or somekind of  domain knowledege then just  run models  like a  black box.\n\nIf  you    think better , what is  the purpose to  do a  good prediction if  you dont know  why it   was good ?\nIt  is  like  you go  to a    ,, vodo '' expert assuming it exists  ,    if  it will tell you are goin to   win lotery   or   going to   be dead in   the next month ,   you end  up  home  getting drunk  because you just  have a prediction without   any explanation",
    "274976": "Interpretation would be useful. In fact, I think many people believe that the features created for this competition would be useful outside of this problem and could drive some useful biological discovery. However, the problem itself doesn’t necessitate interpretation since image segmentation and object detection isn’t directly responsible for any discovery (i.e. “black boxes” are totally fine). The output of a model (or algorithm) for this problem only helps biologists identify a nucleus. The biologist would be expected to perform some work to extract useful features from the detected pixels.",
    "275168": "I'd add that solving the segmentation problem that is presented is by far the main goal. I personally would be surprised if any biological insight came from solving the challenge even if models were interpretable. After all, the challenge is to match human accuracy so we already know what we are looking for (even if experts disagree a bit!)  If we had some alternate source of ground truth other than human judgment, THEN interpretability would be a lot more interesting.\n\nSo I would advise to solve the problem as best you can, don't worry at all about whether the solution offers interpretability."
  },
  "source": "meta"
}