{
  "id": 49355,
  "title": "Has Anyone Figured Out What Were The CNNs Learning?",
  "url": "/competitions/sp-society-camera-model-identification/discussion/49355",
  "author_name": "",
  "post_date": "2018-02-09T19:12:53.292139500Z",
  "votes": 5,
  "comment_count": 3,
  "views": 0,
  "content": "<p>I was wondering if anyone has taken a look at some of the intermediate layers and seen what were they detecting? I may look into it myself, but wanted to see if someone else has not already investigated this.</p>",
  "messages": [
    {
      "id": "280351",
      "postDate": "02/09/2018 19:12:53",
      "content": "<p>I was wondering if anyone has taken a look at some of the intermediate layers and seen what were they detecting? I may look into it myself, but wanted to see if someone else has not already investigated this.</p>",
      "rawMarkdown": "I was wondering if anyone has taken a look at some of the intermediate layers and seen what were they detecting? I may look into it myself, but wanted to see if someone else has not already investigated this.",
      "votes": null
    },
    {
      "id": "280420",
      "postDate": "02/09/2018 23:41:20",
      "content": "<p>It would be a great post (or even a nice paper to publish) if you pursue in this direction. </p>",
      "rawMarkdown": "It would be a great post (or even a nice paper to publish) if you pursue in this direction.",
      "votes": null
    },
    {
      "id": "280425",
      "postDate": "02/09/2018 23:48:14",
      "content": "<p>It must have learned some features about lens, e.g., lens distortion, LED mechnisms(How RGB pixels are formed). This is something interesting. I also want to know the answer.</p>",
      "rawMarkdown": "It must have learned some features about lens, e.g., lens distortion, LED mechnisms(How RGB pixels are formed). This is something interesting. I also want to know the answer.",
      "votes": null
    },
    {
      "id": "280561",
      "postDate": "02/10/2018 09:09:22",
      "content": "<p>Very interesting to know. I find it a bit funny we finish this competition with close to ~0.99 accuracy and we don't know what features the architectures extracted. There's so much research going on re: explainability, either direct or with saliency maps (left for us to interpret). </p>\n\n<p>Here's a few references which may be interesting:</p>\n\n<p><a href=\"https://distill.pub/2017/feature-visualization/\">Feature visualization</a></p>\n\n<p><a href=\"http://netdissect.csail.mit.edu/\">Network dissection</a></p>",
      "rawMarkdown": "Very interesting to know. I find it a bit funny we finish this competition with close to ~0.99 accuracy and we don't know what features the architectures extracted. There's so much research going on re: explainability, either direct or with saliency maps (left for us to interpret). \n\nHere's a few references which may be interesting:\n\n[Feature visualization][1]\n\n[Network dissection][2]\n\n  [1]: https://distill.pub/2017/feature-visualization/\n  [2]: http://netdissect.csail.mit.edu/",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 280420,
      "author_name": "ceperaang",
      "author_url": "",
      "post_date": "02/09/2018 23:41:20",
      "content": "<p>It would be a great post (or even a nice paper to publish) if you pursue in this direction. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 280425,
      "author_name": "zhaoyangma",
      "author_url": "",
      "post_date": "02/09/2018 23:48:14",
      "content": "<p>It must have learned some features about lens, e.g., lens distortion, LED mechnisms(How RGB pixels are formed). This is something interesting. I also want to know the answer.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 280561,
      "author_name": "antorsae",
      "author_url": "",
      "post_date": "02/10/2018 09:09:22",
      "content": "<p>Very interesting to know. I find it a bit funny we finish this competition with close to ~0.99 accuracy and we don't know what features the architectures extracted. There's so much research going on re: explainability, either direct or with saliency maps (left for us to interpret). </p>\n\n<p>Here's a few references which may be interesting:</p>\n\n<p><a href=\"https://distill.pub/2017/feature-visualization/\">Feature visualization</a></p>\n\n<p><a href=\"http://netdissect.csail.mit.edu/\">Network dissection</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "280351": "I was wondering if anyone has taken a look at some of the intermediate layers and seen what were they detecting? I may look into it myself, but wanted to see if someone else has not already investigated this.",
    "280420": "It would be a great post (or even a nice paper to publish) if you pursue in this direction.",
    "280425": "It must have learned some features about lens, e.g., lens distortion, LED mechnisms(How RGB pixels are formed). This is something interesting. I also want to know the answer.",
    "280561": "Very interesting to know. I find it a bit funny we finish this competition with close to ~0.99 accuracy and we don't know what features the architectures extracted. There's so much research going on re: explainability, either direct or with saliency maps (left for us to interpret). \n\nHere's a few references which may be interesting:\n\n[Feature visualization][1]\n\n[Network dissection][2]\n\n  [1]: https://distill.pub/2017/feature-visualization/\n  [2]: http://netdissect.csail.mit.edu/"
  },
  "source": "meta"
}