{
  "id": 128222,
  "title": "Reading the Robot Mind",
  "url": "/competitions/bengaliai-cv19/discussion/128222",
  "author_name": "",
  "post_date": "2020-01-29T17:47:54.443457800Z",
  "votes": 4,
  "comment_count": 1,
  "views": 0,
  "content": "<p>It's difficult to choose parameters for your network, and much of it is left to trial and error. </p>\n\n<p>What if you could \"read the robot mind\" and see the internal representation of the input within the network as it is running?</p>\n\n<p>This example Kernel does just that. Analyzing one of the more popular public Kernels in this competition, this notebook works through the multi-layer convolution network backwards - creating the input based on the output.</p>\n\n<p>By visually seeing the filters, as well as a recreation of the input, it is easy to see when important details are lost (resolution), cut off (clipping), and when too many or not enough filters are included in a particular layer.</p>\n\n<p>Check it out; and let me know if you think it would be a useful tool to make standard within Keras...</p>\n\n<p><a href=\"https://www.kaggle.com/pnussbaum/grapheme-mind-reader-panv12-nogpu\">https://www.kaggle.com/pnussbaum/grapheme-mind-reader-panv12-nogpu</a></p>",
  "messages": [
    {
      "id": "732341",
      "postDate": "01/29/2020 17:47:54",
      "content": "<p>It's difficult to choose parameters for your network, and much of it is left to trial and error. </p>\n\n<p>What if you could \"read the robot mind\" and see the internal representation of the input within the network as it is running?</p>\n\n<p>This example Kernel does just that. Analyzing one of the more popular public Kernels in this competition, this notebook works through the multi-layer convolution network backwards - creating the input based on the output.</p>\n\n<p>By visually seeing the filters, as well as a recreation of the input, it is easy to see when important details are lost (resolution), cut off (clipping), and when too many or not enough filters are included in a particular layer.</p>\n\n<p>Check it out; and let me know if you think it would be a useful tool to make standard within Keras...</p>\n\n<p><a href=\"https://www.kaggle.com/pnussbaum/grapheme-mind-reader-panv12-nogpu\">https://www.kaggle.com/pnussbaum/grapheme-mind-reader-panv12-nogpu</a></p>",
      "rawMarkdown": "It's difficult to choose parameters for your network, and much of it is left to trial and error. \n\nWhat if you could \"read the robot mind\" and see the internal representation of the input within the network as it is running?\n\nThis example Kernel does just that. Analyzing one of the more popular public Kernels in this competition, this notebook works through the multi-layer convolution network backwards - creating the input based on the output.\n\nBy visually seeing the filters, as well as a recreation of the input, it is easy to see when important details are lost (resolution), cut off (clipping), and when too many or not enough filters are included in a particular layer.\n\nCheck it out; and let me know if you think it would be a useful tool to make standard within Keras...\n\nhttps://www.kaggle.com/pnussbaum/grapheme-mind-reader-panv12-nogpu",
      "votes": null
    },
    {
      "id": "743378",
      "postDate": "02/12/2020 02:39:23",
      "content": "<p>Current version has cleaned up code, and now recreates input all the way to the deepest Conv2D layer. This answers some questions, but raises others. See comments after recreation images for more details. Enjoy!</p>",
      "rawMarkdown": "Current version has cleaned up code, and now recreates input all the way to the deepest Conv2D layer. This answers some questions, but raises others. See comments after recreation images for more details. Enjoy!",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 743378,
      "author_name": "pnussbaum",
      "author_url": "",
      "post_date": "02/12/2020 02:39:23",
      "content": "<p>Current version has cleaned up code, and now recreates input all the way to the deepest Conv2D layer. This answers some questions, but raises others. See comments after recreation images for more details. Enjoy!</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "732341": "It's difficult to choose parameters for your network, and much of it is left to trial and error. \n\nWhat if you could \"read the robot mind\" and see the internal representation of the input within the network as it is running?\n\nThis example Kernel does just that. Analyzing one of the more popular public Kernels in this competition, this notebook works through the multi-layer convolution network backwards - creating the input based on the output.\n\nBy visually seeing the filters, as well as a recreation of the input, it is easy to see when important details are lost (resolution), cut off (clipping), and when too many or not enough filters are included in a particular layer.\n\nCheck it out; and let me know if you think it would be a useful tool to make standard within Keras...\n\nhttps://www.kaggle.com/pnussbaum/grapheme-mind-reader-panv12-nogpu",
    "743378": "Current version has cleaned up code, and now recreates input all the way to the deepest Conv2D layer. This answers some questions, but raises others. See comments after recreation images for more details. Enjoy!"
  },
  "source": "meta"
}