{
  "id": 47889,
  "title": "A feeling of incompleteness",
  "url": "/competitions/tensorflow-speech-recognition-challenge/discussion/47889",
  "author_name": "",
  "post_date": "2018-01-20T07:32:19.150091Z",
  "votes": 6,
  "comment_count": 1,
  "views": 0,
  "content": "<p>Hi all,\n   I have been a software engineer for past ~15 years, but never studied neural networks. Only lately I got enthusiastic about it when I read that a wide variety of problems were getting solved using neural networks. I took Andrew Ng's course series, and joined this competition as a way to get hands on experience. It has been great participating here.</p>\n\n<p>However, there is a feeling of ignorance, emptiness and incompleteness. Together, we have come up with solutions which perhaps improve the state of the art, but our solutions hardly throw any insight on the problem. Our methods involve running lots of different models, do variety of data augmentations, going through a hyper parameters search, and then ensemble plenty of them. Some approaches work, others don't. We go in the direction of what works and discard what doesn't. However, we do not know (and don't even try) to find why something works.</p>\n\n<p>When a conventional program of mine produces a wrong output, I call it a \"bug\". I go through my program carefully, find the reason of incorrect output, and correct it. However, when a model of mine produces a wrong label, I am out of my wits. Only ideas I have are like: train larger model, train deeper model, do augmentation etc, and hope that accuracy will improve. </p>\n\n<p>Perhaps as researchers work more on neural networks, we might get to peek inside why some models work, and that will give us better ways to construct good models, than the current approach which is like probing dark.</p>\n\n<p>Thanks for reading through this rant. Hope to see you in other competitions.</p>",
  "messages": [
    {
      "id": "271337",
      "postDate": "01/20/2018 07:32:19",
      "content": "<p>Hi all,\n   I have been a software engineer for past ~15 years, but never studied neural networks. Only lately I got enthusiastic about it when I read that a wide variety of problems were getting solved using neural networks. I took Andrew Ng's course series, and joined this competition as a way to get hands on experience. It has been great participating here.</p>\n\n<p>However, there is a feeling of ignorance, emptiness and incompleteness. Together, we have come up with solutions which perhaps improve the state of the art, but our solutions hardly throw any insight on the problem. Our methods involve running lots of different models, do variety of data augmentations, going through a hyper parameters search, and then ensemble plenty of them. Some approaches work, others don't. We go in the direction of what works and discard what doesn't. However, we do not know (and don't even try) to find why something works.</p>\n\n<p>When a conventional program of mine produces a wrong output, I call it a \"bug\". I go through my program carefully, find the reason of incorrect output, and correct it. However, when a model of mine produces a wrong label, I am out of my wits. Only ideas I have are like: train larger model, train deeper model, do augmentation etc, and hope that accuracy will improve. </p>\n\n<p>Perhaps as researchers work more on neural networks, we might get to peek inside why some models work, and that will give us better ways to construct good models, than the current approach which is like probing dark.</p>\n\n<p>Thanks for reading through this rant. Hope to see you in other competitions.</p>",
      "rawMarkdown": "Hi all,\n   I have been a software engineer for past ~15 years, but never studied neural networks. Only lately I got enthusiastic about it when I read that a wide variety of problems were getting solved using neural networks. I took Andrew Ng's course series, and joined this competition as a way to get hands on experience. It has been great participating here.\n\n   However, there is a feeling of ignorance, emptiness and incompleteness. Together, we have come up with solutions which perhaps improve the state of the art, but our solutions hardly throw any insight on the problem. Our methods involve running lots of different models, do variety of data augmentations, going through a hyper parameters search, and then ensemble plenty of them. Some approaches work, others don't. We go in the direction of what works and discard what doesn't. However, we do not know (and don't even try) to find why something works.\n\n When a conventional program of mine produces a wrong output, I call it a \"bug\". I go through my program carefully, find the reason of incorrect output, and correct it. However, when a model of mine produces a wrong label, I am out of my wits. Only ideas I have are like: train larger model, train deeper model, do augmentation etc, and hope that accuracy will improve. \n\n   Perhaps as researchers work more on neural networks, we might get to peek inside why some models work, and that will give us better ways to construct good models, than the current approach which is like probing dark.\n\n   Thanks for reading through this rant. Hope to see you in other competitions.",
      "votes": null
    },
    {
      "id": "271503",
      "postDate": "01/20/2018 15:56:19",
      "content": "<p>Hi right now there is now answer to a simple question how does it work this reward presentation on youtube tells the story\nAli Rahimi's talk at NIPS</p>",
      "rawMarkdown": "Hi right now there is now answer to a simple question how does it work this reward presentation on youtube tells the story\nAli Rahimi's talk at NIPS",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 271503,
      "author_name": "andre1233",
      "author_url": "",
      "post_date": "01/20/2018 15:56:19",
      "content": "<p>Hi right now there is now answer to a simple question how does it work this reward presentation on youtube tells the story\nAli Rahimi's talk at NIPS</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "271337": "Hi all,\n   I have been a software engineer for past ~15 years, but never studied neural networks. Only lately I got enthusiastic about it when I read that a wide variety of problems were getting solved using neural networks. I took Andrew Ng's course series, and joined this competition as a way to get hands on experience. It has been great participating here.\n\n   However, there is a feeling of ignorance, emptiness and incompleteness. Together, we have come up with solutions which perhaps improve the state of the art, but our solutions hardly throw any insight on the problem. Our methods involve running lots of different models, do variety of data augmentations, going through a hyper parameters search, and then ensemble plenty of them. Some approaches work, others don't. We go in the direction of what works and discard what doesn't. However, we do not know (and don't even try) to find why something works.\n\n When a conventional program of mine produces a wrong output, I call it a \"bug\". I go through my program carefully, find the reason of incorrect output, and correct it. However, when a model of mine produces a wrong label, I am out of my wits. Only ideas I have are like: train larger model, train deeper model, do augmentation etc, and hope that accuracy will improve. \n\n   Perhaps as researchers work more on neural networks, we might get to peek inside why some models work, and that will give us better ways to construct good models, than the current approach which is like probing dark.\n\n   Thanks for reading through this rant. Hope to see you in other competitions.",
    "271503": "Hi right now there is now answer to a simple question how does it work this reward presentation on youtube tells the story\nAli Rahimi's talk at NIPS"
  },
  "source": "meta"
}