{
  "id": 111661,
  "title": "Optimisation when it comes to NN",
  "url": "/competitions/Kannada-MNIST/discussion/111661",
  "author_name": "",
  "post_date": "2019-10-07T19:06:51.820596100Z",
  "votes": 1,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Does anybody have any good starting points for optimising the model? There does not seem to be an obvious way to do this with NN, it seems like it is more trial and error than other types of machine learning. </p>\n\n<p>If it was a classification problem I would use something like gridsearch. Is there an equivalent for NNs?</p>",
  "messages": [
    {
      "id": "643673",
      "postDate": "10/07/2019 19:06:51",
      "content": "<p>Does anybody have any good starting points for optimising the model? There does not seem to be an obvious way to do this with NN, it seems like it is more trial and error than other types of machine learning. </p>\n\n<p>If it was a classification problem I would use something like gridsearch. Is there an equivalent for NNs?</p>",
      "rawMarkdown": "Does anybody have any good starting points for optimising the model? There does not seem to be an obvious way to do this with NN, it seems like it is more trial and error than other types of machine learning. \n\nIf it was a classification problem I would use something like gridsearch. Is there an equivalent for NNs?",
      "votes": null
    },
    {
      "id": "644240",
      "postDate": "10/08/2019 14:24:21",
      "content": "<p>Yes, I realized that too. The symbols are too much spread in the 28x28 area. \nNot sure about the NNs equivalent - I am using fast,ai and looks better for identifying the MNIST like data.\nyou can try too.\nthx\nA</p>",
      "rawMarkdown": "Yes, I realized that too. The symbols are too much spread in the 28x28 area. \nNot sure about the NNs equivalent - I am using fast,ai and looks better for identifying the MNIST like data.\nyou can try too.\nthx\nA",
      "votes": null
    },
    {
      "id": "664388",
      "postDate": "11/03/2019 15:54:58",
      "content": "<p>Just found this kernel which is good for explaining CNN architecture <a href=\"https://www.kaggle.com/cdeotte/how-to-choose-cnn-architecture-mnist\">https://www.kaggle.com/cdeotte/how-to-choose-cnn-architecture-mnist</a></p>",
      "rawMarkdown": "Just found this kernel which is good for explaining CNN architecture https://www.kaggle.com/cdeotte/how-to-choose-cnn-architecture-mnist",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 644240,
      "author_name": "zinovadr",
      "author_url": "",
      "post_date": "10/08/2019 14:24:21",
      "content": "<p>Yes, I realized that too. The symbols are too much spread in the 28x28 area. \nNot sure about the NNs equivalent - I am using fast,ai and looks better for identifying the MNIST like data.\nyou can try too.\nthx\nA</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 664388,
      "author_name": "jakelj",
      "author_url": "",
      "post_date": "11/03/2019 15:54:58",
      "content": "<p>Just found this kernel which is good for explaining CNN architecture <a href=\"https://www.kaggle.com/cdeotte/how-to-choose-cnn-architecture-mnist\">https://www.kaggle.com/cdeotte/how-to-choose-cnn-architecture-mnist</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "643673": "Does anybody have any good starting points for optimising the model? There does not seem to be an obvious way to do this with NN, it seems like it is more trial and error than other types of machine learning. \n\nIf it was a classification problem I would use something like gridsearch. Is there an equivalent for NNs?",
    "644240": "Yes, I realized that too. The symbols are too much spread in the 28x28 area. \nNot sure about the NNs equivalent - I am using fast,ai and looks better for identifying the MNIST like data.\nyou can try too.\nthx\nA",
    "664388": "Just found this kernel which is good for explaining CNN architecture https://www.kaggle.com/cdeotte/how-to-choose-cnn-architecture-mnist"
  },
  "source": "meta"
}