{
  "id": 237000,
  "title": "Optimal network depth",
  "url": "/competitions/digit-recognizer/discussion/237000",
  "author_name": "Georgii Sirotenko",
  "post_date": "2021-05-06T16:39:09.762000",
  "votes": 0,
  "comment_count": 1,
  "views": null,
  "content": "<p>How deep do you think the network should be to solve MNIST? Is it worth using too deep pre-trained models, for example, DenseNet121, or is it too much in this case?</p>",
  "messages": [
    {
      "id": 1305851,
      "postDate": "2021-05-13T14:08:44.197Z",
      "content": "<p>Hi Georgii, you can make it as deep as you want as long as you use ResNet to deal with the vanishing of the gradient. But for a simple data set like digit recognizer, you don't need to go so deep. You can utilize a shallow network with nonlinear activation functions and unorthodox layers, like, utilizing batch normalization layer and dropout layer.</p>",
      "rawMarkdown": "Hi Georgii, you can make it as deep as you want as long as you use ResNet to deal with the vanishing of the gradient. But for a simple data set like digit recognizer, you don't need to go so deep. You can utilize a shallow network with nonlinear activation functions and unorthodox layers, like, utilizing batch normalization layer and dropout layer.",
      "votes": 1
    },
    {
      "id": 1295700,
      "postDate": "2021-05-06T16:39:09.763Z",
      "content": "<p>How deep do you think the network should be to solve MNIST? Is it worth using too deep pre-trained models, for example, DenseNet121, or is it too much in this case?</p>",
      "rawMarkdown": "How deep do you think the network should be to solve MNIST? Is it worth using too deep pre-trained models, for example, DenseNet121, or is it too much in this case?"
    }
  ],
  "comments": [
    {
      "id": 1305851,
      "author_name": "Rasheed AZ",
      "author_url": "",
      "post_date": "2021-05-13T14:08:44.197000",
      "content": "<p>Hi Georgii, you can make it as deep as you want as long as you use ResNet to deal with the vanishing of the gradient. But for a simple data set like digit recognizer, you don't need to go so deep. You can utilize a shallow network with nonlinear activation functions and unorthodox layers, like, utilizing batch normalization layer and dropout layer.</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1305851": "Hi Georgii, you can make it as deep as you want as long as you use ResNet to deal with the vanishing of the gradient. But for a simple data set like digit recognizer, you don't need to go so deep. You can utilize a shallow network with nonlinear activation functions and unorthodox layers, like, utilizing batch normalization layer and dropout layer.",
    "1295700": "How deep do you think the network should be to solve MNIST? Is it worth using too deep pre-trained models, for example, DenseNet121, or is it too much in this case?"
  }
}