{
  "id": 296038,
  "title": "Update ideas for network construction.",
  "url": "/competitions/digit-recognizer/discussion/296038",
  "author_name": "Buqi_",
  "post_date": "2021-12-19T13:59:00.703000",
  "votes": 4,
  "comment_count": 0,
  "views": null,
  "content": "<p>Conv can be regarded as a superset of pool, mean pooling is a conv with a weight of 1/num, and max pooling can be regarded as a conv with a weight of 1 only at the position of the maximum value. So theoretically any pooling layer can be replaced with The corresponding conv layer. Then why use the pooling layer? In fact, this is a relatively strong prior artificially added, such as adding max-pooling after the first layer of convolution, which is believed to reduce some noise interference, and adding mean-pooling to the last layer, which is considered to be a weighting of features. In fact, I think how to use it. And now that there is more data, the pooling layer uses less. The first several layers (one or two) may use max-pooling, and the latter layers use less pooling.</p>",
  "messages": [
    {
      "id": 1623089,
      "postDate": "2021-12-19T13:59:00.703Z",
      "content": "<p>Conv can be regarded as a superset of pool, mean pooling is a conv with a weight of 1/num, and max pooling can be regarded as a conv with a weight of 1 only at the position of the maximum value. So theoretically any pooling layer can be replaced with The corresponding conv layer. Then why use the pooling layer? In fact, this is a relatively strong prior artificially added, such as adding max-pooling after the first layer of convolution, which is believed to reduce some noise interference, and adding mean-pooling to the last layer, which is considered to be a weighting of features. In fact, I think how to use it. And now that there is more data, the pooling layer uses less. The first several layers (one or two) may use max-pooling, and the latter layers use less pooling.</p>",
      "rawMarkdown": "Conv can be regarded as a superset of pool, mean pooling is a conv with a weight of 1/num, and max pooling can be regarded as a conv with a weight of 1 only at the position of the maximum value. So theoretically any pooling layer can be replaced with The corresponding conv layer. Then why use the pooling layer? In fact, this is a relatively strong prior artificially added, such as adding max-pooling after the first layer of convolution, which is believed to reduce some noise interference, and adding mean-pooling to the last layer, which is considered to be a weighting of features. In fact, I think how to use it. And now that there is more data, the pooling layer uses less. The first several layers (one or two) may use max-pooling, and the latter layers use less pooling.",
      "votes": 4
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "1623089": "Conv can be regarded as a superset of pool, mean pooling is a conv with a weight of 1/num, and max pooling can be regarded as a conv with a weight of 1 only at the position of the maximum value. So theoretically any pooling layer can be replaced with The corresponding conv layer. Then why use the pooling layer? In fact, this is a relatively strong prior artificially added, such as adding max-pooling after the first layer of convolution, which is believed to reduce some noise interference, and adding mean-pooling to the last layer, which is considered to be a weighting of features. In fact, I think how to use it. And now that there is more data, the pooling layer uses less. The first several layers (one or two) may use max-pooling, and the latter layers use less pooling."
  }
}