{
  "id": 163675,
  "title": "Is it important to keep a cover-stego pair in a minibatch?",
  "url": "/competitions/alaska2-image-steganalysis/discussion/163675",
  "author_name": "",
  "post_date": "2020-07-03T03:22:56.371603800Z",
  "votes": null,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Because lots of papers claimed that it is  important to keep a cover-stego pair in a minibatch.\nBut the 0.921 notebook seems fine without it?</p>",
  "messages": [
    {
      "id": "913167",
      "postDate": "07/03/2020 03:22:56",
      "content": "<p>Because lots of papers claimed that it is  important to keep a cover-stego pair in a minibatch.\nBut the 0.921 notebook seems fine without it?</p>",
      "rawMarkdown": "Because lots of papers claimed that it is  important to keep a cover-stego pair in a minibatch.\nBut the 0.921 notebook seems fine without it?",
      "votes": null
    },
    {
      "id": "913723",
      "postDate": "07/03/2020 11:25:53",
      "content": "<p>my experiments show pairwise training overfits more.</p>",
      "rawMarkdown": "my experiments show pairwise training overfits more.",
      "votes": null
    },
    {
      "id": "913814",
      "postDate": "07/03/2020 12:56:10",
      "content": "<p>Think of the objective / use case. For test set, you will not be provided the image pair, so best to create a model pipeline that caters to that. If you include the pair, you run the risk of your classifier learning to identify between the specific pairs, as opposed to being able to distinguish between steganographicly altered and non altered images.</p>",
      "rawMarkdown": "Think of the objective / use case. For test set, you will not be provided the image pair, so best to create a model pipeline that caters to that. If you include the pair, you run the risk of your classifier learning to identify between the specific pairs, as opposed to being able to distinguish between steganographicly altered and non altered images.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 913723,
      "author_name": "sharksbeer",
      "author_url": "",
      "post_date": "07/03/2020 11:25:53",
      "content": "<p>my experiments show pairwise training overfits more.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 913814,
      "author_name": "authman",
      "author_url": "",
      "post_date": "07/03/2020 12:56:10",
      "content": "<p>Think of the objective / use case. For test set, you will not be provided the image pair, so best to create a model pipeline that caters to that. If you include the pair, you run the risk of your classifier learning to identify between the specific pairs, as opposed to being able to distinguish between steganographicly altered and non altered images.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "913167": "Because lots of papers claimed that it is  important to keep a cover-stego pair in a minibatch.\nBut the 0.921 notebook seems fine without it?",
    "913723": "my experiments show pairwise training overfits more.",
    "913814": "Think of the objective / use case. For test set, you will not be provided the image pair, so best to create a model pipeline that caters to that. If you include the pair, you run the risk of your classifier learning to identify between the specific pairs, as opposed to being able to distinguish between steganographicly altered and non altered images."
  },
  "source": "meta"
}