{
  "id": 89369,
  "title": "why there is no traditional kernel or discussion for this competiton ",
  "url": "/competitions/freesound-audio-tagging-2019/discussion/89369",
  "author_name": "er9213",
  "post_date": "2019-04-13T13:36:42.724000",
  "votes": 1,
  "comment_count": 6,
  "views": 0,
  "content": "<p>as title</p>",
  "messages": [
    {
      "id": 516350,
      "postDate": "2019-04-14T03:37:32.643Z",
      "content": "<p>Did you mean not deep neural network models? I guess this could partially because you could hardly find traditional models in top 50 of last year's similar competition. Even the last year official baseline model is a CNN. This year is MobileNet. Modern deep models most likely to perform better than traditional ones. </p>\n\n<p>Just curious, why ask?</p>",
      "rawMarkdown": "Did you mean not deep neural network models? I guess this could partially because you could hardly find traditional models in top 50 of last year's similar competition. Even the last year official baseline model is a CNN. This year is MobileNet. Modern deep models most likely to perform better than traditional ones. \n\nJust curious, why ask?",
      "votes": 1,
      "replies": [
        {
          "id": 516778,
          "postDate": "2019-04-15T00:46:15.130Z",
          "content": "<p>thanks for ur reply，despite of the deep learning tech performance，i think traditional method can be perform as good as dl at least if better data represention . Also if there are total dl methods,we may ignore the domain knowledge where still important </p>",
          "rawMarkdown": "thanks for ur reply，despite of the deep learning tech performance，i think traditional method can be perform as good as dl at least if better data represention . Also if there are total dl methods,we may ignore the domain knowledge where still important "
        },
        {
          "id": 517707,
          "postDate": "2019-04-16T12:05:56.463Z",
          "content": "<p>Domain knowledge is not ignored completely. STFT, mel spectrograms, data-specific augmentations - all these come from domain knowledge.</p>",
          "rawMarkdown": "Domain knowledge is not ignored completely. STFT, mel spectrograms, data-specific augmentations - all these come from domain knowledge."
        },
        {
          "id": 518976,
          "postDate": "2019-04-18T07:10:00.367Z",
          "content": "<p>Exactly. I think people are ignoring (despite copy-paste feature extraction) modeling with GMM-HMM, although this part of traditional systems probably performs worse and is harder to train than DNN.</p>",
          "rawMarkdown": "Exactly. I think people are ignoring (despite copy-paste feature extraction) modeling with GMM-HMM, although this part of traditional systems probably performs worse and is harder to train than DNN."
        },
        {
          "id": 521722,
          "postDate": "2019-04-23T10:36:57.167Z",
          "content": "<p>I guess, people ignores that this is audio competition, but image competition...\nOne of my intention to provide kernels is making audio conversion to image easier, but I don't see conversation regarding that for now...\nI'm digging in audio &amp; label features...</p>",
          "rawMarkdown": "I guess, people ignores that this is audio competition, but image competition...\nOne of my intention to provide kernels is making audio conversion to image easier, but I don't see conversation regarding that for now...\nI'm digging in audio &amp; label features...",
          "votes": 1
        },
        {
          "id": 522186,
          "postDate": "2019-04-24T01:50:43.150Z",
          "rawMarkdown": "",
          "votes": 1,
          "isDeleted": true
        }
      ]
    },
    {
      "id": 515983,
      "postDate": "2019-04-13T13:36:42.723Z",
      "content": "<p>as title</p>",
      "rawMarkdown": "as title",
      "votes": 1
    }
  ],
  "comments": [
    {
      "id": 516350,
      "author_name": "Sean XIONG",
      "author_url": "",
      "post_date": "2019-04-14T03:37:32.643000",
      "content": "<p>Did you mean not deep neural network models? I guess this could partially because you could hardly find traditional models in top 50 of last year's similar competition. Even the last year official baseline model is a CNN. This year is MobileNet. Modern deep models most likely to perform better than traditional ones. </p>\n\n<p>Just curious, why ask?</p>",
      "votes": 1,
      "replies": [
        {
          "id": 516778,
          "author_name": "er9213",
          "author_url": "",
          "post_date": "2019-04-15T00:46:15.130000",
          "content": "<p>thanks for ur reply，despite of the deep learning tech performance，i think traditional method can be perform as good as dl at least if better data represention . Also if there are total dl methods,we may ignore the domain knowledge where still important </p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 517707,
          "author_name": "Dmytro Danevskyi",
          "author_url": "",
          "post_date": "2019-04-16T12:05:56.463000",
          "content": "<p>Domain knowledge is not ignored completely. STFT, mel spectrograms, data-specific augmentations - all these come from domain knowledge.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 518976,
          "author_name": "DavidS",
          "author_url": "",
          "post_date": "2019-04-18T07:10:00.367000",
          "content": "<p>Exactly. I think people are ignoring (despite copy-paste feature extraction) modeling with GMM-HMM, although this part of traditional systems probably performs worse and is harder to train than DNN.</p>",
          "votes": 0,
          "replies": []
        },
        {
          "id": 521722,
          "author_name": "daisukelab",
          "author_url": "",
          "post_date": "2019-04-23T10:36:57.167000",
          "content": "<p>I guess, people ignores that this is audio competition, but image competition...\nOne of my intention to provide kernels is making audio conversion to image easier, but I don't see conversation regarding that for now...\nI'm digging in audio &amp; label features...</p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 522186,
          "author_name": "",
          "author_url": "",
          "post_date": "2019-04-24T01:50:43.150000",
          "content": "",
          "votes": 1,
          "replies": []
        }
      ]
    }
  ],
  "raw_markdown_by_id": {
    "516350": "Did you mean not deep neural network models? I guess this could partially because you could hardly find traditional models in top 50 of last year's similar competition. Even the last year official baseline model is a CNN. This year is MobileNet. Modern deep models most likely to perform better than traditional ones. \n\nJust curious, why ask?",
    "515983": "as title"
  }
}