{
  "id": 468963,
  "title": "Why is this kind of sequence completion used in all CNN networks?",
  "url": "/competitions/hms-harmful-brain-activity-classification/discussion/468963",
  "author_name": "",
  "post_date": "2024-01-18T14:02:47.965367100Z",
  "votes": 2,
  "comment_count": 2,
  "views": 0,
  "content": "<p>Currently, the highest-scoring public notebook utilizes ResNet (LB 0.46), followed closely by EfficientNetB2 (LB0.57). Why do they choose CNN networks, and why do they work so well? Has anyone tried using an RNN network and achieved good results?</p>",
  "messages": [
    {
      "id": "2607921",
      "postDate": "01/18/2024 14:02:47",
      "content": "<p>Currently, the highest-scoring public notebook utilizes ResNet (LB 0.46), followed closely by EfficientNetB2 (LB0.57). Why do they choose CNN networks, and why do they work so well? Has anyone tried using an RNN network and achieved good results?</p>",
      "rawMarkdown": "Currently, the highest-scoring public notebook utilizes ResNet (LB 0.46), followed closely by EfficientNetB2 (LB0.57). Why do they choose CNN networks, and why do they work so well? Has anyone tried using an RNN network and achieved good results?",
      "votes": null
    },
    {
      "id": "2607936",
      "postDate": "01/18/2024 14:05:38",
      "content": "<p><a href=\"https://www.kaggle.com/mengvision\" target=\"_blank\">@mengvision</a> for both ResNet &amp; EfficientNetB2 used spectrograms. these CNN models performed well for spectrograms. </p>",
      "rawMarkdown": "mengvision for both ResNet & EfficientNetB2 used spectrograms. these CNN models performed well for spectrograms.",
      "votes": null
    },
    {
      "id": "2607958",
      "postDate": "01/18/2024 14:19:37",
      "content": "<p>The original raw data is EEG waveforms (these are like audio signals). If we use these waveforms, then time based models like RNN, CNN, Transformer work best. A common technique (when working with signals) is to transform the raw waveform into image spectrograms. These are images which represent the presence of time of frequency. If we use spectrograms as input to our models, then image models work best.</p>\n<p>Above I describe DL deep learning. If we use ML machine learning then we engineer our own features. In this case, we can use GBT to perform well with both spectrogram and waveform.</p>",
      "rawMarkdown": "The original raw data is EEG waveforms (these are like audio signals). If we use these waveforms, then time based models like RNN, CNN, Transformer work best. A common technique (when working with signals) is to transform the raw waveform into image spectrograms. These are images which represent the presence of time of frequency. If we use spectrograms as input to our models, then image models work best.\n\nAbove I describe DL deep learning. If we use ML machine learning then we engineer our own features. In this case, we can use GBT to perform well with both spectrogram and waveform.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 2607936,
      "author_name": "seshurajup",
      "author_url": "",
      "post_date": "01/18/2024 14:05:38",
      "content": "<p><a href=\"https://www.kaggle.com/mengvision\" target=\"_blank\">@mengvision</a> for both ResNet &amp; EfficientNetB2 used spectrograms. these CNN models performed well for spectrograms. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 2607958,
      "author_name": "cdeotte",
      "author_url": "",
      "post_date": "01/18/2024 14:19:37",
      "content": "<p>The original raw data is EEG waveforms (these are like audio signals). If we use these waveforms, then time based models like RNN, CNN, Transformer work best. A common technique (when working with signals) is to transform the raw waveform into image spectrograms. These are images which represent the presence of time of frequency. If we use spectrograms as input to our models, then image models work best.</p>\n<p>Above I describe DL deep learning. If we use ML machine learning then we engineer our own features. In this case, we can use GBT to perform well with both spectrogram and waveform.</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "2607921": "Currently, the highest-scoring public notebook utilizes ResNet (LB 0.46), followed closely by EfficientNetB2 (LB0.57). Why do they choose CNN networks, and why do they work so well? Has anyone tried using an RNN network and achieved good results?",
    "2607936": "mengvision for both ResNet & EfficientNetB2 used spectrograms. these CNN models performed well for spectrograms.",
    "2607958": "The original raw data is EEG waveforms (these are like audio signals). If we use these waveforms, then time based models like RNN, CNN, Transformer work best. A common technique (when working with signals) is to transform the raw waveform into image spectrograms. These are images which represent the presence of time of frequency. If we use spectrograms as input to our models, then image models work best.\n\nAbove I describe DL deep learning. If we use ML machine learning then we engineer our own features. In this case, we can use GBT to perform well with both spectrogram and waveform."
  },
  "source": "meta"
}