{
  "id": 174267,
  "title": "Merging manual features with neural-network ones?",
  "url": "/competitions/birdsong-recognition/discussion/174267",
  "author_name": "leodav",
  "post_date": "2020-08-12T22:54:06.242000",
  "votes": 1,
  "comment_count": 0,
  "views": 0,
  "content": "<p>I have tried the following two approaches but they are not giving much improvement and the third is too hard for me to implement would involve into doing something like DeepMind have done for alpha go.  Do you have any other suggestions?</p>\n<p>1) Include manual feature into spectrogram images and train on it</p>\n<p>2) Train on spectrograms and then combine those outputs with manual features.  Using regular boosting methods does not provide sufficient boost.</p>\n<p>3) those familiar with party with the mob approach, realize similar things can be done on neural networks; the key point here is that soft variables are used for propogation of the signal while hard for branch selection.  That is also somewhat similar for the DeepMind approach for alpha go.  However this approach is way to involved.</p>\n<p>This is my first Kaggle competition so I do not know the etiquette here.  Would it be proper to double this question in general discussion on kaggle? </p>",
  "messages": [
    {
      "id": 968343,
      "postDate": "2020-08-12T22:54:06.243Z",
      "content": "<p>I have tried the following two approaches but they are not giving much improvement and the third is too hard for me to implement would involve into doing something like DeepMind have done for alpha go.  Do you have any other suggestions?</p>\n<p>1) Include manual feature into spectrogram images and train on it</p>\n<p>2) Train on spectrograms and then combine those outputs with manual features.  Using regular boosting methods does not provide sufficient boost.</p>\n<p>3) those familiar with party with the mob approach, realize similar things can be done on neural networks; the key point here is that soft variables are used for propogation of the signal while hard for branch selection.  That is also somewhat similar for the DeepMind approach for alpha go.  However this approach is way to involved.</p>\n<p>This is my first Kaggle competition so I do not know the etiquette here.  Would it be proper to double this question in general discussion on kaggle? </p>",
      "rawMarkdown": "I have tried the following two approaches but they are not giving much improvement and the third is too hard for me to implement would involve into doing something like DeepMind have done for alpha go.  Do you have any other suggestions?\n\n1) Include manual feature into spectrogram images and train on it\n\n2) Train on spectrograms and then combine those outputs with manual features.  Using regular boosting methods does not provide sufficient boost.\n\n3) those familiar with party with the mob approach, realize similar things can be done on neural networks; the key point here is that soft variables are used for propogation of the signal while hard for branch selection.  That is also somewhat similar for the DeepMind approach for alpha go.  However this approach is way to involved.\n\nThis is my first Kaggle competition so I do not know the etiquette here.  Would it be proper to double this question in general discussion on kaggle? ",
      "votes": 1
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "968343": "I have tried the following two approaches but they are not giving much improvement and the third is too hard for me to implement would involve into doing something like DeepMind have done for alpha go.  Do you have any other suggestions?\n\n1) Include manual feature into spectrogram images and train on it\n\n2) Train on spectrograms and then combine those outputs with manual features.  Using regular boosting methods does not provide sufficient boost.\n\n3) those familiar with party with the mob approach, realize similar things can be done on neural networks; the key point here is that soft variables are used for propogation of the signal while hard for branch selection.  That is also somewhat similar for the DeepMind approach for alpha go.  However this approach is way to involved.\n\nThis is my first Kaggle competition so I do not know the etiquette here.  Would it be proper to double this question in general discussion on kaggle? "
  }
}