{
  "id": 46674,
  "title": "Transfer Learning: LogReg > NN: what am I doing wrong??",
  "url": "/competitions/dog-breed-identification/discussion/46674",
  "author_name": "",
  "post_date": "2018-01-01T11:34:01.059407600Z",
  "votes": 1,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I have tried many different models for generating bottleneck features. I have tried training Neural Network tops off of those bottleneck features, and also off of sets of bottleneck features concatenated together. My Neural Networks seem stall almost a full point below just running Logistic Regression on the training bottlenecks.</p>\n\n<p>For my NN tops, I have tried various combinations of large dense layers and dropouts leading to a final softmax layer for the output. What am I doing wrong here??</p>",
  "messages": [
    {
      "id": "263860",
      "postDate": "01/01/2018 11:34:01",
      "content": "<p>I have tried many different models for generating bottleneck features. I have tried training Neural Network tops off of those bottleneck features, and also off of sets of bottleneck features concatenated together. My Neural Networks seem stall almost a full point below just running Logistic Regression on the training bottlenecks.</p>\n\n<p>For my NN tops, I have tried various combinations of large dense layers and dropouts leading to a final softmax layer for the output. What am I doing wrong here??</p>",
      "rawMarkdown": "I have tried many different models for generating bottleneck features. I have tried training Neural Network tops off of those bottleneck features, and also off of sets of bottleneck features concatenated together. My Neural Networks seem stall almost a full point below just running Logistic Regression on the training bottlenecks.\n\nFor my NN tops, I have tried various combinations of large dense layers and dropouts leading to a final softmax layer for the output. What am I doing wrong here??",
      "votes": null
    },
    {
      "id": "264159",
      "postDate": "01/02/2018 12:40:42",
      "content": "<p>Why do you think something is wrong? Simple logistic regression might be better than overcomplicated NN. \nDense layers and dropout do not work? Well, then you might already achieved theoretical max with this model/approach. Try another one. </p>",
      "rawMarkdown": "Why do you think something is wrong? Simple logistic regression might be better than overcomplicated NN. \nDense layers and dropout do not work? Well, then you might already achieved theoretical max with this model/approach. Try another one.",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 264159,
      "author_name": "sermakarevich",
      "author_url": "",
      "post_date": "01/02/2018 12:40:42",
      "content": "<p>Why do you think something is wrong? Simple logistic regression might be better than overcomplicated NN. \nDense layers and dropout do not work? Well, then you might already achieved theoretical max with this model/approach. Try another one. </p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "263860": "I have tried many different models for generating bottleneck features. I have tried training Neural Network tops off of those bottleneck features, and also off of sets of bottleneck features concatenated together. My Neural Networks seem stall almost a full point below just running Logistic Regression on the training bottlenecks.\n\nFor my NN tops, I have tried various combinations of large dense layers and dropouts leading to a final softmax layer for the output. What am I doing wrong here??",
    "264159": "Why do you think something is wrong? Simple logistic regression might be better than overcomplicated NN. \nDense layers and dropout do not work? Well, then you might already achieved theoretical max with this model/approach. Try another one."
  },
  "source": "meta"
}