{
  "id": 20802,
  "title": "What values should I choose for my neural network weight initialization?",
  "url": "/competitions/digit-recognizer/discussion/20802",
  "author_name": "VZ",
  "post_date": "2016-05-08T09:29:14.190000",
  "votes": 0,
  "comment_count": 0,
  "views": null,
  "content": "<p>I am trying to write a 2-layer neural net using theano. What is the best weight initialization values? I am setting them randomly between 0 and 0.1, but I noticed that if I set them randomly between -0.1 and +0.1 I get a much worse performance. Why having initial negative weights would affect performance so much?</p>",
  "messages": [
    {
      "id": 119232,
      "postDate": "2016-05-08T09:29:14.190Z",
      "content": "<p>I am trying to write a 2-layer neural net using theano. What is the best weight initialization values? I am setting them randomly between 0 and 0.1, but I noticed that if I set them randomly between -0.1 and +0.1 I get a much worse performance. Why having initial negative weights would affect performance so much?</p>",
      "rawMarkdown": "I am trying to write a 2-layer neural net using theano. What is the best weight initialization values? I am setting them randomly between 0 and 0.1, but I noticed that if I set them randomly between -0.1 and +0.1 I get a much worse performance. Why having initial negative weights would affect performance so much?"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "119232": "I am trying to write a 2-layer neural net using theano. What is the best weight initialization values? I am setting them randomly between 0 and 0.1, but I noticed that if I set them randomly between -0.1 and +0.1 I get a much worse performance. Why having initial negative weights would affect performance so much?"
  }
}