{
  "id": 32789,
  "title": "How to Improve?",
  "url": "/competitions/digit-recognizer/discussion/32789",
  "author_name": "erikjwagner",
  "post_date": "2017-05-10T09:39:01.054000",
  "votes": 0,
  "comment_count": 1,
  "views": null,
  "content": "<p>Hi, this is a 'got any tips?' sort of post.  I am pretty new to ML (completed a course but this is my first 'official' on my own try).  I built a Multi-class neural network that showed a 97.9% accuracy on the test data and a 96.8% accuracy on the actual submission.  Does anyone have any tips on how I can increase accuracy?  I tried several different models and the multi-class neural network appeared to show the most promise.  Did I totally misread which model type would be the best?\nAny help/tips would be great (only asking since this is a 'free' one).  Thanks!</p>",
  "messages": [
    {
      "id": 182590,
      "postDate": "2017-05-14T18:55:11.720Z",
      "content": "<p>Hey you there!</p>\n\n<p>I think if you do understand well Neural Networks the next step is to try a Convolutional Neural Network (CNN) and you can easily get 99% of accuracy.</p>\n\n<p>There a some great introductions to CNNs, but I will give you one of my favorite intro videos (Siraj videos are incredible for an introduction):  <a href=\"https://www.youtube.com/watch?v=cAICT4Al5Ow\">https://www.youtube.com/watch?v=cAICT4Al5Ow</a></p>\n\n<p>After watch that, you could read the CS231n course and understand better: <a href=\"http://cs231n.github.io/\">http://cs231n.github.io/</a>\nIf you program in Python, I think the Keras API will be helpful for you. Here is an example and you just need to adapt to use the Kaggle data: <a href=\"https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py\">https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py</a></p>\n\n<p>Good luck,</p>",
      "rawMarkdown": "Hey you there!\n\nI think if you do understand well Neural Networks the next step is to try a Convolutional Neural Network (CNN) and you can easily get 99% of accuracy.\n\nThere a some great introductions to CNNs, but I will give you one of my favorite intro videos (Siraj videos are incredible for an introduction):  https://www.youtube.com/watch?v=cAICT4Al5Ow\n\nAfter watch that, you could read the CS231n course and understand better: http://cs231n.github.io/\nIf you program in Python, I think the Keras API will be helpful for you. Here is an example and you just need to adapt to use the Kaggle data: https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py\n\nGood luck,",
      "votes": 1
    },
    {
      "id": 181625,
      "postDate": "2017-05-10T09:39:01.053Z",
      "content": "<p>Hi, this is a 'got any tips?' sort of post.  I am pretty new to ML (completed a course but this is my first 'official' on my own try).  I built a Multi-class neural network that showed a 97.9% accuracy on the test data and a 96.8% accuracy on the actual submission.  Does anyone have any tips on how I can increase accuracy?  I tried several different models and the multi-class neural network appeared to show the most promise.  Did I totally misread which model type would be the best?\nAny help/tips would be great (only asking since this is a 'free' one).  Thanks!</p>",
      "rawMarkdown": "Hi, this is a 'got any tips?' sort of post.  I am pretty new to ML (completed a course but this is my first 'official' on my own try).  I built a Multi-class neural network that showed a 97.9% accuracy on the test data and a 96.8% accuracy on the actual submission.  Does anyone have any tips on how I can increase accuracy?  I tried several different models and the multi-class neural network appeared to show the most promise.  Did I totally misread which model type would be the best?\nAny help/tips would be great (only asking since this is a 'free' one).  Thanks!"
    }
  ],
  "comments": [
    {
      "id": 182590,
      "author_name": "",
      "author_url": "",
      "post_date": "2017-05-14T18:55:11.720000",
      "content": "<p>Hey you there!</p>\n\n<p>I think if you do understand well Neural Networks the next step is to try a Convolutional Neural Network (CNN) and you can easily get 99% of accuracy.</p>\n\n<p>There a some great introductions to CNNs, but I will give you one of my favorite intro videos (Siraj videos are incredible for an introduction):  <a href=\"https://www.youtube.com/watch?v=cAICT4Al5Ow\">https://www.youtube.com/watch?v=cAICT4Al5Ow</a></p>\n\n<p>After watch that, you could read the CS231n course and understand better: <a href=\"http://cs231n.github.io/\">http://cs231n.github.io/</a>\nIf you program in Python, I think the Keras API will be helpful for you. Here is an example and you just need to adapt to use the Kaggle data: <a href=\"https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py\">https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py</a></p>\n\n<p>Good luck,</p>",
      "votes": 1,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "182590": "Hey you there!\n\nI think if you do understand well Neural Networks the next step is to try a Convolutional Neural Network (CNN) and you can easily get 99% of accuracy.\n\nThere a some great introductions to CNNs, but I will give you one of my favorite intro videos (Siraj videos are incredible for an introduction):  https://www.youtube.com/watch?v=cAICT4Al5Ow\n\nAfter watch that, you could read the CS231n course and understand better: http://cs231n.github.io/\nIf you program in Python, I think the Keras API will be helpful for you. Here is an example and you just need to adapt to use the Kaggle data: https://github.com/fchollet/keras/blob/master/examples/mnist_cnn.py\n\nGood luck,",
    "181625": "Hi, this is a 'got any tips?' sort of post.  I am pretty new to ML (completed a course but this is my first 'official' on my own try).  I built a Multi-class neural network that showed a 97.9% accuracy on the test data and a 96.8% accuracy on the actual submission.  Does anyone have any tips on how I can increase accuracy?  I tried several different models and the multi-class neural network appeared to show the most promise.  Did I totally misread which model type would be the best?\nAny help/tips would be great (only asking since this is a 'free' one).  Thanks!"
  }
}