{
  "id": 71467,
  "title": "Explore limits & Error Analysis of SRk’s Glove+GRU",
  "url": "/competitions/quora-insincere-questions-classification/discussion/71467",
  "author_name": "Neuron Engineer",
  "post_date": "2018-11-13T23:38:37.598000",
  "votes": 3,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Hi everyone, this is my first kernel for competition. Hope it will be helpful and can produce more ideas : <br>\n<a href=\"https://www.kaggle.com/ratthachat/explore-limits-error-analysis-of-srk-s-glove-gru\">https://www.kaggle.com/ratthachat/explore-limits-error-analysis-of-srk-s-glove-gru</a></p>\n\n<p>In the kernel , we would like to explore the limit of this neural architecture (we choose GloVe as the representative embedding since all of the embedding methods have similar performances) by answering the following questions :</p>\n\n<p><strong>1) Is this architecture's capacity overfit or underfit the problem?</strong>\nCan we improve the performance by running more and more epochs?\nShould we increase the capacity of the network, i.e. increases more latent dimension or add more layers? (underfitting case)</p>\n\n<p><strong>2) At its best, what kind of predictions do the network trying to make?</strong>\nSince the 'insincere' class has only a small number of data, our network has to be careful when it will predict 'class 1'.</p>\n\n<p>Does it try to make only a sure prediction? (try a small number of class-1 prediction, but each one of them is precisely correct)\nOR, Does it make class-1 prediction a lot (in order to cover most class 1 data), but hopefully not make too much mistakes to predict class 0 as class 1.</p>\n\n<p><strong>3) At its best, what kind of errors do the network make?</strong>\nwhat are insincere topics where the network strongly believe to be sincere ?\nwhat are sincere topics where the network strongly believe to be insincere ?\nwhat are insincere topics where the network are most uncertain how to classify ?\nwhat are sincere topics where the network are most uncertain how to classify?\n(Not error, but good to see) what are insincere topics where the network strongly believe correctly ?</p>",
  "messages": [
    {
      "id": 420640,
      "postDate": "2018-11-13T23:38:37.600Z",
      "content": "<p>Hi everyone, this is my first kernel for competition. Hope it will be helpful and can produce more ideas : <br>\n<a href=\"https://www.kaggle.com/ratthachat/explore-limits-error-analysis-of-srk-s-glove-gru\">https://www.kaggle.com/ratthachat/explore-limits-error-analysis-of-srk-s-glove-gru</a></p>\n\n<p>In the kernel , we would like to explore the limit of this neural architecture (we choose GloVe as the representative embedding since all of the embedding methods have similar performances) by answering the following questions :</p>\n\n<p><strong>1) Is this architecture's capacity overfit or underfit the problem?</strong>\nCan we improve the performance by running more and more epochs?\nShould we increase the capacity of the network, i.e. increases more latent dimension or add more layers? (underfitting case)</p>\n\n<p><strong>2) At its best, what kind of predictions do the network trying to make?</strong>\nSince the 'insincere' class has only a small number of data, our network has to be careful when it will predict 'class 1'.</p>\n\n<p>Does it try to make only a sure prediction? (try a small number of class-1 prediction, but each one of them is precisely correct)\nOR, Does it make class-1 prediction a lot (in order to cover most class 1 data), but hopefully not make too much mistakes to predict class 0 as class 1.</p>\n\n<p><strong>3) At its best, what kind of errors do the network make?</strong>\nwhat are insincere topics where the network strongly believe to be sincere ?\nwhat are sincere topics where the network strongly believe to be insincere ?\nwhat are insincere topics where the network are most uncertain how to classify ?\nwhat are sincere topics where the network are most uncertain how to classify?\n(Not error, but good to see) what are insincere topics where the network strongly believe correctly ?</p>",
      "rawMarkdown": "Hi everyone, this is my first kernel for competition. Hope it will be helpful and can produce more ideas :  \nhttps://www.kaggle.com/ratthachat/explore-limits-error-analysis-of-srk-s-glove-gru\n\nIn the kernel , we would like to explore the limit of this neural architecture (we choose GloVe as the representative embedding since all of the embedding methods have similar performances) by answering the following questions :\n\n**1) Is this architecture's capacity overfit or underfit the problem?**\nCan we improve the performance by running more and more epochs?\nShould we increase the capacity of the network, i.e. increases more latent dimension or add more layers? (underfitting case)\n\n**2) At its best, what kind of predictions do the network trying to make?**\nSince the 'insincere' class has only a small number of data, our network has to be careful when it will predict 'class 1'.\n\nDoes it try to make only a sure prediction? (try a small number of class-1 prediction, but each one of them is precisely correct)\nOR, Does it make class-1 prediction a lot (in order to cover most class 1 data), but hopefully not make too much mistakes to predict class 0 as class 1.\n\n**3) At its best, what kind of errors do the network make?**\nwhat are insincere topics where the network strongly believe to be sincere ?\nwhat are sincere topics where the network strongly believe to be insincere ?\nwhat are insincere topics where the network are most uncertain how to classify ?\nwhat are sincere topics where the network are most uncertain how to classify?\n(Not error, but good to see) what are insincere topics where the network strongly believe correctly ?\n",
      "votes": 3
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "420640": "Hi everyone, this is my first kernel for competition. Hope it will be helpful and can produce more ideas :  \nhttps://www.kaggle.com/ratthachat/explore-limits-error-analysis-of-srk-s-glove-gru\n\nIn the kernel , we would like to explore the limit of this neural architecture (we choose GloVe as the representative embedding since all of the embedding methods have similar performances) by answering the following questions :\n\n**1) Is this architecture's capacity overfit or underfit the problem?**\nCan we improve the performance by running more and more epochs?\nShould we increase the capacity of the network, i.e. increases more latent dimension or add more layers? (underfitting case)\n\n**2) At its best, what kind of predictions do the network trying to make?**\nSince the 'insincere' class has only a small number of data, our network has to be careful when it will predict 'class 1'.\n\nDoes it try to make only a sure prediction? (try a small number of class-1 prediction, but each one of them is precisely correct)\nOR, Does it make class-1 prediction a lot (in order to cover most class 1 data), but hopefully not make too much mistakes to predict class 0 as class 1.\n\n**3) At its best, what kind of errors do the network make?**\nwhat are insincere topics where the network strongly believe to be sincere ?\nwhat are sincere topics where the network strongly believe to be insincere ?\nwhat are insincere topics where the network are most uncertain how to classify ?\nwhat are sincere topics where the network are most uncertain how to classify?\n(Not error, but good to see) what are insincere topics where the network strongly believe correctly ?\n"
  }
}