{
  "id": 202581,
  "title": "Deep Learning on Controlled Noisy Labels",
  "url": "/competitions/cassava-leaf-disease-classification/discussion/202581",
  "author_name": "",
  "post_date": "2020-12-10T21:32:36.534255700Z",
  "votes": 16,
  "comment_count": 1,
  "views": 0,
  "content": "<p>I've encountered great research work about addressing noisy labels.  <a href=\"https://arxiv.org/abs/1911.09781\" target=\"_blank\">Beyond Synthetic Noise: Deep Learning on Controlled Noisy Labels</a>. Practical recommendations for training deep neural networks on noisy data, read the whole blog <a href=\"https://ai.googleblog.com/2020/08/understanding-deep-learning-on.html#:~:text=A%20simple%20way%20to%20deal,label%20noise%20from%20the%20web.\" target=\"_blank\">here</a>. <a href=\"https://github.com/google-research/google-research/tree/master/mentormix/code\" target=\"_blank\">Code</a>. Here are the key findings:</p>\n<ul>\n<li>A simple way to deal with <strong>noisy labels</strong> is to fine-tune a model that is pre-trained on clean datasets, like ImageNet. The better the pre-trained model is, the better it may generalize on downstream noisy training tasks.</li>\n<li>Early stopping may not be effective on the real-world label noise from the web.</li>\n<li>Methods that perform well on synthetic noise may not work as well on the real-world noisy labels from the web.</li>\n<li>The label noise from the web appears to be less harmful, yet it is more difficult for our current robust learning methods to tackle. This encourages more future research to be carried out on controlled real-world label noise.</li>\n<li>The proposed <strong>MentorMix</strong> can better overcome both synthetic and real-world noisy labels.</li>\n</ul>",
  "messages": [
    {
      "id": "1108658",
      "postDate": "12/10/2020 21:32:36",
      "content": "<p>I've encountered great research work about addressing noisy labels.  <a href=\"https://arxiv.org/abs/1911.09781\" target=\"_blank\">Beyond Synthetic Noise: Deep Learning on Controlled Noisy Labels</a>. Practical recommendations for training deep neural networks on noisy data, read the whole blog <a href=\"https://ai.googleblog.com/2020/08/understanding-deep-learning-on.html#:~:text=A%20simple%20way%20to%20deal,label%20noise%20from%20the%20web.\" target=\"_blank\">here</a>. <a href=\"https://github.com/google-research/google-research/tree/master/mentormix/code\" target=\"_blank\">Code</a>. Here are the key findings:</p>\n<ul>\n<li>A simple way to deal with <strong>noisy labels</strong> is to fine-tune a model that is pre-trained on clean datasets, like ImageNet. The better the pre-trained model is, the better it may generalize on downstream noisy training tasks.</li>\n<li>Early stopping may not be effective on the real-world label noise from the web.</li>\n<li>Methods that perform well on synthetic noise may not work as well on the real-world noisy labels from the web.</li>\n<li>The label noise from the web appears to be less harmful, yet it is more difficult for our current robust learning methods to tackle. This encourages more future research to be carried out on controlled real-world label noise.</li>\n<li>The proposed <strong>MentorMix</strong> can better overcome both synthetic and real-world noisy labels.</li>\n</ul>",
      "rawMarkdown": "I've encountered great research work about addressing noisy labels.  [Beyond Synthetic Noise: Deep Learning on Controlled Noisy Labels](https://arxiv.org/abs/1911.09781). Practical recommendations for training deep neural networks on noisy data, read the whole blog [here](https://ai.googleblog.com/2020/08/understanding-deep-learning-on.html#:~:text=A%20simple%20way%20to%20deal,label%20noise%20from%20the%20web.). [Code](https://github.com/google-research/google-research/tree/master/mentormix/code). Here are the key findings:\n\n- A simple way to deal with **noisy labels** is to fine-tune a model that is pre-trained on clean datasets, like ImageNet. The better the pre-trained model is, the better it may generalize on downstream noisy training tasks.\n- Early stopping may not be effective on the real-world label noise from the web.\n- Methods that perform well on synthetic noise may not work as well on the real-world noisy labels from the web.\n- The label noise from the web appears to be less harmful, yet it is more difficult for our current robust learning methods to tackle. This encourages more future research to be carried out on controlled real-world label noise.\n- The proposed **MentorMix** can better overcome both synthetic and real-world noisy labels.",
      "votes": null
    },
    {
      "id": "1110913",
      "postDate": "12/13/2020 07:38:41",
      "content": "<p>Thanks for this! I had a doubt that in all this research was it trained on noisy labels but tested on good ones, or both the training and testing labels were noisy?</p>",
      "rawMarkdown": "Thanks for this! I had a doubt that in all this research was it trained on noisy labels but tested on good ones, or both the training and testing labels were noisy?",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 1110913,
      "author_name": "mrinath",
      "author_url": "",
      "post_date": "12/13/2020 07:38:41",
      "content": "<p>Thanks for this! I had a doubt that in all this research was it trained on noisy labels but tested on good ones, or both the training and testing labels were noisy?</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1108658": "I've encountered great research work about addressing noisy labels.  [Beyond Synthetic Noise: Deep Learning on Controlled Noisy Labels](https://arxiv.org/abs/1911.09781). Practical recommendations for training deep neural networks on noisy data, read the whole blog [here](https://ai.googleblog.com/2020/08/understanding-deep-learning-on.html#:~:text=A%20simple%20way%20to%20deal,label%20noise%20from%20the%20web.). [Code](https://github.com/google-research/google-research/tree/master/mentormix/code). Here are the key findings:\n\n- A simple way to deal with **noisy labels** is to fine-tune a model that is pre-trained on clean datasets, like ImageNet. The better the pre-trained model is, the better it may generalize on downstream noisy training tasks.\n- Early stopping may not be effective on the real-world label noise from the web.\n- Methods that perform well on synthetic noise may not work as well on the real-world noisy labels from the web.\n- The label noise from the web appears to be less harmful, yet it is more difficult for our current robust learning methods to tackle. This encourages more future research to be carried out on controlled real-world label noise.\n- The proposed **MentorMix** can better overcome both synthetic and real-world noisy labels.",
    "1110913": "Thanks for this! I had a doubt that in all this research was it trained on noisy labels but tested on good ones, or both the training and testing labels were noisy?"
  },
  "source": "meta"
}