{
  "id": 52845,
  "title": "Paper: Generating and designing DNA with deep GANs",
  "url": "/competitions/data-science-bowl-2018/discussion/52845",
  "author_name": "James Requa",
  "post_date": "2018-03-23T21:24:07.013000",
  "votes": 0,
  "comment_count": 0,
  "views": 0,
  "content": "<p>Came across this paper from a recent tweet made by Ian Goodfellow. While we aren't dealing with DNA sequences per se, but it still presents a good method with application subject matter relevant to this competition so thought I'd share it here with fellow contestants. \n<a href=\"https://arxiv.org/pdf/1712.06148.pdf\">https://arxiv.org/pdf/1712.06148.pdf</a></p>\n\n<p>The paper introduces some tweaks that can be made to GANs to hopefully make them perform better on generating more specialized/niche data. </p>\n\n<p>One issue with style transfer type GANs (i.e. CycleGAN) technique is that it <code>ignores realism or typicality of data in its pursuit of optimal attributes</code> which can lead to pretty wacky images that probably won't help if used during training.  So the clever solution here was to create a <code>joint architecture with two components: a generator G which transforms latent codes z\ninto synthetic data x (e.g., a trained GAN generator), and a predictor P, mapping data x to the\ncorresponding target attributes t = P(x).</code> So basically we have more control over the generative features so they better retain their \"nuclei-style\" properties. </p>",
  "messages": [
    {
      "id": 302281,
      "postDate": "2018-03-23T21:24:07.013Z",
      "content": "<p>Came across this paper from a recent tweet made by Ian Goodfellow. While we aren't dealing with DNA sequences per se, but it still presents a good method with application subject matter relevant to this competition so thought I'd share it here with fellow contestants. \n<a href=\"https://arxiv.org/pdf/1712.06148.pdf\">https://arxiv.org/pdf/1712.06148.pdf</a></p>\n\n<p>The paper introduces some tweaks that can be made to GANs to hopefully make them perform better on generating more specialized/niche data. </p>\n\n<p>One issue with style transfer type GANs (i.e. CycleGAN) technique is that it <code>ignores realism or typicality of data in its pursuit of optimal attributes</code> which can lead to pretty wacky images that probably won't help if used during training.  So the clever solution here was to create a <code>joint architecture with two components: a generator G which transforms latent codes z\ninto synthetic data x (e.g., a trained GAN generator), and a predictor P, mapping data x to the\ncorresponding target attributes t = P(x).</code> So basically we have more control over the generative features so they better retain their \"nuclei-style\" properties. </p>",
      "rawMarkdown": "Came across this paper from a recent tweet made by Ian Goodfellow. While we aren't dealing with DNA sequences per se, but it still presents a good method with application subject matter relevant to this competition so thought I'd share it here with fellow contestants. \nhttps://arxiv.org/pdf/1712.06148.pdf\n\nThe paper introduces some tweaks that can be made to GANs to hopefully make them perform better on generating more specialized/niche data. \n\nOne issue with style transfer type GANs (i.e. CycleGAN) technique is that it `ignores realism or typicality of data in its pursuit of optimal attributes` which can lead to pretty wacky images that probably won't help if used during training.  So the clever solution here was to create a `joint architecture with two components: a generator G which transforms latent codes z\ninto synthetic data x (e.g., a trained GAN generator), and a predictor P, mapping data x to the\ncorresponding target attributes t = P(x).` So basically we have more control over the generative features so they better retain their \"nuclei-style\" properties. \n"
    }
  ],
  "comments": [],
  "raw_markdown_by_id": {
    "302281": "Came across this paper from a recent tweet made by Ian Goodfellow. While we aren't dealing with DNA sequences per se, but it still presents a good method with application subject matter relevant to this competition so thought I'd share it here with fellow contestants. \nhttps://arxiv.org/pdf/1712.06148.pdf\n\nThe paper introduces some tweaks that can be made to GANs to hopefully make them perform better on generating more specialized/niche data. \n\nOne issue with style transfer type GANs (i.e. CycleGAN) technique is that it `ignores realism or typicality of data in its pursuit of optimal attributes` which can lead to pretty wacky images that probably won't help if used during training.  So the clever solution here was to create a `joint architecture with two components: a generator G which transforms latent codes z\ninto synthetic data x (e.g., a trained GAN generator), and a predictor P, mapping data x to the\ncorresponding target attributes t = P(x).` So basically we have more control over the generative features so they better retain their \"nuclei-style\" properties. \n"
  }
}