{
  "id": 17929,
  "title": "Deep Learning Tutorial is dissapointing ",
  "url": "/competitions/second-annual-data-science-bowl/discussion/17929",
  "author_name": "",
  "post_date": "2015-12-15T18:07:27.573Z",
  "votes": null,
  "comment_count": 17,
  "views": 4442,
  "content": "<p>I went through the deep learning tutorial, but LB score(LB:0.134308) is disappointing. Did I miss something? Anyone got a similar score of it?   </p>",
  "messages": [
    {
      "id": "101519",
      "postDate": "12/15/2015 18:07:27",
      "content": "<p>I went through the deep learning tutorial, but LB score(LB:0.134308) is disappointing. Did I miss something? Anyone got a similar score of it?   </p>",
      "rawMarkdown": "I went through the deep learning tutorial, but LB score(LB:0.134308) is disappointing. Did I miss something? Anyone got a similar score of it?",
      "votes": null
    },
    {
      "id": "101530",
      "postDate": "12/15/2015 19:12:06",
      "content": "<p>I have not test it, but I tried the Fourier tutorial and even it is a starting point, it is a very low one as well, so, better to work on your own solution :(</p>",
      "rawMarkdown": "I have not test it, but I tried the Fourier tutorial and even it is a starting point, it is a very low one as well, so, better to work on your own solution :(",
      "votes": null
    },
    {
      "id": "101532",
      "postDate": "12/15/2015 19:41:12",
      "content": "<p>Both deep learning tutorial and Fourier tutorial have similar performance in term of RMSE on train set. It does make sense they both have bad LB score. What score did you get for Fourier? It is around 0.13?</p>",
      "rawMarkdown": "Both deep learning tutorial and Fourier tutorial have similar performance in term of RMSE on train set. It does make sense they both have bad LB score. What score did you get for Fourier? It is around 0.13?",
      "votes": null
    },
    {
      "id": "101535",
      "postDate": "12/15/2015 19:45:22",
      "content": "<p>I got 0.120050 without any parameter tuning. Seems like tuning might improve it a bit, but I didn't bother as the score is very bad.</p>",
      "rawMarkdown": "I got 0.120050 without any parameter tuning. Seems like tuning might improve it a bit, but I didn't bother as the score is very bad.",
      "votes": null
    },
    {
      "id": "101538",
      "postDate": "12/15/2015 19:51:18",
      "content": "<p>Thanks NxGTR. Guess it is time to build our own model. </p>",
      "rawMarkdown": "Thanks NxGTR. Guess it is time to build our own model.",
      "votes": null
    },
    {
      "id": "101540",
      "postDate": "12/15/2015 20:05:52",
      "content": "<p>Welcome, gogogogo!</p>",
      "rawMarkdown": "Welcome, gogogogo!",
      "votes": null
    },
    {
      "id": "101553",
      "postDate": "12/15/2015 21:22:27",
      "content": "<p>I'm the author of the deep learning tutorial. In an effort to encourage people to continue exploring a deep learning solution, I will qualify the tutorial's results and clarify its intent. The main goal of the tutorial is to provide the user (novice or otherwise) a well-guided tour of a feasible solution using convolutional nets. The tutorials were never about a competitive solution, as we have clearly stated. Below are two main points to qualify the tutorials (both deep learning and Fourier-based):</p>\n\n<ol>\n<li>One can use the median value, file size, and other heuristics to &quot;beat the benchmark&quot; in this competition. However, those heuristics are not applicable in the clinical setting. In the clinical setting, the physician is given MRI images and is asked to automatically provide an accurate EF measure. Providing the median value of EF from the training set would be a terrible way of diagnosing cardiovascular health. Both tutorials work directly with MRI images to extract LV contours and make EF predictions, similar to how a cardiac physician performs the task. The tutorials provide sensible methods to achieve the goal based on measurable features.</li>\n<li>The goal of the tutorials is to provide a baseline model, a starting point, from which upon the reader is encouraged to expand. If you read carefully in the tutorials, we provide ideas on how one can reduce error from our baseline models. This is to say that the reader must provide some work, using the tutorials as a starting point, to be competitive. I will personally say this: if you follow my suggestions and recommendations, you will achieve significant improvement from the initial reported result.</li>\n</ol>\n\n<p>Good luck and have fun.</p>",
      "rawMarkdown": "I'm the author of the deep learning tutorial. In an effort to encourage people to continue exploring a deep learning solution, I will qualify the tutorial's results and clarify its intent. The main goal of the tutorial is to provide the user (novice or otherwise) a well-guided tour of a feasible solution using convolutional nets. The tutorials were never about a competitive solution, as we have clearly stated. Below are two main points to qualify the tutorials (both deep learning and Fourier-based):\r\n\r\n 1. One can use the median value, file size, and other heuristics to \"beat the benchmark\" in this competition. However, those heuristics are not applicable in the clinical setting. In the clinical setting, the physician is given MRI images and is asked to automatically provide an accurate EF measure. Providing the median value of EF from the training set would be a terrible way of diagnosing cardiovascular health. Both tutorials work directly with MRI images to extract LV contours and make EF predictions, similar to how a cardiac physician performs the task. The tutorials provide sensible methods to achieve the goal based on measurable features.\r\n 2. The goal of the tutorials is to provide a baseline model, a starting point, from which upon the reader is encouraged to expand. If you read carefully in the tutorials, we provide ideas on how one can reduce error from our baseline models. This is to say that the reader must provide some work, using the tutorials as a starting point, to be competitive. I will personally say this: if you follow my suggestions and recommendations, you will achieve significant improvement from the initial reported result.\r\n\r\nGood luck and have fun.",
      "votes": null
    },
    {
      "id": "101564",
      "postDate": "12/15/2015 22:00:29",
      "content": "<p>I guess no one here is complaining that both tutorial are not competitive. We certainly understood that provide the median of EF from training set would be terrible. However, if a method is sensible and bases on measurable features, at least it should be better than the useless median value &quot;beat the benchmark&quot; even without any tuning.    </p>\n\n<p>Thanks for providing tutorials anyway.</p>",
      "rawMarkdown": "I guess no one here is complaining that both tutorial are not competitive. We certainly understood that provide the median of EF from training set would be terrible. However, if a method is sensible and bases on measurable features, at least it should be better than the useless median value \"beat the benchmark\" even without any tuning.    \r\n\r\nThanks for providing tutorials anyway.",
      "votes": null
    },
    {
      "id": "101630",
      "postDate": "12/16/2015 07:05:32",
      "content": "<p>Please take the tutorial as it is to me: domain knowledge, features and potential automated feature engineering ideas. Kaggle has taught me even the poorest scoring ideas are often just a minor tweak away from exceptional scoring ideas. Even if I wrote a fairly decent scoring benchmark, Kagglers would surpass it in probably a week (maybe days or hours) or so. Hence whether you have a strong scoring tutorial benchmark or not is somewhat irrelevant in the grand scheme of things. The top 10 will always score much higher. I've never seen or participated in a Kaggle where even the strongest early models were demolished by the end of the competition. So have fun, share, and work towards great models with a high potential social impact.</p>",
      "rawMarkdown": "Please take the tutorial as it is to me: domain knowledge, features and potential automated feature engineering ideas. Kaggle has taught me even the poorest scoring ideas are often just a minor tweak away from exceptional scoring ideas. Even if I wrote a fairly decent scoring benchmark, Kagglers would surpass it in probably a week (maybe days or hours) or so. Hence whether you have a strong scoring tutorial benchmark or not is somewhat irrelevant in the grand scheme of things. The top 10 will always score much higher. I've never seen or participated in a Kaggle where even the strongest early models were demolished by the end of the competition. So have fun, share, and work towards great models with a high potential social impact.",
      "votes": null
    },
    {
      "id": "101688",
      "postDate": "12/16/2015 14:09:13",
      "content": "<p>I can't see the tutorial. When I open the web page, it just shows nothing.</p>",
      "rawMarkdown": "I can't see the tutorial. When I open the web page, it just shows nothing.",
      "votes": null
    },
    {
      "id": "101724",
      "postDate": "12/16/2015 17:48:45",
      "content": "<p>[quote=Jiming Ye;101688]</p>\n\n<p>I can't see the tutorial. When I open the web page, it just shows nothing.</p>\n\n<p>[/quote]</p>\n\n<p>Me too, you need to 'climb the wall'</p>",
      "rawMarkdown": "[quote=Jiming Ye;101688]\r\n\r\nI can't see the tutorial. When I open the web page, it just shows nothing.\r\n\r\n[/quote]\r\n\r\nMe too, you need to 'climb the wall'",
      "votes": null
    },
    {
      "id": "101732",
      "postDate": "12/16/2015 19:00:00",
      "content": "<p>The problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.</p>",
      "rawMarkdown": "The problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.",
      "votes": null
    },
    {
      "id": "101807",
      "postDate": "12/17/2015 06:37:01",
      "content": "<p>[quote=newebug;101724]</p>\n\n<p>[quote=Jiming Ye;101688]</p>\n\n<p>I can't see the tutorial. When I open the web page, it just shows nothing.</p>\n\n<p>[/quote]</p>\n\n<p>Me too, you need to 'climb the wall'</p>\n\n<p>[/quote]</p>\n\n<p>lol</p>",
      "rawMarkdown": "[quote=newebug;101724]\r\n\r\n[quote=Jiming Ye;101688]\r\n\r\nI can't see the tutorial. When I open the web page, it just shows nothing.\r\n\r\n[/quote]\r\n\r\nMe too, you need to 'climb the wall'\r\n\r\n[/quote]\r\n\r\nlol",
      "votes": null
    },
    {
      "id": "101813",
      "postDate": "12/17/2015 08:18:48",
      "content": "<p>[quote=DavidGbodiOdaibo;101732]</p>\n\n<p>The problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.</p>\n\n<p>[/quote] <br>\nI think the idea of the competition is that's we would have to do. Maybe a package was chosen but others would be relevant: <br>\n<a href=\"http://deeplearning.net/software_links/\">http://deeplearning.net/software_links/</a></p>",
      "rawMarkdown": "[quote=DavidGbodiOdaibo;101732]\r\n\r\nThe problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.\r\n\r\n[/quote]   \r\nI think the idea of the competition is that's we would have to do. Maybe a package was chosen but others would be relevant:  \r\nhttp://deeplearning.net/software_links/",
      "votes": null
    },
    {
      "id": "102117",
      "postDate": "12/19/2015 16:26:09",
      "content": "<p>I have successfully climbed the wall and seen the tutorials. It's a lot of work and really awesome. </p>\n\n<p>It doesn't use any information regarding the ground truth in train.csv, that's the reason why it performs bad. </p>",
      "rawMarkdown": "I have successfully climbed the wall and seen the tutorials. It's a lot of work and really awesome. \r\n\r\nIt doesn't use any information regarding the ground truth in train.csv, that's the reason why it performs bad.",
      "votes": null
    },
    {
      "id": "102205",
      "postDate": "12/20/2015 12:20:21",
      "content": "<p>I am trying to wrap my head around the deep learning tutoral. Am i right to understand (and please correct me where wrong) that the input file that the tutoral uses already has labelled LV surfaces / &quot;blobs&quot; / areas for the LV? And subsequently, that the deep learning is merely trying to supervised learning to try to predict these blobs from the images / pixels? </p>\n\n<p>I have not downloaded the full dataset yet - but the real data-set has no labellled LV surfaces - no? Thus - how much use is the supervised learning approach in the tutoral?</p>\n\n<p>I feel that i am missing something - please correct me!  </p>",
      "rawMarkdown": "I am trying to wrap my head around the deep learning tutoral. Am i right to understand (and please correct me where wrong) that the input file that the tutoral uses already has labelled LV surfaces / \"blobs\" / areas for the LV? And subsequently, that the deep learning is merely trying to supervised learning to try to predict these blobs from the images / pixels? \r\n\r\nI have not downloaded the full dataset yet - but the real data-set has no labellled LV surfaces - no? Thus - how much use is the supervised learning approach in the tutoral?\r\n\r\nI feel that i am missing something - please correct me!",
      "votes": null
    },
    {
      "id": "102219",
      "postDate": "12/20/2015 15:38:25",
      "content": "<p>You're description is correct</p>",
      "rawMarkdown": "You're description is correct",
      "votes": null
    },
    {
      "id": "102232",
      "postDate": "12/20/2015 19:25:47",
      "content": "<p>Nevermind, there's a much faster and compact registration at the original MICCAI page here: </p>\n\n<p><a href=\"http://smial.sri.utoronto.ca/LV_Challenge/Downloads.html\">http://smial.sri.utoronto.ca/LV_Challenge/Downloads.html</a></p>\n\n<p>Register there and they give you a download link right away.</p>\n\n<p>Hope this helps</p>\n\n<blockquote>\n  <blockquote>\n    <p>&gt;</p>\n  </blockquote>\n</blockquote>\n\n<p>Er, how are people getting access to the Sunnybrook dataset?  It appears to be behind a research registration wall &#8211; are they fairly permissive?</p>\n\n<p><a href=\"http://www.cardiacatlas.org/data-access/request-cap-access/\">http://www.cardiacatlas.org/data-access/request-cap-access/</a></p>",
      "rawMarkdown": "Nevermind, there's a much faster and compact registration at the original MICCAI page here: \r\n\r\nhttp://smial.sri.utoronto.ca/LV_Challenge/Downloads.html\r\n\r\nRegister there and they give you a download link right away.\r\n\r\nHope this helps\r\n\r\n\r\n>>>\r\n\r\nEr, how are people getting access to the Sunnybrook dataset?  It appears to be behind a research registration wall – are they fairly permissive?\r\n\r\nhttp://www.cardiacatlas.org/data-access/request-cap-access/",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 101530,
      "author_name": "carloshuertas",
      "author_url": "",
      "post_date": "12/15/2015 19:12:06",
      "content": "<p>I have not test it, but I tried the Fourier tutorial and even it is a starting point, it is a very low one as well, so, better to work on your own solution :(</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101532,
      "author_name": "jianminsun",
      "author_url": "",
      "post_date": "12/15/2015 19:41:12",
      "content": "<p>Both deep learning tutorial and Fourier tutorial have similar performance in term of RMSE on train set. It does make sense they both have bad LB score. What score did you get for Fourier? It is around 0.13?</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101535,
      "author_name": "carloshuertas",
      "author_url": "",
      "post_date": "12/15/2015 19:45:22",
      "content": "<p>I got 0.120050 without any parameter tuning. Seems like tuning might improve it a bit, but I didn't bother as the score is very bad.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101538,
      "author_name": "jianminsun",
      "author_url": "",
      "post_date": "12/15/2015 19:51:18",
      "content": "<p>Thanks NxGTR. Guess it is time to build our own model. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101540,
      "author_name": "carloshuertas",
      "author_url": "",
      "post_date": "12/15/2015 20:05:52",
      "content": "<p>Welcome, gogogogo!</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101553,
      "author_name": "vuptran",
      "author_url": "",
      "post_date": "12/15/2015 21:22:27",
      "content": "<p>I'm the author of the deep learning tutorial. In an effort to encourage people to continue exploring a deep learning solution, I will qualify the tutorial's results and clarify its intent. The main goal of the tutorial is to provide the user (novice or otherwise) a well-guided tour of a feasible solution using convolutional nets. The tutorials were never about a competitive solution, as we have clearly stated. Below are two main points to qualify the tutorials (both deep learning and Fourier-based):</p>\n\n<ol>\n<li>One can use the median value, file size, and other heuristics to &quot;beat the benchmark&quot; in this competition. However, those heuristics are not applicable in the clinical setting. In the clinical setting, the physician is given MRI images and is asked to automatically provide an accurate EF measure. Providing the median value of EF from the training set would be a terrible way of diagnosing cardiovascular health. Both tutorials work directly with MRI images to extract LV contours and make EF predictions, similar to how a cardiac physician performs the task. The tutorials provide sensible methods to achieve the goal based on measurable features.</li>\n<li>The goal of the tutorials is to provide a baseline model, a starting point, from which upon the reader is encouraged to expand. If you read carefully in the tutorials, we provide ideas on how one can reduce error from our baseline models. This is to say that the reader must provide some work, using the tutorials as a starting point, to be competitive. I will personally say this: if you follow my suggestions and recommendations, you will achieve significant improvement from the initial reported result.</li>\n</ol>\n\n<p>Good luck and have fun.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101564,
      "author_name": "jianminsun",
      "author_url": "",
      "post_date": "12/15/2015 22:00:29",
      "content": "<p>I guess no one here is complaining that both tutorial are not competitive. We certainly understood that provide the median of EF from training set would be terrible. However, if a method is sensible and bases on measurable features, at least it should be better than the useless median value &quot;beat the benchmark&quot; even without any tuning.    </p>\n\n<p>Thanks for providing tutorials anyway.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101630,
      "author_name": "mikeskim",
      "author_url": "",
      "post_date": "12/16/2015 07:05:32",
      "content": "<p>Please take the tutorial as it is to me: domain knowledge, features and potential automated feature engineering ideas. Kaggle has taught me even the poorest scoring ideas are often just a minor tweak away from exceptional scoring ideas. Even if I wrote a fairly decent scoring benchmark, Kagglers would surpass it in probably a week (maybe days or hours) or so. Hence whether you have a strong scoring tutorial benchmark or not is somewhat irrelevant in the grand scheme of things. The top 10 will always score much higher. I've never seen or participated in a Kaggle where even the strongest early models were demolished by the end of the competition. So have fun, share, and work towards great models with a high potential social impact.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101688,
      "author_name": "yejiming",
      "author_url": "",
      "post_date": "12/16/2015 14:09:13",
      "content": "<p>I can't see the tutorial. When I open the web page, it just shows nothing.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101724,
      "author_name": "newebug",
      "author_url": "",
      "post_date": "12/16/2015 17:48:45",
      "content": "<p>[quote=Jiming Ye;101688]</p>\n\n<p>I can't see the tutorial. When I open the web page, it just shows nothing.</p>\n\n<p>[/quote]</p>\n\n<p>Me too, you need to 'climb the wall'</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101732,
      "author_name": "godaibo",
      "author_url": "",
      "post_date": "12/16/2015 19:00:00",
      "content": "<p>The problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101807,
      "author_name": "yejiming",
      "author_url": "",
      "post_date": "12/17/2015 06:37:01",
      "content": "<p>[quote=newebug;101724]</p>\n\n<p>[quote=Jiming Ye;101688]</p>\n\n<p>I can't see the tutorial. When I open the web page, it just shows nothing.</p>\n\n<p>[/quote]</p>\n\n<p>Me too, you need to 'climb the wall'</p>\n\n<p>[/quote]</p>\n\n<p>lol</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 101813,
      "author_name": "wildwizard",
      "author_url": "",
      "post_date": "12/17/2015 08:18:48",
      "content": "<p>[quote=DavidGbodiOdaibo;101732]</p>\n\n<p>The problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.</p>\n\n<p>[/quote] <br>\nI think the idea of the competition is that's we would have to do. Maybe a package was chosen but others would be relevant: <br>\n<a href=\"http://deeplearning.net/software_links/\">http://deeplearning.net/software_links/</a></p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 102117,
      "author_name": "yejiming",
      "author_url": "",
      "post_date": "12/19/2015 16:26:09",
      "content": "<p>I have successfully climbed the wall and seen the tutorials. It's a lot of work and really awesome. </p>\n\n<p>It doesn't use any information regarding the ground truth in train.csv, that's the reason why it performs bad. </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 102205,
      "author_name": "wouterd1",
      "author_url": "",
      "post_date": "12/20/2015 12:20:21",
      "content": "<p>I am trying to wrap my head around the deep learning tutoral. Am i right to understand (and please correct me where wrong) that the input file that the tutoral uses already has labelled LV surfaces / &quot;blobs&quot; / areas for the LV? And subsequently, that the deep learning is merely trying to supervised learning to try to predict these blobs from the images / pixels? </p>\n\n<p>I have not downloaded the full dataset yet - but the real data-set has no labellled LV surfaces - no? Thus - how much use is the supervised learning approach in the tutoral?</p>\n\n<p>I feel that i am missing something - please correct me!  </p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 102219,
      "author_name": "udibr1",
      "author_url": "",
      "post_date": "12/20/2015 15:38:25",
      "content": "<p>You're description is correct</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 102232,
      "author_name": "leotam",
      "author_url": "",
      "post_date": "12/20/2015 19:25:47",
      "content": "<p>Nevermind, there's a much faster and compact registration at the original MICCAI page here: </p>\n\n<p><a href=\"http://smial.sri.utoronto.ca/LV_Challenge/Downloads.html\">http://smial.sri.utoronto.ca/LV_Challenge/Downloads.html</a></p>\n\n<p>Register there and they give you a download link right away.</p>\n\n<p>Hope this helps</p>\n\n<blockquote>\n  <blockquote>\n    <p>&gt;</p>\n  </blockquote>\n</blockquote>\n\n<p>Er, how are people getting access to the Sunnybrook dataset?  It appears to be behind a research registration wall &#8211; are they fairly permissive?</p>\n\n<p><a href=\"http://www.cardiacatlas.org/data-access/request-cap-access/\">http://www.cardiacatlas.org/data-access/request-cap-access/</a></p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "101519": "I went through the deep learning tutorial, but LB score(LB:0.134308) is disappointing. Did I miss something? Anyone got a similar score of it?",
    "101530": "I have not test it, but I tried the Fourier tutorial and even it is a starting point, it is a very low one as well, so, better to work on your own solution :(",
    "101532": "Both deep learning tutorial and Fourier tutorial have similar performance in term of RMSE on train set. It does make sense they both have bad LB score. What score did you get for Fourier? It is around 0.13?",
    "101535": "I got 0.120050 without any parameter tuning. Seems like tuning might improve it a bit, but I didn't bother as the score is very bad.",
    "101538": "Thanks NxGTR. Guess it is time to build our own model.",
    "101540": "Welcome, gogogogo!",
    "101553": "I'm the author of the deep learning tutorial. In an effort to encourage people to continue exploring a deep learning solution, I will qualify the tutorial's results and clarify its intent. The main goal of the tutorial is to provide the user (novice or otherwise) a well-guided tour of a feasible solution using convolutional nets. The tutorials were never about a competitive solution, as we have clearly stated. Below are two main points to qualify the tutorials (both deep learning and Fourier-based):\r\n\r\n 1. One can use the median value, file size, and other heuristics to \"beat the benchmark\" in this competition. However, those heuristics are not applicable in the clinical setting. In the clinical setting, the physician is given MRI images and is asked to automatically provide an accurate EF measure. Providing the median value of EF from the training set would be a terrible way of diagnosing cardiovascular health. Both tutorials work directly with MRI images to extract LV contours and make EF predictions, similar to how a cardiac physician performs the task. The tutorials provide sensible methods to achieve the goal based on measurable features.\r\n 2. The goal of the tutorials is to provide a baseline model, a starting point, from which upon the reader is encouraged to expand. If you read carefully in the tutorials, we provide ideas on how one can reduce error from our baseline models. This is to say that the reader must provide some work, using the tutorials as a starting point, to be competitive. I will personally say this: if you follow my suggestions and recommendations, you will achieve significant improvement from the initial reported result.\r\n\r\nGood luck and have fun.",
    "101564": "I guess no one here is complaining that both tutorial are not competitive. We certainly understood that provide the median of EF from training set would be terrible. However, if a method is sensible and bases on measurable features, at least it should be better than the useless median value \"beat the benchmark\" even without any tuning.    \r\n\r\nThanks for providing tutorials anyway.",
    "101630": "Please take the tutorial as it is to me: domain knowledge, features and potential automated feature engineering ideas. Kaggle has taught me even the poorest scoring ideas are often just a minor tweak away from exceptional scoring ideas. Even if I wrote a fairly decent scoring benchmark, Kagglers would surpass it in probably a week (maybe days or hours) or so. Hence whether you have a strong scoring tutorial benchmark or not is somewhat irrelevant in the grand scheme of things. The top 10 will always score much higher. I've never seen or participated in a Kaggle where even the strongest early models were demolished by the end of the competition. So have fun, share, and work towards great models with a high potential social impact.",
    "101688": "I can't see the tutorial. When I open the web page, it just shows nothing.",
    "101724": "[quote=Jiming Ye;101688]\r\n\r\nI can't see the tutorial. When I open the web page, it just shows nothing.\r\n\r\n[/quote]\r\n\r\nMe too, you need to 'climb the wall'",
    "101732": "The problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.",
    "101807": "[quote=newebug;101724]\r\n\r\n[quote=Jiming Ye;101688]\r\n\r\nI can't see the tutorial. When I open the web page, it just shows nothing.\r\n\r\n[/quote]\r\n\r\nMe too, you need to 'climb the wall'\r\n\r\n[/quote]\r\n\r\nlol",
    "101813": "[quote=DavidGbodiOdaibo;101732]\r\n\r\nThe problem I see with the deep learning tutorial's approach is that the Sunnybrook dataset had very few labeled samples (15 if I recall) granted it was created in 2009 before CNNets and deep learning were hot stuff, so using a model that was trained on that small data-set to extract LV segments on this huge data-set will produce inaccurate results in my opinion. Has a human cardiologist verified that the segments extracted by the deep learning tutorial are reasonable?  Using a model that was trained on 15 training samples to extract segments on a data-set with thousands of images is a little bit optimistic.\r\n\r\n[/quote]   \r\nI think the idea of the competition is that's we would have to do. Maybe a package was chosen but others would be relevant:  \r\nhttp://deeplearning.net/software_links/",
    "102117": "I have successfully climbed the wall and seen the tutorials. It's a lot of work and really awesome. \r\n\r\nIt doesn't use any information regarding the ground truth in train.csv, that's the reason why it performs bad.",
    "102205": "I am trying to wrap my head around the deep learning tutoral. Am i right to understand (and please correct me where wrong) that the input file that the tutoral uses already has labelled LV surfaces / \"blobs\" / areas for the LV? And subsequently, that the deep learning is merely trying to supervised learning to try to predict these blobs from the images / pixels? \r\n\r\nI have not downloaded the full dataset yet - but the real data-set has no labellled LV surfaces - no? Thus - how much use is the supervised learning approach in the tutoral?\r\n\r\nI feel that i am missing something - please correct me!",
    "102219": "You're description is correct",
    "102232": "Nevermind, there's a much faster and compact registration at the original MICCAI page here: \r\n\r\nhttp://smial.sri.utoronto.ca/LV_Challenge/Downloads.html\r\n\r\nRegister there and they give you a download link right away.\r\n\r\nHope this helps\r\n\r\n\r\n>>>\r\n\r\nEr, how are people getting access to the Sunnybrook dataset?  It appears to be behind a research registration wall – are they fairly permissive?\r\n\r\nhttp://www.cardiacatlas.org/data-access/request-cap-access/"
  },
  "source": "meta"
}