{
  "id": 7915,
  "title": "what is the goal?",
  "url": "/competitions/connectomics/discussion/7915",
  "author_name": "",
  "post_date": "2014-04-29T00:48:28.540Z",
  "votes": null,
  "comment_count": 6,
  "views": 1738,
  "content": "<p>Hi everyone, I just found out about the challenge. Certainly, an interesting problem! So my question is, it appears that they are already some existing algorithms that perform reasonably well (~90% accuracy). Seeing that we are dealing with a 1000 neuron network, I am wondering is &nbsp;this challenge about implementing these algorithms in an efficient manner (GPUs, cluster computing) or about coming up with new algorithms? &nbsp;If the latter was the case, why such a large network? It seems I can't even run this stuff on my laptop and I have no access to cluster computing.</p>",
  "messages": [
    {
      "id": "43254",
      "postDate": "04/29/2014 00:48:28",
      "content": "<p>Hi everyone, I just found out about the challenge. Certainly, an interesting problem! So my question is, it appears that they are already some existing algorithms that perform reasonably well (~90% accuracy). Seeing that we are dealing with a 1000 neuron network, I am wondering is &nbsp;this challenge about implementing these algorithms in an efficient manner (GPUs, cluster computing) or about coming up with new algorithms? &nbsp;If the latter was the case, why such a large network? It seems I can't even run this stuff on my laptop and I have no access to cluster computing.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "43277",
      "postDate": "04/29/2014 08:36:56",
      "content": "<p>Hi,</p>\n<p>It's more about the second actually. Also 0.9 might be misleading. Given that the network is sparse, an AUC of 0.9 could still mean that you only get half the links right!</p>\n<p>The size of the network was actually a compromise, some people wanted even bigger! Also, N around 1000 is the typical number of neurons we are able to record in experiments. We also provide a smaller dataset with N=100 for testing purposes.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "43367",
      "postDate": "04/30/2014 15:17:50",
      "content": "<p>Well, I don't think you need a cluster to run the sample code with 1,000 neurons. With some modification to the MATLAB GTE code provided&nbsp;by the organizer, it's possible to run the 1,000 neuron test set within 40 minutes on a Intel i7 2600 Desktop.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "44358",
      "postDate": "05/11/2014 04:31:51",
      "content": "<p>Hi there killertom,</p>\n<p>Are you willing to share your modified Matlab code for GTE? I would love to try it out.</p>\n<p>cheers,</p>\n<p>James</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "44360",
      "postDate": "05/11/2014 05:42:36",
      "content": "<p>Sure. Here you go.</p>\n<p>Please notice the difference in usage compared with the original GTE. You need to provide the discretized sequence and dynamical regime information to the GTEfast instead of providing only the original series. And GTEfast returns the GTE in all dynamical regimes, so a three-dimensional matrix.</p>\n<p>A simple demo would look like:</p>\n<p>[D2, G] = discretizeFluorescenceSignal(M, 'conditioningLevel', 0.25, 'bins', [-10,0.12,10]);</p>\n<p>tic;</p>\n<p>Nsamples = 50000;</p>\n<p>GTE = computeGTEfast(D2(1:Nsamples, :) ,G(1:Nsamples));</p>\n<p>toc;</p>\n\n<p>Please inform me if you have any trouble running this code.&nbsp;</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "44364",
      "postDate": "05/11/2014 09:37:50",
      "content": "<p>Thanks.</p>\n<p>Regarding the ~40 minute runtime you mentioned previously, is that for the full dataset (i.e. ~180k rows)?</p>\n<p>James.</p>",
      "rawMarkdown": "",
      "votes": null
    },
    {
      "id": "44371",
      "postDate": "05/11/2014 11:57:32",
      "content": "<p>Hi James,&nbsp;</p>\n<p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp;40 mins is for the full-size 1,000 neuron dataset. (the size of valid/test, normal 1-4)</p>",
      "rawMarkdown": "",
      "votes": null
    }
  ],
  "comments": [
    {
      "id": 43277,
      "author_name": "orlandi",
      "author_url": "",
      "post_date": "04/29/2014 08:36:56",
      "content": "<p>Hi,</p>\n<p>It's more about the second actually. Also 0.9 might be misleading. Given that the network is sparse, an AUC of 0.9 could still mean that you only get half the links right!</p>\n<p>The size of the network was actually a compromise, some people wanted even bigger! Also, N around 1000 is the typical number of neurons we are able to record in experiments. We also provide a smaller dataset with N=100 for testing purposes.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 43367,
      "author_name": "killertom",
      "author_url": "",
      "post_date": "04/30/2014 15:17:50",
      "content": "<p>Well, I don't think you need a cluster to run the sample code with 1,000 neurons. With some modification to the MATLAB GTE code provided&nbsp;by the organizer, it's possible to run the 1,000 neuron test set within 40 minutes on a Intel i7 2600 Desktop.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 44358,
      "author_name": "jamestee1",
      "author_url": "",
      "post_date": "05/11/2014 04:31:51",
      "content": "<p>Hi there killertom,</p>\n<p>Are you willing to share your modified Matlab code for GTE? I would love to try it out.</p>\n<p>cheers,</p>\n<p>James</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 44360,
      "author_name": "killertom",
      "author_url": "",
      "post_date": "05/11/2014 05:42:36",
      "content": "<p>Sure. Here you go.</p>\n<p>Please notice the difference in usage compared with the original GTE. You need to provide the discretized sequence and dynamical regime information to the GTEfast instead of providing only the original series. And GTEfast returns the GTE in all dynamical regimes, so a three-dimensional matrix.</p>\n<p>A simple demo would look like:</p>\n<p>[D2, G] = discretizeFluorescenceSignal(M, 'conditioningLevel', 0.25, 'bins', [-10,0.12,10]);</p>\n<p>tic;</p>\n<p>Nsamples = 50000;</p>\n<p>GTE = computeGTEfast(D2(1:Nsamples, :) ,G(1:Nsamples));</p>\n<p>toc;</p>\n\n<p>Please inform me if you have any trouble running this code.&nbsp;</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 44364,
      "author_name": "jamestee1",
      "author_url": "",
      "post_date": "05/11/2014 09:37:50",
      "content": "<p>Thanks.</p>\n<p>Regarding the ~40 minute runtime you mentioned previously, is that for the full dataset (i.e. ~180k rows)?</p>\n<p>James.</p>",
      "votes": null,
      "replies": []
    },
    {
      "id": 44371,
      "author_name": "killertom",
      "author_url": "",
      "post_date": "05/11/2014 11:57:32",
      "content": "<p>Hi James,&nbsp;</p>\n<p>&nbsp; &nbsp; &nbsp; &nbsp; &nbsp;40 mins is for the full-size 1,000 neuron dataset. (the size of valid/test, normal 1-4)</p>",
      "votes": null,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "43254": "",
    "43277": "",
    "43367": "",
    "44358": "",
    "44360": "",
    "44364": "",
    "44371": ""
  },
  "source": "meta"
}