See https://twitter.com/WonderMicky/status/1029547418369515520\n>
\n\nI answered this:
\n\n> I wonder why none of CERN affiliates entered that competition to show us how good their approach is. Afraid of something?
\n\nThe discussion goes on on twitter if you're interested.
\n\nEdited as the twitter discussion evolves more positively than I thought: https://twitter.com/WonderMicky/status/1029604119810469888
\n\nA side effect is that we now know that #3 and #4 solutions are from specialists of the topic:
\n\nhttps://twitter.com/dhpmrou/status/1029602591750144000
", "messages": [ { "id": "370596", "postDate": "08/15/2018 05:32:23", "content": "See https://twitter.com/WonderMicky/status/1029547418369515520\n>
\n\nI answered this:
\n\n> I wonder why none of CERN affiliates entered that competition to show us how good their approach is. Afraid of something?
\n\nThe discussion goes on on twitter if you're interested.
\n\nEdited as the twitter discussion evolves more positively than I thought: https://twitter.com/WonderMicky/status/1029604119810469888
\n\nA side effect is that we now know that #3 and #4 solutions are from specialists of the topic:
\n\nhttps://twitter.com/dhpmrou/status/1029602591750144000
", "rawMarkdown": "See https://twitter.com/WonderMicky/status/1029547418369515520\n>\n\nI answered this:\n\n> I wonder why none of CERN affiliates entered that competition to show us how good their approach is. Afraid of something? \n\nThe discussion goes on on twitter if you're interested.\n\nEdited as the twitter discussion evolves more positively than I thought: https://twitter.com/WonderMicky/status/1029604119810469888\n\nA side effect is that we now know that #3 and #4 solutions are from specialists of the topic: \n\nhttps://twitter.com/dhpmrou/status/1029602591750144000", "votes": null }, { "id": "370643", "postDate": "08/15/2018 07:35:42", "content": "\n\n\nA side effect is that we now know that #3 and #4 solutions are from specialists
\n
That's interesting. I'm burning with curiosity who @demelian is. :-) Too bad he did not share (yet?).
", "rawMarkdown": "> A side effect is that we now know that #3 and #4 solutions are from specialists\n\nThat's interesting. I'm burning with curiosity who @demelian is. :-) Too bad he did not share (yet?).", "votes": null }, { "id": "370660", "postDate": "08/15/2018 08:25:44", "content": "In a nutshell - this competition is about accuracy, not throughput, and @Edwin's code is faster than what CERN has right now. I think it's unfair to criticize @outrunner's solution that takes days or DBSCAN solutions that can take hours. I'm sure in the 2nd phase competition people will start with the approach similar to @icecuber's, @sgorbuno (Sergey's) or @edwinst (Edwin's).
\n\nMy 2 cents again:\nI thought @Edwin is a specialist as well, from this post https://profmattstrassler.com/articles-and-posts/particle-physics-basics/quantum-fluctuations-and-their-energy/ if that's the same Edwin :) We knew this a couple months ago when Edwin came out from nowhere as a Kaggle newbie. And I was thrilled to know he's from DACH as well.
\n\nEveryone joins Kaggle competition with different skills, some do better in one and worse in others. When we join this competition, we accept this disadvantage. Like @bestfitting, @outrunner, or you (@CPMP) definitely have an edge over the 2nd time Kagglers like us, but we accepted the challenge. I think it's very fair. We all started as a Kaggle newbie. You never know if a newbie will become the next @bestfitting. :)
", "rawMarkdown": "In a nutshell - this competition is about accuracy, not throughput, and @Edwin's code is faster than what CERN has right now. I think it's unfair to criticize @outrunner's solution that takes days or DBSCAN solutions that can take hours. I'm sure in the 2nd phase competition people will start with the approach similar to @icecuber's, @sgorbuno (Sergey's) or @edwinst (Edwin's). \n\n**My 2 cents again:**\nI thought @Edwin is a specialist as well, from this post https://profmattstrassler.com/articles-and-posts/particle-physics-basics/quantum-fluctuations-and-their-energy/ if that's the same Edwin :) We knew this a couple months ago when Edwin came out from nowhere as a Kaggle newbie. And I was thrilled to know he's from DACH as well.\n\nEveryone joins Kaggle competition with different skills, some do better in one and worse in others. When we join this competition, we accept this disadvantage. Like @bestfitting, @outrunner, or you (@CPMP) definitely have an edge over the 2nd time Kagglers like us, but we accepted the challenge. I think it's very fair. We all started as a Kaggle newbie. You never know if a newbie will become the next @bestfitting. :)", "votes": null }, { "id": "370683", "postDate": "08/15/2018 09:32:55", "content": "\n\n\nif that's the same Edwin :)
\n
Yes, that's me. I'm not a professional physicist but I care a lot about physics and have acquired some beginning technical understanding of topics like quantum field theory, and I hope to learn more. My professional life is as a software developer.
\n\n\n\n\nhe's from DACH as well
\n
Aus dem wunderschönen Österreich um genauer zu sein. :-)
\n\n\n\n\nEveryone joins Kaggle competition with different skills
\n
Yes, and I wish I'd had more (than zero) experience with ML as I started. Even though the top solution is probably quite similar to the current approach in HEP, I think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for ML globally, because the only true correlations in the data are the well-known physical laws of relativistic particles in a magnetic field, and precision is required to get them right. Everything else is basically true randomness.
\n\nHowever, that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems. The best example is @icecuber's smart use of logistic regression for truncating the list of hit pairs to use as seeds. There are many other opportunities for sure, e.g. in the assigning of hits to candidate tracks.
", "rawMarkdown": "> if that's the same Edwin :)\n\nYes, that's me. I'm not a professional physicist but I care a lot about physics and have acquired some beginning technical understanding of topics like quantum field theory, and I hope to learn more. My professional life is as a software developer.\n\n> he's from DACH as well\n\nAus dem wunderschönen Österreich um genauer zu sein. :-)\n\n> Everyone joins Kaggle competition with different skills\n\nYes, and I wish I'd had more (than zero) experience with ML as I started. Even though the top solution is probably quite similar to the current approach in HEP, I think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for ML *globally*, because the only true correlations in the data are the well-known physical laws of relativistic particles in a magnetic field, and precision is required to get them right. Everything else is basically *true* randomness.\n\n*However*, that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems. The best example is @icecuber's smart use of logistic regression for truncating the list of hit pairs to use as seeds. There are many other opportunities for sure, e.g. in the assigning of hits to candidate tracks.", "votes": null }, { "id": "370727", "postDate": "08/15/2018 11:00:17", "content": "Thanks! That's the interesting discussion.
", "rawMarkdown": "Thanks! That's the interesting discussion.", "votes": null }, { "id": "370737", "postDate": "08/15/2018 11:19:18", "content": "\n\n\nAus dem wunderschönen Österreich um genauer zu sein. :-)
\n
Totally agree, my profile picture suggests I spend most of my vacation days in Austria. :D
\n\n\n\n\nI think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for \n ML globally, ...... that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems.
\n
A very good write-up. @andri27 (Andrea Lonza #11) also improved his score from 0.68 to 0.76 using a supervised learning approach (decision tree) to extend tracks, which takes 20 minutes but there's always room for optimization. This is inspiring. @outrunner also showed us finding tracks using NN that alone can reach 0.8. Just like CERN was trying LSTM approaches to replace expensive Kalman filters, which probably is the bottleneck of performance, that's what a ML approach can tackle.
", "rawMarkdown": "> Aus dem wunderschönen Österreich um genauer zu sein. :-)\n\nTotally agree, my profile picture suggests I spend most of my vacation days in Austria. :D\n\n> I think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for \n> ML globally, ...... that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems.\n\nA very good write-up. @andri27 (Andrea Lonza #11) also improved his score from 0.68 to 0.76 using a supervised learning approach (decision tree) to extend tracks, which takes 20 minutes but there's always room for optimization. This is inspiring. @outrunner also showed us finding tracks using NN that alone can reach 0.8. Just like CERN was trying LSTM approaches to replace expensive Kalman filters, which probably is the bottleneck of performance, that's what a ML approach can tackle.", "votes": null }, { "id": "370763", "postDate": "08/15/2018 12:24:28", "content": "I found this facebook page but I doubt it is relevant :)
", "rawMarkdown": "I found t[his facebook page][1] but I doubt it is relevant :)\n\n\n [1]: https://www.facebook.com/DeMelian-147470485919697/", "votes": null }, { "id": "370765", "postDate": "08/15/2018 12:29:00", "content": "It’s a shame neither Kyle or Micky entered the competition and shared their knowledge in the forums. It would have been a lot more helpful than dismissive and condescending tweets.
", "rawMarkdown": "It’s a shame neither Kyle or Micky entered the competition and shared their knowledge in the forums. It would have been a lot more helpful than dismissive and condescending tweets.", "votes": null }, { "id": "370851", "postDate": "08/15/2018 14:50:09", "content": "I would not put them both in the same basket. Kyle provided some facts, while Micky snorted. Big difference.
\n\nAnd it is fine questioning the relevance of this to high physics research, see Edwin's great post about it.
\n\nQuestioning is part of the scientific process. But it has to be done scientifically, looking at facts and analyzing data, not making bold statements without any supportive information.
", "rawMarkdown": "I would not put them both in the same basket. Kyle provided some facts, while Micky snorted. Big difference. \n\nAnd it is fine questioning the relevance of this to high physics research, see Edwin's [great post about it.][1] \n\nQuestioning is part of the scientific process. But it has to be done scientifically, looking at facts and analyzing data, not making bold statements without any supportive information.\n\n\n [1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/63368#latest-370768", "votes": null }, { "id": "370895", "postDate": "08/15/2018 16:02:14", "content": "That's fair, I shouldn't lump them in together. Kyle's criticism of not having a time metric to the competitions is a valid point and I would like to see that integrated into the platform somehow. It looks like there is a move towards adding a speed component with the Airbus comp that is currently active, maybe there have been previous competitions involving a speed component but I haven't been on here that long so that is the only one I am aware of.
", "rawMarkdown": "That's fair, I shouldn't lump them in together. Kyle's criticism of not having a time metric to the competitions is a valid point and I would like to see that integrated into the platform somehow. It looks like there is a move towards adding a speed component with the Airbus comp that is currently active, maybe there have been previous competitions involving a speed component but I haven't been on here that long so that is the only one I am aware of.", "votes": null }, { "id": "370907", "postDate": "08/15/2018 16:21:14", "content": "It is not really a valid point as we've always advertised that this competition on Kaggle will be followed by a second phase with a strong cpu incentive Point 3 https://www.kaggle.com/c/trackml-particle-identification/discussion/63289 \nNote that we did discuss with kaggle the possibility to have the second phase on kaggle as well. They are looking into this, but not in the time scale we needed.
", "rawMarkdown": "It is not really a valid point as we've always advertised that this competition on Kaggle will be followed by a second phase with a strong cpu incentive Point 3 https://www.kaggle.com/c/trackml-particle-identification/discussion/63289 \nNote that we did discuss with kaggle the possibility to have the second phase on kaggle as well. They are looking into this, but not in the time scale we needed.", "votes": null }, { "id": "370918", "postDate": "08/15/2018 16:46:16", "content": "I was referring to this part of the thread, second sentence in particular:
\n\n\n\n\nI wasn't criticizing... I'm just pointing to the challenge and the\n large gap to be closed. It's also interesting that if you don't\n include time in the metric that you get solutions that may never be\n candidates for solving the real problem because of the algorithmic\n complexity.
\n
I read it as a general criticism rather than specific to this competition.
", "rawMarkdown": "I was referring to this part of the thread, second sentence in particular:\n\n> I wasn't criticizing... I'm just pointing to the challenge and the\n> large gap to be closed. It's also interesting that if you don't\n> include time in the metric that you get solutions that may never be\n> candidates for solving the real problem because of the algorithmic\n> complexity.\n\nI read it as a general criticism rather than specific to this competition.", "votes": null }, { "id": "371286", "postDate": "08/16/2018 13:05:02", "content": "And now we learn that #1 works at/with CERN, see https://twitter.com/pietrovischia/status/1030048246918533121
\n\nThis makes @outrunner performance even more impressive, he is the only non specialist among top 4.
\n\nEdit: this was wrong, see https://twitter.com/alheld_/status/1030088014784028672
", "rawMarkdown": "And now we learn that #1 works at/with CERN, see https://twitter.com/pietrovischia/status/1030048246918533121\n\nThis makes @outrunner performance even more impressive, he is the only non specialist among top 4.\n\nEdit: this was wrong, see https://twitter.com/alheld_/status/1030088014784028672", "votes": null } ], "comments": [ { "id": 370643, "author_name": "edwinst", "author_url": "", "post_date": "08/15/2018 07:35:42", "content": "\n\n\nA side effect is that we now know that #3 and #4 solutions are from specialists
\n
That's interesting. I'm burning with curiosity who @demelian is. :-) Too bad he did not share (yet?).
", "votes": null, "replies": [ { "id": 370763, "author_name": "cpmpml", "author_url": "", "post_date": "08/15/2018 12:24:28", "content": "I found this facebook page but I doubt it is relevant :)
", "votes": null, "replies": [] } ] }, { "id": 370660, "author_name": "nicolefinnie", "author_url": "", "post_date": "08/15/2018 08:25:44", "content": "In a nutshell - this competition is about accuracy, not throughput, and @Edwin's code is faster than what CERN has right now. I think it's unfair to criticize @outrunner's solution that takes days or DBSCAN solutions that can take hours. I'm sure in the 2nd phase competition people will start with the approach similar to @icecuber's, @sgorbuno (Sergey's) or @edwinst (Edwin's).
\n\nMy 2 cents again:\nI thought @Edwin is a specialist as well, from this post https://profmattstrassler.com/articles-and-posts/particle-physics-basics/quantum-fluctuations-and-their-energy/ if that's the same Edwin :) We knew this a couple months ago when Edwin came out from nowhere as a Kaggle newbie. And I was thrilled to know he's from DACH as well.
\n\nEveryone joins Kaggle competition with different skills, some do better in one and worse in others. When we join this competition, we accept this disadvantage. Like @bestfitting, @outrunner, or you (@CPMP) definitely have an edge over the 2nd time Kagglers like us, but we accepted the challenge. I think it's very fair. We all started as a Kaggle newbie. You never know if a newbie will become the next @bestfitting. :)
", "votes": null, "replies": [ { "id": 370683, "author_name": "edwinst", "author_url": "", "post_date": "08/15/2018 09:32:55", "content": "\n\n\nif that's the same Edwin :)
\n
Yes, that's me. I'm not a professional physicist but I care a lot about physics and have acquired some beginning technical understanding of topics like quantum field theory, and I hope to learn more. My professional life is as a software developer.
\n\n\n\n\nhe's from DACH as well
\n
Aus dem wunderschönen Österreich um genauer zu sein. :-)
\n\n\n\n\nEveryone joins Kaggle competition with different skills
\n
Yes, and I wish I'd had more (than zero) experience with ML as I started. Even though the top solution is probably quite similar to the current approach in HEP, I think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for ML globally, because the only true correlations in the data are the well-known physical laws of relativistic particles in a magnetic field, and precision is required to get them right. Everything else is basically true randomness.
\n\nHowever, that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems. The best example is @icecuber's smart use of logistic regression for truncating the list of hit pairs to use as seeds. There are many other opportunities for sure, e.g. in the assigning of hits to candidate tracks.
", "votes": null, "replies": [] }, { "id": 370737, "author_name": "nicolefinnie", "author_url": "", "post_date": "08/15/2018 11:19:18", "content": "\n\n\nAus dem wunderschönen Österreich um genauer zu sein. :-)
\n
Totally agree, my profile picture suggests I spend most of my vacation days in Austria. :D
\n\n\n\n\nI think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for \n ML globally, ...... that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems.
\n
A very good write-up. @andri27 (Andrea Lonza #11) also improved his score from 0.68 to 0.76 using a supervised learning approach (decision tree) to extend tracks, which takes 20 minutes but there's always room for optimization. This is inspiring. @outrunner also showed us finding tracks using NN that alone can reach 0.8. Just like CERN was trying LSTM approaches to replace expensive Kalman filters, which probably is the bottleneck of performance, that's what a ML approach can tackle.
", "votes": null, "replies": [] } ] }, { "id": 370727, "author_name": "sergeyzlobin", "author_url": "", "post_date": "08/15/2018 11:00:17", "content": "Thanks! That's the interesting discussion.
", "votes": null, "replies": [] }, { "id": 370765, "author_name": "jackvial", "author_url": "", "post_date": "08/15/2018 12:29:00", "content": "It’s a shame neither Kyle or Micky entered the competition and shared their knowledge in the forums. It would have been a lot more helpful than dismissive and condescending tweets.
", "votes": null, "replies": [ { "id": 370851, "author_name": "cpmpml", "author_url": "", "post_date": "08/15/2018 14:50:09", "content": "I would not put them both in the same basket. Kyle provided some facts, while Micky snorted. Big difference.
\n\nAnd it is fine questioning the relevance of this to high physics research, see Edwin's great post about it.
\n\nQuestioning is part of the scientific process. But it has to be done scientifically, looking at facts and analyzing data, not making bold statements without any supportive information.
", "votes": null, "replies": [] }, { "id": 370895, "author_name": "jackvial", "author_url": "", "post_date": "08/15/2018 16:02:14", "content": "That's fair, I shouldn't lump them in together. Kyle's criticism of not having a time metric to the competitions is a valid point and I would like to see that integrated into the platform somehow. It looks like there is a move towards adding a speed component with the Airbus comp that is currently active, maybe there have been previous competitions involving a speed component but I haven't been on here that long so that is the only one I am aware of.
", "votes": null, "replies": [] }, { "id": 370907, "author_name": "droussea", "author_url": "", "post_date": "08/15/2018 16:21:14", "content": "It is not really a valid point as we've always advertised that this competition on Kaggle will be followed by a second phase with a strong cpu incentive Point 3 https://www.kaggle.com/c/trackml-particle-identification/discussion/63289 \nNote that we did discuss with kaggle the possibility to have the second phase on kaggle as well. They are looking into this, but not in the time scale we needed.
", "votes": null, "replies": [] }, { "id": 370918, "author_name": "jackvial", "author_url": "", "post_date": "08/15/2018 16:46:16", "content": "I was referring to this part of the thread, second sentence in particular:
\n\n\n\n\nI wasn't criticizing... I'm just pointing to the challenge and the\n large gap to be closed. It's also interesting that if you don't\n include time in the metric that you get solutions that may never be\n candidates for solving the real problem because of the algorithmic\n complexity.
\n
I read it as a general criticism rather than specific to this competition.
", "votes": null, "replies": [] } ] }, { "id": 371286, "author_name": "cpmpml", "author_url": "", "post_date": "08/16/2018 13:05:02", "content": "And now we learn that #1 works at/with CERN, see https://twitter.com/pietrovischia/status/1030048246918533121
\n\nThis makes @outrunner performance even more impressive, he is the only non specialist among top 4.
\n\nEdit: this was wrong, see https://twitter.com/alheld_/status/1030088014784028672
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "370596": "See https://twitter.com/WonderMicky/status/1029547418369515520\n>\n\nI answered this:\n\n> I wonder why none of CERN affiliates entered that competition to show us how good their approach is. Afraid of something? \n\nThe discussion goes on on twitter if you're interested.\n\nEdited as the twitter discussion evolves more positively than I thought: https://twitter.com/WonderMicky/status/1029604119810469888\n\nA side effect is that we now know that #3 and #4 solutions are from specialists of the topic: \n\nhttps://twitter.com/dhpmrou/status/1029602591750144000", "370643": "> A side effect is that we now know that #3 and #4 solutions are from specialists\n\nThat's interesting. I'm burning with curiosity who @demelian is. :-) Too bad he did not share (yet?).", "370660": "In a nutshell - this competition is about accuracy, not throughput, and @Edwin's code is faster than what CERN has right now. I think it's unfair to criticize @outrunner's solution that takes days or DBSCAN solutions that can take hours. I'm sure in the 2nd phase competition people will start with the approach similar to @icecuber's, @sgorbuno (Sergey's) or @edwinst (Edwin's). \n\n**My 2 cents again:**\nI thought @Edwin is a specialist as well, from this post https://profmattstrassler.com/articles-and-posts/particle-physics-basics/quantum-fluctuations-and-their-energy/ if that's the same Edwin :) We knew this a couple months ago when Edwin came out from nowhere as a Kaggle newbie. And I was thrilled to know he's from DACH as well.\n\nEveryone joins Kaggle competition with different skills, some do better in one and worse in others. When we join this competition, we accept this disadvantage. Like @bestfitting, @outrunner, or you (@CPMP) definitely have an edge over the 2nd time Kagglers like us, but we accepted the challenge. I think it's very fair. We all started as a Kaggle newbie. You never know if a newbie will become the next @bestfitting. :)", "370683": "> if that's the same Edwin :)\n\nYes, that's me. I'm not a professional physicist but I care a lot about physics and have acquired some beginning technical understanding of topics like quantum field theory, and I hope to learn more. My professional life is as a software developer.\n\n> he's from DACH as well\n\nAus dem wunderschönen Österreich um genauer zu sein. :-)\n\n> Everyone joins Kaggle competition with different skills\n\nYes, and I wish I'd had more (than zero) experience with ML as I started. Even though the top solution is probably quite similar to the current approach in HEP, I think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for ML *globally*, because the only true correlations in the data are the well-known physical laws of relativistic particles in a magnetic field, and precision is required to get them right. Everything else is basically *true* randomness.\n\n*However*, that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems. The best example is @icecuber's smart use of logistic regression for truncating the list of hit pairs to use as seeds. There are many other opportunities for sure, e.g. in the assigning of hits to candidate tracks.", "370727": "Thanks! That's the interesting discussion.", "370737": "> Aus dem wunderschönen Österreich um genauer zu sein. :-)\n\nTotally agree, my profile picture suggests I spend most of my vacation days in Austria. :D\n\n> I think it would be too early to discard ML. My opinion is that the trackml problem is not a good fit for \n> ML globally, ...... that doesn't mean algorithms cannot greatly benefit from ML in solving sub-problems.\n\nA very good write-up. @andri27 (Andrea Lonza #11) also improved his score from 0.68 to 0.76 using a supervised learning approach (decision tree) to extend tracks, which takes 20 minutes but there's always room for optimization. This is inspiring. @outrunner also showed us finding tracks using NN that alone can reach 0.8. Just like CERN was trying LSTM approaches to replace expensive Kalman filters, which probably is the bottleneck of performance, that's what a ML approach can tackle.", "370763": "I found t[his facebook page][1] but I doubt it is relevant :)\n\n\n [1]: https://www.facebook.com/DeMelian-147470485919697/", "370765": "It’s a shame neither Kyle or Micky entered the competition and shared their knowledge in the forums. It would have been a lot more helpful than dismissive and condescending tweets.", "370851": "I would not put them both in the same basket. Kyle provided some facts, while Micky snorted. Big difference. \n\nAnd it is fine questioning the relevance of this to high physics research, see Edwin's [great post about it.][1] \n\nQuestioning is part of the scientific process. But it has to be done scientifically, looking at facts and analyzing data, not making bold statements without any supportive information.\n\n\n [1]: https://www.kaggle.com/c/trackml-particle-identification/discussion/63368#latest-370768", "370895": "That's fair, I shouldn't lump them in together. Kyle's criticism of not having a time metric to the competitions is a valid point and I would like to see that integrated into the platform somehow. It looks like there is a move towards adding a speed component with the Airbus comp that is currently active, maybe there have been previous competitions involving a speed component but I haven't been on here that long so that is the only one I am aware of.", "370907": "It is not really a valid point as we've always advertised that this competition on Kaggle will be followed by a second phase with a strong cpu incentive Point 3 https://www.kaggle.com/c/trackml-particle-identification/discussion/63289 \nNote that we did discuss with kaggle the possibility to have the second phase on kaggle as well. They are looking into this, but not in the time scale we needed.", "370918": "I was referring to this part of the thread, second sentence in particular:\n\n> I wasn't criticizing... I'm just pointing to the challenge and the\n> large gap to be closed. It's also interesting that if you don't\n> include time in the metric that you get solutions that may never be\n> candidates for solving the real problem because of the algorithmic\n> complexity.\n\nI read it as a general criticism rather than specific to this competition.", "371286": "And now we learn that #1 works at/with CERN, see https://twitter.com/pietrovischia/status/1030048246918533121\n\nThis makes @outrunner performance even more impressive, he is the only non specialist among top 4.\n\nEdit: this was wrong, see https://twitter.com/alheld_/status/1030088014784028672" }, "source": "meta" }