![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F644036%2Fe0406162927cdec7c1db546d0d957acc%2Flb_Tr.png?generation=1564930492180256&alt=media\")
It's clear one team has pulled ahead from the rest on the public leaderboard. Certainly it's possible that other teams have similarly strong models but are just holding back on submitting. Apart from that, right now it appears they are blowing the rest of the field away.
\n\nAre there theories as to how one team could be so much stronger than the rest?
\n\n
![\"\"](\"https://i.imgur.com/d4z0vdl.gif\")
It's clear one team has pulled ahead from the rest on the public leaderboard. Certainly it's possible that other teams have similarly strong models but are just holding back on submitting. Apart from that, right now it appears they are blowing the rest of the field away.
\n\nAre there theories as to how one team could be so much stronger than the rest?
\n\n
\n\n\nAre there theories as to how one team could be so much stronger than the rest?
\n
You provided the answer:
\n\n\n", "rawMarkdown": "> Are there theories as to how one team could be so much stronger than the rest?\n\nYou provided the answer:\n\n> Certainly it's possible that other teams have similarly strong models but are just holding back on submitting.", "votes": null }, { "id": "591990", "postDate": "08/04/2019 15:37:30", "content": "Certainly it's possible that other teams have similarly strong models but are just holding back on submitting.
\n
We have developed a quantum computer and achieved quantum supremacy.
", "rawMarkdown": "We have developed a quantum computer and achieved quantum supremacy.", "votes": null }, { "id": "591999", "postDate": "08/04/2019 15:50:01", "content": "Great, you'll be removed then because you use quantum computation :D
", "rawMarkdown": "Great, you'll be removed then because you use quantum computation :D", "votes": null }, { "id": "592068", "postDate": "08/04/2019 18:39:33", "content": "Maybe they ensemble models?
", "rawMarkdown": "Maybe they ensemble models?", "votes": null }, { "id": "592084", "postDate": "08/04/2019 19:20:20", "content": "Out of curiosity, for those of you who has average log MAE less than -2.5 (by which I assume is achieved by ensembling different types of models), which one is better 1JHC and 1JHN? GB-based (LGB XGB) or NN-based (NN, GNN, GCN)?
", "rawMarkdown": "Out of curiosity, for those of you who has average log MAE less than -2.5 (by which I assume is achieved by ensembling different types of models), which one is better 1JHC and 1JHN? GB-based (LGB XGB) or NN-based (NN, GNN, GCN)?", "votes": null }, { "id": "592458", "postDate": "08/05/2019 10:16:11", "content": "It seems some poeple were motivated to close some of the gap after 1st team...
", "rawMarkdown": "It seems some poeple were motivated to close some of the gap after 1st team...", "votes": null }, { "id": "592479", "postDate": "08/05/2019 11:07:41", "content": "We basically work with one model for now, hence cannot really answer your question ;)
\n\nI would be surprised if lgb is better than NNs here.
", "rawMarkdown": "We basically work with one model for now, hence cannot really answer your question ;) \n\nI would be surprised if lgb is better than NNs here.", "votes": null }, { "id": "592601", "postDate": "08/05/2019 14:37:02", "content": "or to be ahead of uncle
", "rawMarkdown": "or to be ahead of uncle", "votes": null }, { "id": "592648", "postDate": "08/05/2019 16:00:39", "content": "ok, you convinced me to start a MPNN.
", "rawMarkdown": "ok, you convinced me to start a MPNN.", "votes": null }, { "id": "592713", "postDate": "08/05/2019 18:02:57", "content": "I really want to do quantum jump to top...:)
", "rawMarkdown": "I really want to do quantum jump to top...:)", "votes": null }, { "id": "592763", "postDate": "08/05/2019 19:37:59", "content": "Im rooting for you YouHan. you are always so close to gold. Good luck
", "rawMarkdown": "Im rooting for you YouHan. you are always so close to gold. Good luck", "votes": null }, { "id": "592826", "postDate": "08/05/2019 21:17:19", "content": "Pics or it didn't happen
", "rawMarkdown": "Pics or it didn't happen", "votes": null }, { "id": "593138", "postDate": "08/06/2019 08:08:37", "content": "As requested, here's a pic of our quantum computer. We're worried b/c it's getting AGI-like qualities and we don't know how to contain it:
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F8507%2Ffee4f2f20de135798baf02d6f6a41252%2Fall-your-atoms.png?generation=1565078913929451&alt=media\")
Thanks for kind words, Corey! \nI'm also rooting for you. I hope both of us get golds, then I became competition master, and you became so close to grand master! Good luck :)
", "rawMarkdown": "Thanks for kind words, Corey! \nI'm also rooting for you. I hope both of us get golds, then I became competition master, and you became so close to grand master! Good luck :)", "votes": null }, { "id": "593258", "postDate": "08/06/2019 11:24:55", "content": "\n\n\nWe're worried b/c it's getting AGI-like qualities and we don't know how to contain it
\n
AGI getting out of control? hmm\nIt is going to decide humankind's fate? hmm
\n\nThats a serious situation rather than try to fool it with adversarial samples and give humanity another chance, I would just pull the plug out.
\n\n![\"\"](\"https://i.imgflip.com/37gd4k.jpg\")
My theory: they use quantum tunneling to go through the -3.0 barrier.
", "rawMarkdown": "My theory: they use quantum tunneling to go through the -3.0 barrier.", "votes": null }, { "id": "594476", "postDate": "08/08/2019 04:53:32", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F644036%2F3b4a78877ca26fef9de98d987f1a68ea%2F37ib37.jpg?generation=1565239995447448&alt=media\")
Nice one!
", "rawMarkdown": "Nice one!", "votes": null }, { "id": "594511", "postDate": "08/08/2019 05:48:17", "content": "BTW I found it also interesting that there is a team named Walter White while another team is named Pinkman, and what is more interesting is that a team-member of Pinkman has Bojack horseman's character as his avatar (which was voiced by Pinkman's actor).
\n\nI think someone should form a team named DEA.
", "rawMarkdown": "BTW I found it also interesting that there is a team named Walter White while another team is named Pinkman, and what is more interesting is that a team-member of Pinkman has Bojack horseman's character as his avatar (which was voiced by Pinkman's actor). \n\nI think someone should form a team named DEA.", "votes": null }, { "id": "594519", "postDate": "08/08/2019 06:05:28", "content": "A bit off-topic but Todd Chavez is my favorite TV character. I started watching Breaking Bad much later and didn't even notice it is the same voice until a random google search.
\n\nIn this competition, there are a lot of cartels on top of us and we couldn't get our desired purity yet. It is not blue enough:)
", "rawMarkdown": "A bit off-topic but Todd Chavez is my favorite TV character. I started watching Breaking Bad much later and didn't even notice it is the same voice until a random google search.\n\nIn this competition, there are a lot of cartels on top of us and we couldn't get our desired purity yet. It is not blue enough:)", "votes": null }, { "id": "595056", "postDate": "08/08/2019 19:42:08", "content": "I agree. \"4 GM and the brain\" sounds pretty villain'ish.
", "rawMarkdown": "I agree. \"4 GM and the brain\" sounds pretty villain'ish.", "votes": null }, { "id": "602896", "postDate": "08/19/2019 16:07:22", "content": "\n\n\nor to be ahead of uncle
\n
Seems this motivates you indeed ;)
", "rawMarkdown": "> or to be ahead of uncle\n\nSeems this motivates you indeed ;)", "votes": null }, { "id": "602898", "postDate": "08/19/2019 16:10:14", "content": "Seems like you found someone who can handle GNN ;)
", "rawMarkdown": "Seems like you found someone who can handle GNN ;)", "votes": null }, { "id": "602900", "postDate": "08/19/2019 16:12:19", "content": "We still miss a quantum chemist and a fourth scientist.
", "rawMarkdown": "We still miss a quantum chemist and a fourth scientist.", "votes": null }, { "id": "605884", "postDate": "08/22/2019 22:33:20", "content": "Maybe the top team is primarily training on the first 29 % of the data
", "rawMarkdown": "Maybe the top team is primarily training on the first 29 % of the data", "votes": null }, { "id": "605897", "postDate": "08/22/2019 23:21:41", "content": "Why do you think so ? Do you revealed the way to devide public/private dataset ?
", "rawMarkdown": "Why do you think so ? Do you revealed the way to devide public/private dataset ?", "votes": null }, { "id": "609271", "postDate": "08/27/2019 14:53:25", "content": "The gap is being closed, and there is a continuum of scores now. Largest gap is only a out 0.085. It will be very interesting to read writeups.
", "rawMarkdown": "The gap is being closed, and there is a continuum of scores now. Largest gap is only a out 0.085. It will be very interesting to read writeups.", "votes": null }, { "id": "610444", "postDate": "08/28/2019 20:02:14", "content": "The gap has been closed!
\n\n![\"\"](\"https://www.kaggleusercontent.com/kf/19738636/eyJhbGciOiJkaXIiLCJlbmMiOiJBMTI4Q0JDLUhTMjU2In0..Ttx4j8D1MPHOs-nXPY_Fpw.QH8itByxYz9gXyNluODXmmn8fXr5CPARbK6kXtA2Dq8gfPj5FdeK9CgS4EN-ULpSX2jjSiuYW54QNzMLa5msD0nJybURlui10js1r4sOG6Hpyrq-udXIfpi8qzFPqMDQWnM-3hr3Shs0lcvuuVak7-4hgMTw-j4IRFMI3HRRmes.mFGC2H25HBZlpuOfQLu2Kg/__results___files/__results___10_0.png\")
Well it's over. I can put my computer to sleep ;-)
\n\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F8507%2F143bc3851ec23a95417ba91642250e88%2FScreen%20Shot%202019-08-29%20at%202.07.46%20AM.png?generation=1567037287027219&alt=media\")
This is very fitting 😄
", "rawMarkdown": "This is very fitting 😄", "votes": null }, { "id": "615715", "postDate": "09/02/2019 09:35:18", "content": "> Seems like you found someone who can handle GNN ;)
\n\nOutrunner NN is not a graph NN by any mean actually. I couldn't say it during the competition obviously ;)
\n\nOur models are closer in spirit to 2nd team, except twe use distance and angles instead of xyz coordinates. I'm still puzzled that the order of atom index and xyz coordinates are significant given how model that use it outperform models that are rotation and atom order independent.
", "rawMarkdown": "> Seems like you found someone who can handle GNN ;)\n\nOutrunner NN is not a graph NN by any mean actually. I couldn't say it during the competition obviously ;) \n\nOur models are closer in spirit to 2nd team, except twe use distance and angles instead of xyz coordinates. I'm still puzzled that the order of atom index and xyz coordinates are significant given how model that use it outperform models that are rotation and atom order independent.", "votes": null }, { "id": "615719", "postDate": "09/02/2019 09:38:56", "content": "We do not use atom index (position in the array) 9999?9
", "rawMarkdown": "We do not use atom index (position in the array) 9999?9", "votes": null }, { "id": "615755", "postDate": "09/02/2019 10:36:34", "content": "3rd team uses the atoms as a sequence.
", "rawMarkdown": "3rd team uses the atoms as a sequence.", "votes": null }, { "id": "615764", "postDate": "09/02/2019 11:00:32", "content": "I bet if you remove the positional info their model performs equal or better
", "rawMarkdown": "I bet if you remove the positional info their model performs equal or better", "votes": null }, { "id": "615772", "postDate": "09/02/2019 11:06:55", "content": "Maybe. My point is that models that are rotation invariant and translation invariant and atom order invariant perform worse than those that miss one or more of these properties.
\n\nThis is surprising to me (and few others).
\n\nIt does not mean that the models that perform better are bad, to the contrary. They use data better than the other models.
\n\nHaving preconceived assumptions about what the models should be led to worse performance. Mind openness led to better performance.
", "rawMarkdown": "Maybe. My point is that models that are rotation invariant and translation invariant and atom order invariant perform worse than those that miss one or more of these properties. \n\nThis is surprising to me (and few others). \n\nIt does not mean that the models that perform better are bad, to the contrary. They use data better than the other models. \n\nHaving preconceived assumptions about what the models should be led to worse performance. Mind openness led to better performance.", "votes": null }, { "id": "615780", "postDate": "09/02/2019 11:24:21", "content": "![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-user-content/o/inbox%2F8507%2F187e12a764956e7e758b6c44514fe6db%2FIMG_20190902_132158.jpg?generation=1567423420659081&alt=media\")
\nFrom https://www.quantamagazine.org/how-artificial-intelligence-is-changing-science-20190311/
This was an inspiration to approach this challenge
", "rawMarkdown": "\nFrom https://www.quantamagazine.org/how-artificial-intelligence-is-changing-science-20190311/\n\nThis was an inspiration to approach this challenge", "votes": null } ], "comments": [ { "id": 591978, "author_name": "cpmpml", "author_url": "", "post_date": "08/04/2019 15:26:49", "content": "\n\n\nAre there theories as to how one team could be so much stronger than the rest?
\n
You provided the answer:
\n\n\n", "votes": null, "replies": [] }, { "id": 591990, "author_name": "antorsae", "author_url": "", "post_date": "08/04/2019 15:37:30", "content": "Certainly it's possible that other teams have similarly strong models but are just holding back on submitting.
\n
We have developed a quantum computer and achieved quantum supremacy.
", "votes": null, "replies": [ { "id": 591999, "author_name": "cpmpml", "author_url": "", "post_date": "08/04/2019 15:50:01", "content": "Great, you'll be removed then because you use quantum computation :D
", "votes": null, "replies": [] }, { "id": 592826, "author_name": "summationnotations", "author_url": "", "post_date": "08/05/2019 21:17:19", "content": "Pics or it didn't happen
", "votes": null, "replies": [] }, { "id": 593138, "author_name": "antorsae", "author_url": "", "post_date": "08/06/2019 08:08:37", "content": "As requested, here's a pic of our quantum computer. We're worried b/c it's getting AGI-like qualities and we don't know how to contain it:
\n\n
\n\n\nWe're worried b/c it's getting AGI-like qualities and we don't know how to contain it
\n
AGI getting out of control? hmm\nIt is going to decide humankind's fate? hmm
\n\nThats a serious situation rather than try to fool it with adversarial samples and give humanity another chance, I would just pull the plug out.
\n\nMaybe they ensemble models?
", "votes": null, "replies": [] }, { "id": 592084, "author_name": "scaomath", "author_url": "", "post_date": "08/04/2019 19:20:20", "content": "Out of curiosity, for those of you who has average log MAE less than -2.5 (by which I assume is achieved by ensembling different types of models), which one is better 1JHC and 1JHN? GB-based (LGB XGB) or NN-based (NN, GNN, GCN)?
", "votes": null, "replies": [ { "id": 592479, "author_name": "cpmpml", "author_url": "", "post_date": "08/05/2019 11:07:41", "content": "We basically work with one model for now, hence cannot really answer your question ;)
\n\nI would be surprised if lgb is better than NNs here.
", "votes": null, "replies": [] } ] }, { "id": 592458, "author_name": "cpmpml", "author_url": "", "post_date": "08/05/2019 10:16:11", "content": "It seems some poeple were motivated to close some of the gap after 1st team...
", "votes": null, "replies": [ { "id": 592601, "author_name": "christofhenkel", "author_url": "", "post_date": "08/05/2019 14:37:02", "content": "or to be ahead of uncle
", "votes": null, "replies": [] }, { "id": 592648, "author_name": "cpmpml", "author_url": "", "post_date": "08/05/2019 16:00:39", "content": "ok, you convinced me to start a MPNN.
", "votes": null, "replies": [] }, { "id": 594110, "author_name": "christofhenkel", "author_url": "", "post_date": "08/07/2019 15:22:43", "content": "
\n\n\nor to be ahead of uncle
\n
Seems this motivates you indeed ;)
", "votes": null, "replies": [] }, { "id": 602898, "author_name": "christofhenkel", "author_url": "", "post_date": "08/19/2019 16:10:14", "content": "Seems like you found someone who can handle GNN ;)
", "votes": null, "replies": [] }, { "id": 602900, "author_name": "cpmpml", "author_url": "", "post_date": "08/19/2019 16:12:19", "content": "We still miss a quantum chemist and a fourth scientist.
", "votes": null, "replies": [] }, { "id": 615715, "author_name": "cpmpml", "author_url": "", "post_date": "09/02/2019 09:35:18", "content": "> Seems like you found someone who can handle GNN ;)
\n\nOutrunner NN is not a graph NN by any mean actually. I couldn't say it during the competition obviously ;)
\n\nOur models are closer in spirit to 2nd team, except twe use distance and angles instead of xyz coordinates. I'm still puzzled that the order of atom index and xyz coordinates are significant given how model that use it outperform models that are rotation and atom order independent.
", "votes": null, "replies": [] }, { "id": 615719, "author_name": "antorsae", "author_url": "", "post_date": "09/02/2019 09:38:56", "content": "We do not use atom index (position in the array) 9999?9
", "votes": null, "replies": [] }, { "id": 615755, "author_name": "cpmpml", "author_url": "", "post_date": "09/02/2019 10:36:34", "content": "3rd team uses the atoms as a sequence.
", "votes": null, "replies": [] }, { "id": 615764, "author_name": "antorsae", "author_url": "", "post_date": "09/02/2019 11:00:32", "content": "I bet if you remove the positional info their model performs equal or better
", "votes": null, "replies": [] }, { "id": 615772, "author_name": "cpmpml", "author_url": "", "post_date": "09/02/2019 11:06:55", "content": "Maybe. My point is that models that are rotation invariant and translation invariant and atom order invariant perform worse than those that miss one or more of these properties.
\n\nThis is surprising to me (and few others).
\n\nIt does not mean that the models that perform better are bad, to the contrary. They use data better than the other models.
\n\nHaving preconceived assumptions about what the models should be led to worse performance. Mind openness led to better performance.
", "votes": null, "replies": [] }, { "id": 615780, "author_name": "antorsae", "author_url": "", "post_date": "09/02/2019 11:24:21", "content": "\nFrom https://www.quantamagazine.org/how-artificial-intelligence-is-changing-science-20190311/
This was an inspiration to approach this challenge
", "votes": null, "replies": [] } ] }, { "id": 592713, "author_name": "youhanlee", "author_url": "", "post_date": "08/05/2019 18:02:57", "content": "I really want to do quantum jump to top...:)
", "votes": null, "replies": [ { "id": 592763, "author_name": "returnofsputnik", "author_url": "", "post_date": "08/05/2019 19:37:59", "content": "Im rooting for you YouHan. you are always so close to gold. Good luck
", "votes": null, "replies": [] }, { "id": 593171, "author_name": "youhanlee", "author_url": "", "post_date": "08/06/2019 08:48:10", "content": "Thanks for kind words, Corey! \nI'm also rooting for you. I hope both of us get golds, then I became competition master, and you became so close to grand master! Good luck :)
", "votes": null, "replies": [] } ] }, { "id": 594459, "author_name": "tunguz", "author_url": "", "post_date": "08/08/2019 04:05:45", "content": "My theory: they use quantum tunneling to go through the -3.0 barrier.
", "votes": null, "replies": [] }, { "id": 594476, "author_name": "robikscube", "author_url": "", "post_date": "08/08/2019 04:53:32", "content": "
Nice one!
", "votes": null, "replies": [] } ] }, { "id": 594511, "author_name": "scaomath", "author_url": "", "post_date": "08/08/2019 05:48:17", "content": "BTW I found it also interesting that there is a team named Walter White while another team is named Pinkman, and what is more interesting is that a team-member of Pinkman has Bojack horseman's character as his avatar (which was voiced by Pinkman's actor).
\n\nI think someone should form a team named DEA.
", "votes": null, "replies": [ { "id": 594519, "author_name": "aerdem4", "author_url": "", "post_date": "08/08/2019 06:05:28", "content": "A bit off-topic but Todd Chavez is my favorite TV character. I started watching Breaking Bad much later and didn't even notice it is the same voice until a random google search.
\n\nIn this competition, there are a lot of cartels on top of us and we couldn't get our desired purity yet. It is not blue enough:)
", "votes": null, "replies": [] }, { "id": 595056, "author_name": "scaomath", "author_url": "", "post_date": "08/08/2019 19:42:08", "content": "I agree. \"4 GM and the brain\" sounds pretty villain'ish.
", "votes": null, "replies": [] } ] }, { "id": 605884, "author_name": "eladwar", "author_url": "", "post_date": "08/22/2019 22:33:20", "content": "Maybe the top team is primarily training on the first 29 % of the data
", "votes": null, "replies": [ { "id": 605897, "author_name": "toshik", "author_url": "", "post_date": "08/22/2019 23:21:41", "content": "Why do you think so ? Do you revealed the way to devide public/private dataset ?
", "votes": null, "replies": [] } ] }, { "id": 609271, "author_name": "cpmpml", "author_url": "", "post_date": "08/27/2019 14:53:25", "content": "The gap is being closed, and there is a continuum of scores now. Largest gap is only a out 0.085. It will be very interesting to read writeups.
", "votes": null, "replies": [] }, { "id": 610444, "author_name": "robikscube", "author_url": "", "post_date": "08/28/2019 20:02:14", "content": "The gap has been closed!
\n\n
Well it's over. I can put my computer to sleep ;-)
\n\n
This is very fitting 😄
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "591959": "It's clear one team has pulled ahead from the rest on the public leaderboard. Certainly it's possible that other teams have similarly strong models but are just holding back on submitting. Apart from that, right now it appears they are blowing the rest of the field away.\n\nAre there theories as to how one team could be so much stronger than the rest?\n\n\n\n", "591978": "> Are there theories as to how one team could be so much stronger than the rest?\n\nYou provided the answer:\n\n> Certainly it's possible that other teams have similarly strong models but are just holding back on submitting.", "591990": "We have developed a quantum computer and achieved quantum supremacy.", "591999": "Great, you'll be removed then because you use quantum computation :D", "592068": "Maybe they ensemble models?", "592084": "Out of curiosity, for those of you who has average log MAE less than -2.5 (by which I assume is achieved by ensembling different types of models), which one is better 1JHC and 1JHN? GB-based (LGB XGB) or NN-based (NN, GNN, GCN)?", "592458": "It seems some poeple were motivated to close some of the gap after 1st team...", "592479": "We basically work with one model for now, hence cannot really answer your question ;) \n\nI would be surprised if lgb is better than NNs here.", "592601": "or to be ahead of uncle", "592648": "ok, you convinced me to start a MPNN.", "592713": "I really want to do quantum jump to top...:)", "592763": "Im rooting for you YouHan. you are always so close to gold. Good luck", "592826": "Pics or it didn't happen", "593138": "As requested, here's a pic of our quantum computer. We're worried b/c it's getting AGI-like qualities and we don't know how to contain it:\n\n", "593171": "Thanks for kind words, Corey! \nI'm also rooting for you. I hope both of us get golds, then I became competition master, and you became so close to grand master! Good luck :)", "593258": "> We're worried b/c it's getting AGI-like qualities and we don't know how to contain it\n \nAGI getting out of control? hmm\nIt is going to decide humankind's fate? hmm\n \nThats a serious situation rather than try to fool it with adversarial samples and give humanity another chance, I would just pull the plug out.\n \n", "594110": "", "594459": "My theory: they use quantum tunneling to go through the -3.0 barrier.", "594476": "", "594483": "Nice one!", "594511": "BTW I found it also interesting that there is a team named Walter White while another team is named Pinkman, and what is more interesting is that a team-member of Pinkman has Bojack horseman's character as his avatar (which was voiced by Pinkman's actor). \n\nI think someone should form a team named DEA.", "594519": "A bit off-topic but Todd Chavez is my favorite TV character. I started watching Breaking Bad much later and didn't even notice it is the same voice until a random google search.\n\nIn this competition, there are a lot of cartels on top of us and we couldn't get our desired purity yet. It is not blue enough:)", "595056": "I agree. \"4 GM and the brain\" sounds pretty villain'ish.", "602896": "> or to be ahead of uncle\n\nSeems this motivates you indeed ;)", "602898": "Seems like you found someone who can handle GNN ;)", "602900": "We still miss a quantum chemist and a fourth scientist.", "605884": "Maybe the top team is primarily training on the first 29 % of the data", "605897": "Why do you think so ? Do you revealed the way to devide public/private dataset ?", "609271": "The gap is being closed, and there is a continuum of scores now. Largest gap is only a out 0.085. It will be very interesting to read writeups.", "610444": "The gap has been closed!\n\n", "610677": "Well it's over. I can put my computer to sleep ;-)\n\n", "610683": "This is very fitting 😄", "615715": "> Seems like you found someone who can handle GNN ;)\n\nOutrunner NN is not a graph NN by any mean actually. I couldn't say it during the competition obviously ;) \n\nOur models are closer in spirit to 2nd team, except twe use distance and angles instead of xyz coordinates. I'm still puzzled that the order of atom index and xyz coordinates are significant given how model that use it outperform models that are rotation and atom order independent.", "615719": "We do not use atom index (position in the array) 9999?9", "615755": "3rd team uses the atoms as a sequence.", "615764": "I bet if you remove the positional info their model performs equal or better", "615772": "Maybe. My point is that models that are rotation invariant and translation invariant and atom order invariant perform worse than those that miss one or more of these properties. \n\nThis is surprising to me (and few others). \n\nIt does not mean that the models that perform better are bad, to the contrary. They use data better than the other models. \n\nHaving preconceived assumptions about what the models should be led to worse performance. Mind openness led to better performance.", "615780": "\nFrom https://www.quantamagazine.org/how-artificial-intelligence-is-changing-science-20190311/\n\nThis was an inspiration to approach this challenge" }, "source": "meta" }