I want to take the opportunity to thank everyone who participated in this competition! This was the first competition that I got to host, and it was a fun ride. \nThe amount of enthusiasm, dedication and expertise that you guys put into this is truly amazing! For us this competition was a huge success, and I hope you enjoyed it as much as I did.
\n\nAt this point, I want to give special thanks to Alexandre Barachant who clearly is an expert in EEG data and has been very active since the very beginning of the competition. He was very open with his knowledge, helped many participants and his scripts and posts provide valuable insight into the data. Thank you Alexandre!
\n\nAlso (preliminary) congratulations to the winners! You beat our in-house models (AUC ~0.95) by quite some margin, and I'm really looking forward to learn how you pulled that off.
", "messages": [ { "id": "91167", "postDate": "09/01/2015 09:32:24", "content": "I want to take the opportunity to thank everyone who participated in this competition! This was the first competition that I got to host, and it was a fun ride. \nThe amount of enthusiasm, dedication and expertise that you guys put into this is truly amazing! For us this competition was a huge success, and I hope you enjoyed it as much as I did.
\n\nAt this point, I want to give special thanks to Alexandre Barachant who clearly is an expert in EEG data and has been very active since the very beginning of the competition. He was very open with his knowledge, helped many participants and his scripts and posts provide valuable insight into the data. Thank you Alexandre!
\n\nAlso (preliminary) congratulations to the winners! You beat our in-house models (AUC ~0.95) by quite some margin, and I'm really looking forward to learn how you pulled that off.
", "rawMarkdown": "I want to take the opportunity to thank everyone who participated in this competition! This was the first competition that I got to host, and it was a fun ride. \r\nThe amount of enthusiasm, dedication and expertise that you guys put into this is truly amazing! For us this competition was a huge success, and I hope you enjoyed it as much as I did. \r\n\r\nAt this point, I want to give special thanks to Alexandre Barachant who clearly is an expert in EEG data and has been very active since the very beginning of the competition. He was very open with his knowledge, helped many participants and his scripts and posts provide valuable insight into the data. Thank you Alexandre!\r\n\r\nAlso (preliminary) congratulations to the winners! You beat our in-house models (AUC ~0.95) by quite some margin, and I'm really looking forward to learn how you pulled that off.", "votes": null }, { "id": "91671", "postDate": "09/05/2015 22:04:40", "content": "@Klaus. I would like to thank you for hosting such a great competition ! This dataset is really good, and it was a pleasure to work on. I also enjoyed the way the problem was defined, with the sequence of (overlapping) events and the prediction for every time sample. it makes this competition refreshing and definitely more challenging.
\n\nI would love to see another challenge on the full version of this dataset. What about something like using EEG to reconstruct EMG power or hand position/velocity/grip force ?
\n\n@everyone. I would like to thank all the participants for making this challenge so fun. A special thanks to Yair, who gave us a lot of challenge during all of the competition, and the Team HEDJ who make a spectacular progression in the last days of the challenge (Honestly, this was not good for our blood pressure ... )
", "rawMarkdown": "Klaus. I would like to thank you for hosting such a great competition ! This dataset is really good, and it was a pleasure to work on. I also enjoyed the way the problem was defined, with the sequence of (overlapping) events and the prediction for every time sample. it makes this competition refreshing and definitely more challenging.\r\n\r\nI would love to see another challenge on the full version of this dataset. What about something like using EEG to reconstruct EMG power or hand position/velocity/grip force ?\r\n\r\n@everyone. I would like to thank all the participants for making this challenge so fun. A special thanks to Yair, who gave us a lot of challenge during all of the competition, and the Team HEDJ who make a spectacular progression in the last days of the challenge (Honestly, this was not good for our blood pressure ... )", "votes": null }, { "id": "93050", "postDate": "09/20/2015 23:33:12", "content": "@Klaus, Now that I've nearly caught up on sleep after our final push, I'd like to second (or perhaps third?) both of Alexandre's comments. I really enjoyed this competition and I have a dozen or so ideas for how I could do better on this or similar data, so I'd also enjoy seeing a followup to this competition. (Of course, most of those ideas won't work, but a few of them are likely to pan out and it would be interesting to see in any event).
\n\n@Alexandre, we're proud to have got you sweating!
", "rawMarkdown": "Klaus, Now that I've nearly caught up on sleep after our final push, I'd like to second (or perhaps third?) both of Alexandre's comments. I really enjoyed this competition and I have a dozen or so ideas for how I could do better on this or similar data, so I'd also enjoy seeing a followup to this competition. (Of course, most of those ideas won't work, but a few of them are likely to pan out and it would be interesting to see in any event).\r\n\r\n@Alexandre, we're proud to have got you sweating!", "votes": null }, { "id": "93213", "postDate": "09/22/2015 21:15:06", "content": "@Klaus It was such an interesting competition. I enjoyed every bit of it. I was so surprised to see convolutional networks, CNN in our case and RCNN in team @daheimao case, beat up traditional hand crafted feature + traditional model pipelines so well in electrophysiological signals! The most impressive part was the beat up didn't involve any per-subject or per-event training. All models were trained using same params for all subjects and events.
\n\nSo, that means, this should have also surprised you as you mentioned in the forums that you would be surprised if one can generalize across subjects in an answer to @rcarson question about per-subject training.
\n\nIn any case, it was so much fun and I really hope you would follow it up with another competition soon with the complete dataset you have published in that Nature paper or any new dataset your group may have!
\n\n@Alex, if a couple of the nets that we were training during the last day but didn't make it by the deadline made it to the ensemble, you'd have sweated more as we at least beat you on the public LB with a score of 98.130 right after the deadline :) I admit it overfitted a bit on the private LB. Aws, thanks so much for sharing valuable domain knowledge thoughout the competition and for making the competition so much fun!
", "rawMarkdown": "Klaus It was such an interesting competition. I enjoyed every bit of it. I was so surprised to see convolutional networks, CNN in our case and RCNN in team @daheimao case, beat up traditional hand crafted feature + traditional model pipelines so well in electrophysiological signals! The most impressive part was the beat up didn't involve any per-subject or per-event training. All models were trained using same params for all subjects and events. \r\n\r\nSo, that means, this should have also surprised you as you mentioned in the forums that you would be surprised if one can generalize across subjects in an answer to @rcarson question about per-subject training.\r\n\r\nIn any case, it was so much fun and I really hope you would follow it up with another competition soon with the complete dataset you have published in that Nature paper or any new dataset your group may have!\r\n\r\n@Alex, if a couple of the nets that we were training during the last day but didn't make it by the deadline made it to the ensemble, you'd have sweated more as we at least beat you on the public LB with a score of 98.130 right after the deadline :) I admit it overfitted a bit on the private LB. Aws, thanks so much for sharing valuable domain knowledge thoughout the competition and for making the competition so much fun!", "votes": null }, { "id": "93693", "postDate": "09/30/2015 12:12:05", "content": "Thank you guys for the kind words, I will forward the praise of the dataset to the people who actually recorded it! \nI agree that it is an interesting dataset, and this competition has only scratched the surface of what could be done with it. For one we completely ignored the weight and surface friction changes of the grasped objects. And detecting that moment of surprise could potentially help to allow a more adaptive control of prosthetics. Also I really @Alexandre's idea of doing a regression on the positions. Given the great results that you achieved during this competition, this very difficult setup almost seems within reach!
\n\nUnfortunately the EU research project within which we recorded and analyzed the data is now over. So I don't think we'll have funding to organize another competition like this in the near future. But the data is out there for anybody! If you have a good idea, don't hesitate to just give it a shot and publish a nice paper about it. In fact, we were hoping to water your mouths a bit with this competition. :-)
\n\n@Deep: I am surprised to see how well ConvNets are doing on this. I was expecting RNNs to dominate the field eventually. (Seems like that was just my personal bias.) \nBut I'm also a bit confused: You said that there was no per-subject training, but from your code it seems to me you are in fact training the networks for different subjects individually. Am I wrong? \nIt is true though, that the architecture and the hyperparameters are all the same for all subjects, and that itself is remarkable.
", "rawMarkdown": "Thank you guys for the kind words, I will forward the praise of the dataset to the people who actually recorded it! \r\nI agree that it is an interesting dataset, and this competition has only scratched the surface of what could be done with it. For one we completely ignored the weight and surface friction changes of the grasped objects. And detecting that moment of surprise could potentially help to allow a more adaptive control of prosthetics. Also I really @Alexandre's idea of doing a regression on the positions. Given the great results that you achieved during this competition, this very difficult setup almost seems within reach!\r\n\r\nUnfortunately the EU research project within which we recorded and analyzed the data is now over. So I don't think we'll have funding to organize another competition like this in the near future. But the data is out there for anybody! If you have a good idea, don't hesitate to just give it a shot and publish a nice paper about it. In fact, we were hoping to water your mouths a bit with this competition. :-)\r\n\r\n@Deep: I am surprised to see how well ConvNets are doing on this. I was expecting RNNs to dominate the field eventually. (Seems like that was just my personal bias.) \r\nBut I'm also a bit confused: You said that there was no per-subject training, but from your code it seems to me you are in fact training the networks for different subjects individually. Am I wrong? \r\nIt is true though, that the architecture and the hyperparameters are all the same for all subjects, and that itself is remarkable.", "votes": null }, { "id": "93717", "postDate": "09/30/2015 17:19:01", "content": "@Klaus, Indeed our mouths are more than watered by convnets performance :) and 4 people from all top 3 teams + another 1 already formed a team to explore either the full dataset and/or another standard dataset in hopes of churning out a good paper.
\n\nWhen I said 'we didn't do per subject training' I meant we didn't tune different parameters for each subject separately. I was referring to your answer to @rcarson question of tuning params for each subject separately.
\n\nOf course to have one single model for all subjects would have been more impressive and ideal, but, having 12 models with same params is the closest thing we could do.
", "rawMarkdown": "Klaus, Indeed our mouths are more than watered by convnets performance :) and 4 people from all top 3 teams + another 1 already formed a team to explore either the full dataset and/or another standard dataset in hopes of churning out a good paper.\r\n\r\nWhen I said 'we didn't do per subject training' I meant we didn't tune different parameters for each subject separately. I was referring to [your answer][1] to @rcarson question of tuning params for each subject separately.\r\n\r\nOf course to have one single model for all subjects would have been more impressive and ideal, but, having 12 models with same params is the closest thing we could do.\r\n\r\n\r\n [1]: https://www.kaggle.com/c/grasp-and-lift-eeg-detection/forums/t/15961/is-per-subject-model-allowed/89493#post89493", "votes": null } ], "comments": [ { "id": 91671, "author_name": "alexandrebarachant", "author_url": "", "post_date": "09/05/2015 22:04:40", "content": "@Klaus. I would like to thank you for hosting such a great competition ! This dataset is really good, and it was a pleasure to work on. I also enjoyed the way the problem was defined, with the sequence of (overlapping) events and the prediction for every time sample. it makes this competition refreshing and definitely more challenging.
\n\nI would love to see another challenge on the full version of this dataset. What about something like using EEG to reconstruct EMG power or hand position/velocity/grip force ?
\n\n@everyone. I would like to thank all the participants for making this challenge so fun. A special thanks to Yair, who gave us a lot of challenge during all of the competition, and the Team HEDJ who make a spectacular progression in the last days of the challenge (Honestly, this was not good for our blood pressure ... )
", "votes": null, "replies": [] }, { "id": 93050, "author_name": "bitsofbits", "author_url": "", "post_date": "09/20/2015 23:33:12", "content": "@Klaus, Now that I've nearly caught up on sleep after our final push, I'd like to second (or perhaps third?) both of Alexandre's comments. I really enjoyed this competition and I have a dozen or so ideas for how I could do better on this or similar data, so I'd also enjoy seeing a followup to this competition. (Of course, most of those ideas won't work, but a few of them are likely to pan out and it would be interesting to see in any event).
\n\n@Alexandre, we're proud to have got you sweating!
", "votes": null, "replies": [] }, { "id": 93213, "author_name": "deepcnn", "author_url": "", "post_date": "09/22/2015 21:15:06", "content": "@Klaus It was such an interesting competition. I enjoyed every bit of it. I was so surprised to see convolutional networks, CNN in our case and RCNN in team @daheimao case, beat up traditional hand crafted feature + traditional model pipelines so well in electrophysiological signals! The most impressive part was the beat up didn't involve any per-subject or per-event training. All models were trained using same params for all subjects and events.
\n\nSo, that means, this should have also surprised you as you mentioned in the forums that you would be surprised if one can generalize across subjects in an answer to @rcarson question about per-subject training.
\n\nIn any case, it was so much fun and I really hope you would follow it up with another competition soon with the complete dataset you have published in that Nature paper or any new dataset your group may have!
\n\n@Alex, if a couple of the nets that we were training during the last day but didn't make it by the deadline made it to the ensemble, you'd have sweated more as we at least beat you on the public LB with a score of 98.130 right after the deadline :) I admit it overfitted a bit on the private LB. Aws, thanks so much for sharing valuable domain knowledge thoughout the competition and for making the competition so much fun!
", "votes": null, "replies": [] }, { "id": 93693, "author_name": "qwlouse", "author_url": "", "post_date": "09/30/2015 12:12:05", "content": "Thank you guys for the kind words, I will forward the praise of the dataset to the people who actually recorded it! \nI agree that it is an interesting dataset, and this competition has only scratched the surface of what could be done with it. For one we completely ignored the weight and surface friction changes of the grasped objects. And detecting that moment of surprise could potentially help to allow a more adaptive control of prosthetics. Also I really @Alexandre's idea of doing a regression on the positions. Given the great results that you achieved during this competition, this very difficult setup almost seems within reach!
\n\nUnfortunately the EU research project within which we recorded and analyzed the data is now over. So I don't think we'll have funding to organize another competition like this in the near future. But the data is out there for anybody! If you have a good idea, don't hesitate to just give it a shot and publish a nice paper about it. In fact, we were hoping to water your mouths a bit with this competition. :-)
\n\n@Deep: I am surprised to see how well ConvNets are doing on this. I was expecting RNNs to dominate the field eventually. (Seems like that was just my personal bias.) \nBut I'm also a bit confused: You said that there was no per-subject training, but from your code it seems to me you are in fact training the networks for different subjects individually. Am I wrong? \nIt is true though, that the architecture and the hyperparameters are all the same for all subjects, and that itself is remarkable.
", "votes": null, "replies": [] }, { "id": 93717, "author_name": "deepcnn", "author_url": "", "post_date": "09/30/2015 17:19:01", "content": "@Klaus, Indeed our mouths are more than watered by convnets performance :) and 4 people from all top 3 teams + another 1 already formed a team to explore either the full dataset and/or another standard dataset in hopes of churning out a good paper.
\n\nWhen I said 'we didn't do per subject training' I meant we didn't tune different parameters for each subject separately. I was referring to your answer to @rcarson question of tuning params for each subject separately.
\n\nOf course to have one single model for all subjects would have been more impressive and ideal, but, having 12 models with same params is the closest thing we could do.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "91167": "I want to take the opportunity to thank everyone who participated in this competition! This was the first competition that I got to host, and it was a fun ride. \r\nThe amount of enthusiasm, dedication and expertise that you guys put into this is truly amazing! For us this competition was a huge success, and I hope you enjoyed it as much as I did. \r\n\r\nAt this point, I want to give special thanks to Alexandre Barachant who clearly is an expert in EEG data and has been very active since the very beginning of the competition. He was very open with his knowledge, helped many participants and his scripts and posts provide valuable insight into the data. Thank you Alexandre!\r\n\r\nAlso (preliminary) congratulations to the winners! You beat our in-house models (AUC ~0.95) by quite some margin, and I'm really looking forward to learn how you pulled that off.", "91671": "Klaus. I would like to thank you for hosting such a great competition ! This dataset is really good, and it was a pleasure to work on. I also enjoyed the way the problem was defined, with the sequence of (overlapping) events and the prediction for every time sample. it makes this competition refreshing and definitely more challenging.\r\n\r\nI would love to see another challenge on the full version of this dataset. What about something like using EEG to reconstruct EMG power or hand position/velocity/grip force ?\r\n\r\n@everyone. I would like to thank all the participants for making this challenge so fun. A special thanks to Yair, who gave us a lot of challenge during all of the competition, and the Team HEDJ who make a spectacular progression in the last days of the challenge (Honestly, this was not good for our blood pressure ... )", "93050": "Klaus, Now that I've nearly caught up on sleep after our final push, I'd like to second (or perhaps third?) both of Alexandre's comments. I really enjoyed this competition and I have a dozen or so ideas for how I could do better on this or similar data, so I'd also enjoy seeing a followup to this competition. (Of course, most of those ideas won't work, but a few of them are likely to pan out and it would be interesting to see in any event).\r\n\r\n@Alexandre, we're proud to have got you sweating!", "93213": "Klaus It was such an interesting competition. I enjoyed every bit of it. I was so surprised to see convolutional networks, CNN in our case and RCNN in team @daheimao case, beat up traditional hand crafted feature + traditional model pipelines so well in electrophysiological signals! The most impressive part was the beat up didn't involve any per-subject or per-event training. All models were trained using same params for all subjects and events. \r\n\r\nSo, that means, this should have also surprised you as you mentioned in the forums that you would be surprised if one can generalize across subjects in an answer to @rcarson question about per-subject training.\r\n\r\nIn any case, it was so much fun and I really hope you would follow it up with another competition soon with the complete dataset you have published in that Nature paper or any new dataset your group may have!\r\n\r\n@Alex, if a couple of the nets that we were training during the last day but didn't make it by the deadline made it to the ensemble, you'd have sweated more as we at least beat you on the public LB with a score of 98.130 right after the deadline :) I admit it overfitted a bit on the private LB. Aws, thanks so much for sharing valuable domain knowledge thoughout the competition and for making the competition so much fun!", "93693": "Thank you guys for the kind words, I will forward the praise of the dataset to the people who actually recorded it! \r\nI agree that it is an interesting dataset, and this competition has only scratched the surface of what could be done with it. For one we completely ignored the weight and surface friction changes of the grasped objects. And detecting that moment of surprise could potentially help to allow a more adaptive control of prosthetics. Also I really @Alexandre's idea of doing a regression on the positions. Given the great results that you achieved during this competition, this very difficult setup almost seems within reach!\r\n\r\nUnfortunately the EU research project within which we recorded and analyzed the data is now over. So I don't think we'll have funding to organize another competition like this in the near future. But the data is out there for anybody! If you have a good idea, don't hesitate to just give it a shot and publish a nice paper about it. In fact, we were hoping to water your mouths a bit with this competition. :-)\r\n\r\n@Deep: I am surprised to see how well ConvNets are doing on this. I was expecting RNNs to dominate the field eventually. (Seems like that was just my personal bias.) \r\nBut I'm also a bit confused: You said that there was no per-subject training, but from your code it seems to me you are in fact training the networks for different subjects individually. Am I wrong? \r\nIt is true though, that the architecture and the hyperparameters are all the same for all subjects, and that itself is remarkable.", "93717": "Klaus, Indeed our mouths are more than watered by convnets performance :) and 4 people from all top 3 teams + another 1 already formed a team to explore either the full dataset and/or another standard dataset in hopes of churning out a good paper.\r\n\r\nWhen I said 'we didn't do per subject training' I meant we didn't tune different parameters for each subject separately. I was referring to [your answer][1] to @rcarson question of tuning params for each subject separately.\r\n\r\nOf course to have one single model for all subjects would have been more impressive and ideal, but, having 12 models with same params is the closest thing we could do.\r\n\r\n\r\n [1]: https://www.kaggle.com/c/grasp-and-lift-eeg-detection/forums/t/15961/is-per-subject-model-allowed/89493#post89493" }, "source": "meta" }