After looking at the quick ascension of @rcarson to the top of the leaderboard, I couldn't help it but wonder if that result is just a single NN model or ensemble with other methods.
\n\nMy local validation for current run seems so unreal. Even for the difficult subject 2 data, I got ~92%. Is this for reals? I mean, I haven't even gotton started in modifying the network much, properly preprocessing. Could you share your local Val/LB experience with the NN written by @Tim?
", "messages": [ { "id": "88495", "postDate": "08/05/2015 06:07:53", "content": "After looking at the quick ascension of @rcarson to the top of the leaderboard, I couldn't help it but wonder if that result is just a single NN model or ensemble with other methods.
\n\nMy local validation for current run seems so unreal. Even for the difficult subject 2 data, I got ~92%. Is this for reals? I mean, I haven't even gotton started in modifying the network much, properly preprocessing. Could you share your local Val/LB experience with the NN written by @Tim?
", "rawMarkdown": "After looking at the quick ascension of @rcarson to the top of the leaderboard, I couldn't help it but wonder if that result is just a single NN model or ensemble with other methods.\r\n\r\nMy local validation for current run seems so unreal. Even for the difficult subject 2 data, I got ~92%. Is this for reals? I mean, I haven't even gotton started in modifying the network much, properly preprocessing. Could you share your local Val/LB experience with the NN written by @Tim?", "votes": null }, { "id": "88505", "postDate": "08/05/2015 07:53:08", "content": "Hi there, that is an ensemble of several models. I tried the nn model only once and it got 0.82/0.87, local/lb : )
", "rawMarkdown": "Hi there, that is an ensemble of several models. I tried the nn model only once and it got 0.82/0.87, local/lb : )", "votes": null }, { "id": "88551", "postDate": "08/05/2015 17:13:35", "content": "I have not start trying xgboost or nn yet, but I doubt it will make a big boost to my score. This competition still have a long way to go, we might even need to get to 0.98 or more if want to win.
\n\nAnd also beware this is just series9, the final result is only based on series10, I am not sure how different these two series are, so it is possible some lucky guy unintentionally overfit to series 10 and get a big boost on their score.
", "rawMarkdown": "I have not start trying xgboost or nn yet, but I doubt it will make a big boost to my score. This competition still have a long way to go, we might even need to get to 0.98 or more if want to win.\r\n\r\nAnd also beware this is just series9, the final result is only based on series10, I am not sure how different these two series are, so it is possible some lucky guy unintentionally overfit to series 10 and get a big boost on their score.", "votes": null }, { "id": "88591", "postDate": "08/05/2015 21:43:13", "content": "@rcarson, thanks for the answer!
\n\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference?
", "rawMarkdown": "rcarson, thanks for the answer!\r\n\r\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference?", "votes": null }, { "id": "88601", "postDate": "08/05/2015 23:08:37", "content": "Yes, I didn't ensemble NN, I was still reading relevant papers, try to get a basic idea of what I am doing.
\n\nI am actually more interested in whether or not we can classify what we are grasping or what's the weight of the object we are grasping, it would be cool if we can do this by just looking at the eeg signal.
\n\n[quote=Deep;88591]
\n\n@rcarson, thanks for the answer!
\n\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference?
\n\n[/quote]
", "rawMarkdown": "Yes, I didn't ensemble NN, I was still reading relevant papers, try to get a basic idea of what I am doing.\r\n\r\nI am actually more interested in whether or not we can classify what we are grasping or what's the weight of the object we are grasping, it would be cool if we can do this by just looking at the eeg signal.\r\n\r\n\r\n[quote=Deep;88591]\r\n\r\n@rcarson, thanks for the answer!\r\n\r\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference? \r\n\r\n[/quote]", "votes": null } ], "comments": [ { "id": 88505, "author_name": "jiweiliu", "author_url": "", "post_date": "08/05/2015 07:53:08", "content": "Hi there, that is an ensemble of several models. I tried the nn model only once and it got 0.82/0.87, local/lb : )
", "votes": null, "replies": [] }, { "id": 88551, "author_name": "skylibrary", "author_url": "", "post_date": "08/05/2015 17:13:35", "content": "I have not start trying xgboost or nn yet, but I doubt it will make a big boost to my score. This competition still have a long way to go, we might even need to get to 0.98 or more if want to win.
\n\nAnd also beware this is just series9, the final result is only based on series10, I am not sure how different these two series are, so it is possible some lucky guy unintentionally overfit to series 10 and get a big boost on their score.
", "votes": null, "replies": [] }, { "id": 88591, "author_name": "deepcnn", "author_url": "", "post_date": "08/05/2015 21:43:13", "content": "@rcarson, thanks for the answer!
\n\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference?
", "votes": null, "replies": [] }, { "id": 88601, "author_name": "skylibrary", "author_url": "", "post_date": "08/05/2015 23:08:37", "content": "Yes, I didn't ensemble NN, I was still reading relevant papers, try to get a basic idea of what I am doing.
\n\nI am actually more interested in whether or not we can classify what we are grasping or what's the weight of the object we are grasping, it would be cool if we can do this by just looking at the eeg signal.
\n\n[quote=Deep;88591]
\n\n@rcarson, thanks for the answer!
\n\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference?
\n\n[/quote]
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "88495": "After looking at the quick ascension of @rcarson to the top of the leaderboard, I couldn't help it but wonder if that result is just a single NN model or ensemble with other methods.\r\n\r\nMy local validation for current run seems so unreal. Even for the difficult subject 2 data, I got ~92%. Is this for reals? I mean, I haven't even gotton started in modifying the network much, properly preprocessing. Could you share your local Val/LB experience with the NN written by @Tim?", "88505": "Hi there, that is an ensemble of several models. I tried the nn model only once and it got 0.82/0.87, local/lb : )", "88551": "I have not start trying xgboost or nn yet, but I doubt it will make a big boost to my score. This competition still have a long way to go, we might even need to get to 0.98 or more if want to win.\r\n\r\nAnd also beware this is just series9, the final result is only based on series10, I am not sure how different these two series are, so it is possible some lucky guy unintentionally overfit to series 10 and get a big boost on their score.", "88591": "rcarson, thanks for the answer!\r\n\r\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference?", "88601": "Yes, I didn't ensemble NN, I was still reading relevant papers, try to get a basic idea of what I am doing.\r\n\r\nI am actually more interested in whether or not we can classify what we are grasping or what's the weight of the object we are grasping, it would be cool if we can do this by just looking at the eeg signal.\r\n\r\n\r\n[quote=Deep;88591]\r\n\r\n@rcarson, thanks for the answer!\r\n\r\n@skylibrary, so you haven't NN to your ensemble yet on your current rank? If so, is preprocessing and feature extraction that is making the difference? \r\n\r\n[/quote]" }, "source": "meta" }