Just for fun, and to see how far it can go, I was thinking about borrowing and stacking on top of a few boundary representation techniques from the image recognition community (see 1 and 2) to approach this challenge. In 1 their shape context approach gets 96.2%. I'm happy to use neural nets as well but alternative representations/approaches could serve as input and inspire a more creative solution than a typical straight forward neural network ensemble.
\n\nFeel free to contact me if there's any interest in collaborating over github, etc. I can contribute about 8 hours a week on this. I can work in Python or R but prefer Python since I've started developing some prototype code in Python.
\n\nBackground: I'm a machine learning and data science consultant, have a masters in computer science, bachelors in electrical engineering and have 6 years of industry experience prior to starting consulting.
\n\n1 "Shape Context: A new descriptor for shape matching and object recognition," Serge Belongie, et al (2000)\n2 "Lecture 7: Shape Description (Contours) ," Bryan S. Morse, Brigham Young University
", "messages": [ { "id": 107447, "postDate": "2016-02-09T22:42:37.133Z", "content": "Anyone else interested in joining this team? There's a git repo here and I've set up a slack channel. Currently we're doing EDA and I've written up a rough team document on roles and tasks so people can feel at least somewhat organized.
", "rawMarkdown": "Anyone else interested in joining this team? There's a [git repo here][1] and I've set up a slack channel. Currently we're doing EDA and I've written up a rough team document on roles and tasks so people can feel at least somewhat organized.\r\n\r\n\r\n [1]: https://github.com/robinsonkwame/DigitRecognizer" }, { "id": 105807, "postDate": "2016-01-26T20:49:57.887Z", "content": "That's okay, Ben helped me out and I was able to send you an invite. Be sure to edit your post and take out your email address so it's not public. I'll follow up in an email.
", "rawMarkdown": "That's okay, Ben helped me out and I was able to send you an invite. Be sure to edit your post and take out your email address so it's not public. I'll follow up in an email." }, { "id": 105803, "postDate": "2016-01-26T20:34:20.083Z", "content": "Hi, I saw the issue you posted. Do you need me to offer you my email address?
", "rawMarkdown": "Hi, I saw the issue you posted. Do you need me to offer you my email address? " }, { "id": 105802, "postDate": "2016-01-26T20:15:59.613Z", "content": "Yep, still looking. There appears to a bug with team creation (Invite Someone to a Team) so I've contacted @benhammer to add you to the team.
", "rawMarkdown": "Yep, still looking. There appears to a bug with team creation ([Invite Someone to a Team][1]) so I've contacted @benhammer to add you to the team.\r\n\r\n\r\n [1]: https://www.kaggle.com/forums/f/15/kaggle-forum/t/14111/invite-someone-to-a-team/77627" }, { "id": 105768, "postDate": "2016-01-26T16:35:46.080Z", "content": "Hi，are you still looking for collaborators in this problem?I am interested in joining you. I am a Master student focussing on statistics & ML. I can work in Python & R.
", "rawMarkdown": "Hi，are you still looking for collaborators in this problem?I am interested in joining you. I am a Master student focussing on statistics & ML. I can work in Python & R." }, { "id": 105107, "postDate": "2016-01-20T03:03:50.757Z", "content": "Not sure what this is what you are talking about, but I recently tried applying a dynamic time warping solution to the problem of digit recognition to see how successful it could be. I ended up with an 80% success rate on a subset of the training data, but didn't bother tuning the parameters much to see if I could do any better (which I think I could, and might get around to doing later). If you're interested in looking over the results, I posted about it on my blog.
", "rawMarkdown": "Not sure what this is what you are talking about, but I recently tried applying a dynamic time warping solution to the problem of digit recognition to see how successful it could be. I ended up with an 80% success rate on a subset of the training data, but didn't bother tuning the parameters much to see if I could do any better (which I think I could, and might get around to doing later). If you're interested in looking over the results, I posted about it on [my blog][1].\r\n\r\n\r\n [1]: http://tphinkle.github.io/MNIST-Digit-Recognizer/" }, { "id": 105106, "postDate": "2016-01-20T02:54:19.060Z", "content": "Just for fun, and to see how far it can go, I was thinking about borrowing and stacking on top of a few boundary representation techniques from the image recognition community (see 1 and 2) to approach this challenge. In 1 their shape context approach gets 96.2%. I'm happy to use neural nets as well but alternative representations/approaches could serve as input and inspire a more creative solution than a typical straight forward neural network ensemble.
\n\nFeel free to contact me if there's any interest in collaborating over github, etc. I can contribute about 8 hours a week on this. I can work in Python or R but prefer Python since I've started developing some prototype code in Python.
\n\nBackground: I'm a machine learning and data science consultant, have a masters in computer science, bachelors in electrical engineering and have 6 years of industry experience prior to starting consulting.
\n\n1 "Shape Context: A new descriptor for shape matching and object recognition," Serge Belongie, et al (2000)\n2 "Lecture 7: Shape Description (Contours) ," Bryan S. Morse, Brigham Young University
", "rawMarkdown": "Just for fun, and to see how far it can go, I was thinking about borrowing and stacking on top of a few boundary representation techniques from the image recognition community (see [1] and [2]) to approach this challenge. In [1] their shape context approach gets 96.2%. I'm happy to use neural nets as well but alternative representations/approaches could serve as input and inspire a more creative solution than a typical straight forward neural network ensemble.\r\n\r\nFeel free to contact me if there's any interest in collaborating over github, etc. I can contribute about 8 hours a week on this. I can work in Python or R but prefer Python since I've started developing some prototype code in Python.\r\n\r\nBackground: I'm a machine learning and data science consultant, have a masters in computer science, bachelors in electrical engineering and have 6 years of industry experience prior to starting consulting.\r\n\r\n[1] \"Shape Context: A new descriptor for shape matching and object recognition,\" Serge Belongie, et al (2000)\r\n[2] \"Lecture 7: Shape Description (Contours) ,\" Bryan S. Morse, Brigham Young University\r\n\r\n [1]: https://www.researchgate.net/profile/Serge_Belongie/publication/2353535_Shape_Context_A_new_descriptor_for_shape_matching_and_object_recognition/links/0912f50b3498fd04df000000.pdf\r\n\r\n [2]: http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/MORSE/boundary-rep-desc.pdf" } ], "comments": [ { "id": 107447, "author_name": "Kwame", "author_url": "", "post_date": "2016-02-09T22:42:37.133000", "content": "Anyone else interested in joining this team? There's a git repo here and I've set up a slack channel. Currently we're doing EDA and I've written up a rough team document on roles and tasks so people can feel at least somewhat organized.
", "votes": 0, "replies": [] }, { "id": 105807, "author_name": "Kwame", "author_url": "", "post_date": "2016-01-26T20:49:57.887000", "content": "That's okay, Ben helped me out and I was able to send you an invite. Be sure to edit your post and take out your email address so it's not public. I'll follow up in an email.
", "votes": 0, "replies": [] }, { "id": 105803, "author_name": "Lucy Cao", "author_url": "", "post_date": "2016-01-26T20:34:20.083000", "content": "Hi, I saw the issue you posted. Do you need me to offer you my email address?
", "votes": 0, "replies": [] }, { "id": 105802, "author_name": "Kwame", "author_url": "", "post_date": "2016-01-26T20:15:59.613000", "content": "Yep, still looking. There appears to a bug with team creation (Invite Someone to a Team) so I've contacted @benhammer to add you to the team.
", "votes": 0, "replies": [] }, { "id": 105768, "author_name": "Lucy Cao", "author_url": "", "post_date": "2016-01-26T16:35:46.080000", "content": "Hi，are you still looking for collaborators in this problem?I am interested in joining you. I am a Master student focussing on statistics & ML. I can work in Python & R.
", "votes": 0, "replies": [] }, { "id": 105107, "author_name": "PrestonH", "author_url": "", "post_date": "2016-01-20T03:03:50.757000", "content": "Not sure what this is what you are talking about, but I recently tried applying a dynamic time warping solution to the problem of digit recognition to see how successful it could be. I ended up with an 80% success rate on a subset of the training data, but didn't bother tuning the parameters much to see if I could do any better (which I think I could, and might get around to doing later). If you're interested in looking over the results, I posted about it on my blog.
", "votes": 0, "replies": [] } ], "raw_markdown_by_id": { "107447": "Anyone else interested in joining this team? There's a [git repo here][1] and I've set up a slack channel. Currently we're doing EDA and I've written up a rough team document on roles and tasks so people can feel at least somewhat organized.\r\n\r\n\r\n [1]: https://github.com/robinsonkwame/DigitRecognizer", "105807": "That's okay, Ben helped me out and I was able to send you an invite. Be sure to edit your post and take out your email address so it's not public. I'll follow up in an email.", "105803": "Hi, I saw the issue you posted. Do you need me to offer you my email address? ", "105802": "Yep, still looking. There appears to a bug with team creation ([Invite Someone to a Team][1]) so I've contacted @benhammer to add you to the team.\r\n\r\n\r\n [1]: https://www.kaggle.com/forums/f/15/kaggle-forum/t/14111/invite-someone-to-a-team/77627", "105768": "Hi，are you still looking for collaborators in this problem?I am interested in joining you. I am a Master student focussing on statistics & ML. I can work in Python & R.", "105107": "Not sure what this is what you are talking about, but I recently tried applying a dynamic time warping solution to the problem of digit recognition to see how successful it could be. I ended up with an 80% success rate on a subset of the training data, but didn't bother tuning the parameters much to see if I could do any better (which I think I could, and might get around to doing later). If you're interested in looking over the results, I posted about it on [my blog][1].\r\n\r\n\r\n [1]: http://tphinkle.github.io/MNIST-Digit-Recognizer/", "105106": "Just for fun, and to see how far it can go, I was thinking about borrowing and stacking on top of a few boundary representation techniques from the image recognition community (see [1] and [2]) to approach this challenge. In [1] their shape context approach gets 96.2%. I'm happy to use neural nets as well but alternative representations/approaches could serve as input and inspire a more creative solution than a typical straight forward neural network ensemble.\r\n\r\nFeel free to contact me if there's any interest in collaborating over github, etc. I can contribute about 8 hours a week on this. I can work in Python or R but prefer Python since I've started developing some prototype code in Python.\r\n\r\nBackground: I'm a machine learning and data science consultant, have a masters in computer science, bachelors in electrical engineering and have 6 years of industry experience prior to starting consulting.\r\n\r\n[1] \"Shape Context: A new descriptor for shape matching and object recognition,\" Serge Belongie, et al (2000)\r\n[2] \"Lecture 7: Shape Description (Contours) ,\" Bryan S. Morse, Brigham Young University\r\n\r\n [1]: https://www.researchgate.net/profile/Serge_Belongie/publication/2353535_Shape_Context_A_new_descriptor_for_shape_matching_and_object_recognition/links/0912f50b3498fd04df000000.pdf\r\n\r\n [2]: http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/MORSE/boundary-rep-desc.pdf" } }