Thank you all for an exciting competition! It was great to see such activity during the last two months. Quick Excel analysis (sorry not Python) shows total of 686 submissions, of which 582 from local TUT teams.
\n\nIn particular, I would like to congratulate the competition winner @daisukelab and the best TUT team #11: Eetu, Tuukka, Jrm and Lukas.
\n\nI know everyone is eagerly waiting to find out what the top teams did. So please share your insight to the entire community. Not only the two teams mentioned above, but everyone. Either in this thread or in the kernels section.
\n\nWe are also considering a research paper about the results. One possible forum would be IEEE MLSP 2018 conference, with deadline in a few months; similar to the one from two years ago. Huttunen et al., SCIA 2015. Interested top teams are invited to join as authors if there seems to be enough potential for a paper.
", "messages": [ { "id": "289784", "postDate": "03/02/2018 09:02:12", "content": "Thank you all for an exciting competition! It was great to see such activity during the last two months. Quick Excel analysis (sorry not Python) shows total of 686 submissions, of which 582 from local TUT teams.
\n\nIn particular, I would like to congratulate the competition winner @daisukelab and the best TUT team #11: Eetu, Tuukka, Jrm and Lukas.
\n\nI know everyone is eagerly waiting to find out what the top teams did. So please share your insight to the entire community. Not only the two teams mentioned above, but everyone. Either in this thread or in the kernels section.
\n\nWe are also considering a research paper about the results. One possible forum would be IEEE MLSP 2018 conference, with deadline in a few months; similar to the one from two years ago. Huttunen et al., SCIA 2015. Interested top teams are invited to join as authors if there seems to be enough potential for a paper.
", "rawMarkdown": "Thank you all for an exciting competition! It was great to see such activity during the last two months. Quick Excel analysis (sorry not Python) shows total of 686 submissions, of which 582 from local TUT teams.\n\nIn particular, I would like to congratulate the competition winner @daisukelab and the best TUT team #11: Eetu, Tuukka, Jrm and Lukas. \n\nI know everyone is eagerly waiting to find out what the top teams did. So please share your insight to the entire community. Not only the two teams mentioned above, but everyone. Either in this thread or in the kernels section.\n\nWe are also considering a research paper about the results. One possible forum would be [IEEE MLSP 2018 conference][1], with deadline in a few months; similar to the one from two years ago. [Huttunen et al., SCIA 2015][2]. Interested top teams are invited to join as authors if there seems to be enough potential for a paper.\n\n [1]: http://mlsp2018.conwiz.dk/home.htm\n [2]: http://vision.cs.tut.fi/data/publications/scia2015_hpe.pdf", "votes": null }, { "id": "290774", "postDate": "03/04/2018 21:15:30", "content": "TUT team 11: Eetu, Tuukka, Jorma, Lucas
\n\nThanks for the course and competition. There were too much to learn and do for this short time, but it was interesting.
\n\nOur plan: use multiple non-neuronetwork classifiers and combine their probabilities. Input data will be different features by frequencys (discarding time related ones) and possibly some other features used by sound/music analyzers. Convolutional networks can extract more specific time-related patterns. And since data is relatively small compared to testdata, adding confident test data labels to training data should be efficient (semisupervised learning).
\n\nNotes during process: Semisupervision needs some kind of confidence with the minimum probability to be more efficient. The semisupervision process can be run multiple times to be bit more efficient, but the next loops shouldn't overwrite the previous predicts to avoid relabeling anything based on new lower confidence data.\nConvolutional neuronetworks never got over 73% by themselves.
\n\nFinal product: Classifiers (Randomforest, Logisticregression, LDA and SVM with linear and rbf kernels), features by frequency: (mean, deviation, median, maximum and minimum), one small and one larger convnetwork.\nsemisupervision: probablity over 0.5 and no other probability over 0.3 (confidence), looped the semisupervision 3 times. Neuronets weren't teached multiple times.
\n\nTODO for better accuracy: more features and cut out the confusers, teach all networks in the semisupervision loops, try new networks as they should achieve atleast the same results by themselves.
\n\naccuracy progress with new techniques:
\n\n72% using multiple classifiers, data averaged for frequencys.
\n\n75% added new features of data: deviation, median, min, max, skew, kurtosis for each frequency. Kurtosis and skewness turned out to be confusers as they were time dependent.
\n\n77% added semisupervised learning
\n\n78% added confidence threshold to semisupervision
\n\n79% looped the semisupervision more
\n\n79.5% made semisupervised learning not to overwrite its previous predictions
\n\n80% combined with larger convnet (Jormas product)
\n\nMost of the (messy) code can be found on my github: https://github.com/senttula/music-categorizer\nI take credit for the code in my github, but plan and ideas we came up as a team.
\n\n-Tuukka Senttula
\n\nTeam: Eetu, Tuukka, Jorma, Lucas
", "rawMarkdown": "TUT team 11: Eetu, Tuukka, Jorma, Lucas\n\nThanks for the course and competition. There were too much to learn and do for this short time, but it was interesting.\n\nOur plan: use multiple non-neuronetwork classifiers and combine their probabilities. Input data will be different features by frequencys (discarding time related ones) and possibly some other features used by sound/music analyzers. Convolutional networks can extract more specific time-related patterns. And since data is relatively small compared to testdata, adding confident test data labels to training data should be efficient (semisupervised learning).\n\nNotes during process: Semisupervision needs some kind of confidence with the minimum probability to be more efficient. The semisupervision process can be run multiple times to be bit more efficient, but the next loops shouldn't overwrite the previous predicts to avoid relabeling anything based on new lower confidence data.\nConvolutional neuronetworks never got over 73% by themselves.\n\nFinal product: Classifiers (Randomforest, Logisticregression, LDA and SVM with linear and rbf kernels), features by frequency: (mean, deviation, median, maximum and minimum), one small and one larger convnetwork.\nsemisupervision: probablity over 0.5 and no other probability over 0.3 (confidence), looped the semisupervision 3 times. Neuronets weren't teached multiple times.\n\nTODO for better accuracy: more features and cut out the confusers, teach all networks in the semisupervision loops, try new networks as they should achieve atleast the same results by themselves.\n\naccuracy progress with new techniques:\n\n72% using multiple classifiers, data averaged for frequencys.\n\n75% added new features of data: deviation, median, min, max, skew, kurtosis for each frequency. Kurtosis and skewness turned out to be confusers as they were time dependent.\n\n77% added semisupervised learning\n\n78% added confidence threshold to semisupervision\n\n79% looped the semisupervision more\n\n79.5% made semisupervised learning not to overwrite its previous predictions\n\n80% combined with larger convnet (Jormas product)\n\nMost of the (messy) code can be found on my github: https://github.com/senttula/music-categorizer\nI take credit for the code in my github, but plan and ideas we came up as a team.\n\n-Tuukka Senttula\n\nTeam: Eetu, Tuukka, Jorma, Lucas", "votes": null } ], "comments": [ { "id": 290774, "author_name": "senttula", "author_url": "", "post_date": "03/04/2018 21:15:30", "content": "TUT team 11: Eetu, Tuukka, Jorma, Lucas
\n\nThanks for the course and competition. There were too much to learn and do for this short time, but it was interesting.
\n\nOur plan: use multiple non-neuronetwork classifiers and combine their probabilities. Input data will be different features by frequencys (discarding time related ones) and possibly some other features used by sound/music analyzers. Convolutional networks can extract more specific time-related patterns. And since data is relatively small compared to testdata, adding confident test data labels to training data should be efficient (semisupervised learning).
\n\nNotes during process: Semisupervision needs some kind of confidence with the minimum probability to be more efficient. The semisupervision process can be run multiple times to be bit more efficient, but the next loops shouldn't overwrite the previous predicts to avoid relabeling anything based on new lower confidence data.\nConvolutional neuronetworks never got over 73% by themselves.
\n\nFinal product: Classifiers (Randomforest, Logisticregression, LDA and SVM with linear and rbf kernels), features by frequency: (mean, deviation, median, maximum and minimum), one small and one larger convnetwork.\nsemisupervision: probablity over 0.5 and no other probability over 0.3 (confidence), looped the semisupervision 3 times. Neuronets weren't teached multiple times.
\n\nTODO for better accuracy: more features and cut out the confusers, teach all networks in the semisupervision loops, try new networks as they should achieve atleast the same results by themselves.
\n\naccuracy progress with new techniques:
\n\n72% using multiple classifiers, data averaged for frequencys.
\n\n75% added new features of data: deviation, median, min, max, skew, kurtosis for each frequency. Kurtosis and skewness turned out to be confusers as they were time dependent.
\n\n77% added semisupervised learning
\n\n78% added confidence threshold to semisupervision
\n\n79% looped the semisupervision more
\n\n79.5% made semisupervised learning not to overwrite its previous predictions
\n\n80% combined with larger convnet (Jormas product)
\n\nMost of the (messy) code can be found on my github: https://github.com/senttula/music-categorizer\nI take credit for the code in my github, but plan and ideas we came up as a team.
\n\n-Tuukka Senttula
\n\nTeam: Eetu, Tuukka, Jorma, Lucas
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "289784": "Thank you all for an exciting competition! It was great to see such activity during the last two months. Quick Excel analysis (sorry not Python) shows total of 686 submissions, of which 582 from local TUT teams.\n\nIn particular, I would like to congratulate the competition winner @daisukelab and the best TUT team #11: Eetu, Tuukka, Jrm and Lukas. \n\nI know everyone is eagerly waiting to find out what the top teams did. So please share your insight to the entire community. Not only the two teams mentioned above, but everyone. Either in this thread or in the kernels section.\n\nWe are also considering a research paper about the results. One possible forum would be [IEEE MLSP 2018 conference][1], with deadline in a few months; similar to the one from two years ago. [Huttunen et al., SCIA 2015][2]. Interested top teams are invited to join as authors if there seems to be enough potential for a paper.\n\n [1]: http://mlsp2018.conwiz.dk/home.htm\n [2]: http://vision.cs.tut.fi/data/publications/scia2015_hpe.pdf", "290774": "TUT team 11: Eetu, Tuukka, Jorma, Lucas\n\nThanks for the course and competition. There were too much to learn and do for this short time, but it was interesting.\n\nOur plan: use multiple non-neuronetwork classifiers and combine their probabilities. Input data will be different features by frequencys (discarding time related ones) and possibly some other features used by sound/music analyzers. Convolutional networks can extract more specific time-related patterns. And since data is relatively small compared to testdata, adding confident test data labels to training data should be efficient (semisupervised learning).\n\nNotes during process: Semisupervision needs some kind of confidence with the minimum probability to be more efficient. The semisupervision process can be run multiple times to be bit more efficient, but the next loops shouldn't overwrite the previous predicts to avoid relabeling anything based on new lower confidence data.\nConvolutional neuronetworks never got over 73% by themselves.\n\nFinal product: Classifiers (Randomforest, Logisticregression, LDA and SVM with linear and rbf kernels), features by frequency: (mean, deviation, median, maximum and minimum), one small and one larger convnetwork.\nsemisupervision: probablity over 0.5 and no other probability over 0.3 (confidence), looped the semisupervision 3 times. Neuronets weren't teached multiple times.\n\nTODO for better accuracy: more features and cut out the confusers, teach all networks in the semisupervision loops, try new networks as they should achieve atleast the same results by themselves.\n\naccuracy progress with new techniques:\n\n72% using multiple classifiers, data averaged for frequencys.\n\n75% added new features of data: deviation, median, min, max, skew, kurtosis for each frequency. Kurtosis and skewness turned out to be confusers as they were time dependent.\n\n77% added semisupervised learning\n\n78% added confidence threshold to semisupervision\n\n79% looped the semisupervision more\n\n79.5% made semisupervised learning not to overwrite its previous predictions\n\n80% combined with larger convnet (Jormas product)\n\nMost of the (messy) code can be found on my github: https://github.com/senttula/music-categorizer\nI take credit for the code in my github, but plan and ideas we came up as a team.\n\n-Tuukka Senttula\n\nTeam: Eetu, Tuukka, Jorma, Lucas" }, "source": "meta" }