The noisy subset may be a key point for winning the competition, there are some related papers:
\n\nUnsupervised or semi-supervised learning
\n\nOne approach might be to consider the noisy data as totally unlabeled and use some unsupervised technics to learn useful representations.
\n\nPapers:
\n\nhttps://arxiv.org/abs/1803.11095\nhttps://arxiv.org/abs/1808.06670\nhttps://arxiv.org/abs/1507.02672
\n\nLearning from noisy labels
\n\nAnother approach is to use existing labels but also take some steps in order to combat label noise.
\n\nPapers:
\n\nhttps://arxiv.org/abs/1903.07507\nhttps://arxiv.org/abs/1705.10694\nhttps://papers.nips.cc/paper/5073-learning-with-noisy-labels.pdf
\n\nFeel free to contribute.
\n\nGood luck!
", "messages": [ { "id": "523700", "postDate": "04/26/2019 19:08:55", "content": "The noisy subset may be a key point for winning the competition, there are some related papers:
\n\nUnsupervised or semi-supervised learning
\n\nOne approach might be to consider the noisy data as totally unlabeled and use some unsupervised technics to learn useful representations.
\n\nPapers:
\n\nhttps://arxiv.org/abs/1803.11095\nhttps://arxiv.org/abs/1808.06670\nhttps://arxiv.org/abs/1507.02672
\n\nLearning from noisy labels
\n\nAnother approach is to use existing labels but also take some steps in order to combat label noise.
\n\nPapers:
\n\nhttps://arxiv.org/abs/1903.07507\nhttps://arxiv.org/abs/1705.10694\nhttps://papers.nips.cc/paper/5073-learning-with-noisy-labels.pdf
\n\nFeel free to contribute.
\n\nGood luck!
", "rawMarkdown": "The noisy subset may be a key point for winning the competition, there are some related papers:\n\n**Unsupervised or semi-supervised learning**\n\nOne approach might be to consider the noisy data as totally unlabeled and use some unsupervised technics to learn useful representations. \n\nPapers:\n\nhttps://arxiv.org/abs/1803.11095\nhttps://arxiv.org/abs/1808.06670\nhttps://arxiv.org/abs/1507.02672\n\n**Learning from noisy labels**\n\nAnother approach is to use existing labels but also take some steps in order to combat label noise.\n\nPapers:\n\nhttps://arxiv.org/abs/1903.07507\nhttps://arxiv.org/abs/1705.10694\nhttps://papers.nips.cc/paper/5073-learning-with-noisy-labels.pdf\n\nFeel free to contribute.\n\nGood luck!", "votes": null }, { "id": "523706", "postDate": "04/26/2019 19:56:07", "content": "One more to the \"learning from noisy labels\" list:\nhttps://arxiv.org/abs/1901.01189
", "rawMarkdown": "One more to the \"learning from noisy labels\" list:\nhttps://arxiv.org/abs/1901.01189", "votes": null }, { "id": "523752", "postDate": "04/26/2019 23:29:49", "content": "Oh this is written by this competition hosts.
", "rawMarkdown": "Oh this is written by this competition hosts.", "votes": null }, { "id": "523783", "postDate": "04/27/2019 02:08:44", "content": "I viewed all given links those are nice and also helpful for us.... Thanks for sharing!!
", "rawMarkdown": "I viewed all given links those are nice and also helpful for us.... Thanks for sharing!!", "votes": null }, { "id": "525602", "postDate": "05/01/2019 11:24:36", "content": "https://arxiv.org/pdf/1712.09482.pdf
\n\nThis multi-class related paper, but interesting to show that MAE loss is robust under label noise.\n'...we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'
\n\nUsually BCE loss would be used for multi-label solutions by default...
", "rawMarkdown": "https://arxiv.org/pdf/1712.09482.pdf\n\nThis multi-class related paper, but interesting to show that MAE loss is robust under label noise.\n'...we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'\n\nUsually BCE loss would be used for multi-label solutions by default...", "votes": null }, { "id": "525890", "postDate": "05/02/2019 00:01:46", "content": "My summary of paper readings.
\n\nhttps://arxiv.org/pdf/1901.01189.pdf\nLEARNING SOUND EVENT CLASSIFIERS FROM WEB AUDIO WITH NOISY LABELS\n- This is from competition hosts (thanks!), using maybe almost the same data with single labels.\n- Using all training set including noisy data improved performance greatly.\n- Couple of loss functions expecting label noise robustness are compared, which we could derive something.
\n\nhttps://www.researchgate.net/publication/328927908_Audio_Tagging_System_Using_Densely_Connected_Convolutional_Networks\nAUDIO TAGGING SYSTEM USING DENSELY CONNECTED CONVOLUTIONAL NETWORKS, 2018\n- _ Batch-wise loss masking_ is used in last year's Freesound Audio Tagging 2018 1st place solution. This is also applicable.
\n\nhttps://arxiv.org/pdf/1712.09482.pdf\nRobust Loss Functions under Label Noise for Deep Neural Networks, 2017\n- 'we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'
\n\nhttps://arxiv.org/pdf/1802.05031.pdf\nTackling Multilabel Imbalance through Label Decoupling and Data Resampling Hybridization, 2018\nhttps://arxiv.org/pdf/1802.05033.pdf\nDealing with Difficult Minority Labels in Imbalanced Mutilabel Data Sets, 2018\n- Empirical research on combination of decoupling of label concurrence and resampling.\n- Need to be careful using label decoupling, depending on conditions like type of classifiers or metrics. But sounds like applicable.
", "rawMarkdown": "My summary of paper readings.\n\nhttps://arxiv.org/pdf/1901.01189.pdf\nLEARNING SOUND EVENT CLASSIFIERS FROM WEB AUDIO WITH NOISY LABELS\n- This is from competition hosts (thanks!), using maybe almost the same data with single labels.\n- Using all training set including noisy data improved performance greatly.\n- Couple of loss functions expecting label noise robustness are compared, which we could derive something.\n\nhttps://www.researchgate.net/publication/328927908_Audio_Tagging_System_Using_Densely_Connected_Convolutional_Networks\nAUDIO TAGGING SYSTEM USING DENSELY CONNECTED CONVOLUTIONAL NETWORKS, 2018\n- _ Batch-wise loss masking_ is used in last year's Freesound Audio Tagging 2018 1st place solution. This is also applicable.\n\nhttps://arxiv.org/pdf/1712.09482.pdf\nRobust Loss Functions under Label Noise for Deep Neural Networks, 2017\n- 'we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'\n\nhttps://arxiv.org/pdf/1802.05031.pdf\nTackling Multilabel Imbalance through Label Decoupling and Data Resampling Hybridization, 2018\nhttps://arxiv.org/pdf/1802.05033.pdf\nDealing with Difficult Minority Labels in Imbalanced Mutilabel Data Sets, 2018\n- Empirical research on combination of decoupling of label concurrence and resampling.\n- Need to be careful using label decoupling, depending on conditions like type of classifiers or metrics. But sounds like applicable.", "votes": null }, { "id": "526324", "postDate": "05/02/2019 20:13:09", "content": "Anyone made use of label smoothing on the noisy set yet? Been working on doing this with Fastai but running into errors around multi-classification.
", "rawMarkdown": "Anyone made use of label smoothing on the noisy set yet? Been working on doing this with Fastai but running into errors around multi-classification.", "votes": null }, { "id": "616156", "postDate": "09/02/2019 18:46:17", "content": "Did you ever figure this out? Working on a different problem and am struggling with this as well. Thank you!
", "rawMarkdown": "Did you ever figure this out? Working on a different problem and am struggling with this as well. Thank you!", "votes": null }, { "id": "1686206", "postDate": "02/11/2022 21:05:09", "content": "I found the link below with the most implemented papers for audio tagging, including the FSDKaggle2018, FSDKaggle2019 datasets, and more.
\nhttps://paperswithcode.com/task/audio-tagging
\nFor those who’d like to test different model performances, I’d suggest deepchecks which helps you do the data and model checks less effortlessly. Good luck!
", "rawMarkdown": "I found the link below with the most implemented papers for audio tagging, including the FSDKaggle2018, FSDKaggle2019 datasets, and more.\n\nhttps://paperswithcode.com/task/audio-tagging [](url)\n\nFor those who’d like to test different model performances, I’d suggest [deepchecks](https://github.com/deepchecks) which helps you do the data and model checks less effortlessly. Good luck!", "votes": null } ], "comments": [ { "id": 1686206, "author_name": "arnoldholt", "author_url": "", "post_date": "02/11/2022 21:05:09", "content": "I found the link below with the most implemented papers for audio tagging, including the FSDKaggle2018, FSDKaggle2019 datasets, and more.
\nhttps://paperswithcode.com/task/audio-tagging
\nFor those who’d like to test different model performances, I’d suggest deepchecks which helps you do the data and model checks less effortlessly. Good luck!
", "votes": null, "replies": [] }, { "id": 523706, "author_name": "mnpinto", "author_url": "", "post_date": "04/26/2019 19:56:07", "content": "One more to the \"learning from noisy labels\" list:\nhttps://arxiv.org/abs/1901.01189
", "votes": null, "replies": [ { "id": 523752, "author_name": "daisukelab", "author_url": "", "post_date": "04/26/2019 23:29:49", "content": "Oh this is written by this competition hosts.
", "votes": null, "replies": [] } ] }, { "id": 523783, "author_name": "himusoni", "author_url": "", "post_date": "04/27/2019 02:08:44", "content": "I viewed all given links those are nice and also helpful for us.... Thanks for sharing!!
", "votes": null, "replies": [] }, { "id": 525602, "author_name": "daisukelab", "author_url": "", "post_date": "05/01/2019 11:24:36", "content": "https://arxiv.org/pdf/1712.09482.pdf
\n\nThis multi-class related paper, but interesting to show that MAE loss is robust under label noise.\n'...we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'
\n\nUsually BCE loss would be used for multi-label solutions by default...
", "votes": null, "replies": [] }, { "id": 525890, "author_name": "daisukelab", "author_url": "", "post_date": "05/02/2019 00:01:46", "content": "My summary of paper readings.
\n\nhttps://arxiv.org/pdf/1901.01189.pdf\nLEARNING SOUND EVENT CLASSIFIERS FROM WEB AUDIO WITH NOISY LABELS\n- This is from competition hosts (thanks!), using maybe almost the same data with single labels.\n- Using all training set including noisy data improved performance greatly.\n- Couple of loss functions expecting label noise robustness are compared, which we could derive something.
\n\nhttps://www.researchgate.net/publication/328927908_Audio_Tagging_System_Using_Densely_Connected_Convolutional_Networks\nAUDIO TAGGING SYSTEM USING DENSELY CONNECTED CONVOLUTIONAL NETWORKS, 2018\n- _ Batch-wise loss masking_ is used in last year's Freesound Audio Tagging 2018 1st place solution. This is also applicable.
\n\nhttps://arxiv.org/pdf/1712.09482.pdf\nRobust Loss Functions under Label Noise for Deep Neural Networks, 2017\n- 'we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'
\n\nhttps://arxiv.org/pdf/1802.05031.pdf\nTackling Multilabel Imbalance through Label Decoupling and Data Resampling Hybridization, 2018\nhttps://arxiv.org/pdf/1802.05033.pdf\nDealing with Difficult Minority Labels in Imbalanced Mutilabel Data Sets, 2018\n- Empirical research on combination of decoupling of label concurrence and resampling.\n- Need to be careful using label decoupling, depending on conditions like type of classifiers or metrics. But sounds like applicable.
", "votes": null, "replies": [] }, { "id": 526324, "author_name": "zacharyneveu", "author_url": "", "post_date": "05/02/2019 20:13:09", "content": "Anyone made use of label smoothing on the noisy set yet? Been working on doing this with Fastai but running into errors around multi-classification.
", "votes": null, "replies": [ { "id": 616156, "author_name": "zacheism", "author_url": "", "post_date": "09/02/2019 18:46:17", "content": "Did you ever figure this out? Working on a different problem and am struggling with this as well. Thank you!
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "523700": "The noisy subset may be a key point for winning the competition, there are some related papers:\n\n**Unsupervised or semi-supervised learning**\n\nOne approach might be to consider the noisy data as totally unlabeled and use some unsupervised technics to learn useful representations. \n\nPapers:\n\nhttps://arxiv.org/abs/1803.11095\nhttps://arxiv.org/abs/1808.06670\nhttps://arxiv.org/abs/1507.02672\n\n**Learning from noisy labels**\n\nAnother approach is to use existing labels but also take some steps in order to combat label noise.\n\nPapers:\n\nhttps://arxiv.org/abs/1903.07507\nhttps://arxiv.org/abs/1705.10694\nhttps://papers.nips.cc/paper/5073-learning-with-noisy-labels.pdf\n\nFeel free to contribute.\n\nGood luck!", "523706": "One more to the \"learning from noisy labels\" list:\nhttps://arxiv.org/abs/1901.01189", "523752": "Oh this is written by this competition hosts.", "523783": "I viewed all given links those are nice and also helpful for us.... Thanks for sharing!!", "525602": "https://arxiv.org/pdf/1712.09482.pdf\n\nThis multi-class related paper, but interesting to show that MAE loss is robust under label noise.\n'...we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'\n\nUsually BCE loss would be used for multi-label solutions by default...", "525890": "My summary of paper readings.\n\nhttps://arxiv.org/pdf/1901.01189.pdf\nLEARNING SOUND EVENT CLASSIFIERS FROM WEB AUDIO WITH NOISY LABELS\n- This is from competition hosts (thanks!), using maybe almost the same data with single labels.\n- Using all training set including noisy data improved performance greatly.\n- Couple of loss functions expecting label noise robustness are compared, which we could derive something.\n\nhttps://www.researchgate.net/publication/328927908_Audio_Tagging_System_Using_Densely_Connected_Convolutional_Networks\nAUDIO TAGGING SYSTEM USING DENSELY CONNECTED CONVOLUTIONAL NETWORKS, 2018\n- _ Batch-wise loss masking_ is used in last year's Freesound Audio Tagging 2018 1st place solution. This is also applicable.\n\nhttps://arxiv.org/pdf/1712.09482.pdf\nRobust Loss Functions under Label Noise for Deep Neural Networks, 2017\n- 'we showed that the commonly used CCE loss is sensitive to label noise while MAE loss is robust'\n\nhttps://arxiv.org/pdf/1802.05031.pdf\nTackling Multilabel Imbalance through Label Decoupling and Data Resampling Hybridization, 2018\nhttps://arxiv.org/pdf/1802.05033.pdf\nDealing with Difficult Minority Labels in Imbalanced Mutilabel Data Sets, 2018\n- Empirical research on combination of decoupling of label concurrence and resampling.\n- Need to be careful using label decoupling, depending on conditions like type of classifiers or metrics. But sounds like applicable.", "526324": "Anyone made use of label smoothing on the noisy set yet? Been working on doing this with Fastai but running into errors around multi-classification.", "616156": "Did you ever figure this out? Working on a different problem and am struggling with this as well. Thank you!", "1686206": "I found the link below with the most implemented papers for audio tagging, including the FSDKaggle2018, FSDKaggle2019 datasets, and more.\n\nhttps://paperswithcode.com/task/audio-tagging [](url)\n\nFor those who’d like to test different model performances, I’d suggest [deepchecks](https://github.com/deepchecks) which helps you do the data and model checks less effortlessly. Good luck!" }, "source": "meta" }