I've been thinking about how to use the false positive data to improve my classifier, but all I can think is removing the fp from the training data. Is there a way to actually utilize this kind of data in a multi-class classifier?
\nI've read this post: https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/200562 and the tip that was given was to \"use them as hard negative examples\", what that actually means and how can I implement it in practice?
\nThanks.
", "messages": [ { "id": "1135266", "postDate": "01/02/2021 04:57:40", "content": "I've been thinking about how to use the false positive data to improve my classifier, but all I can think is removing the fp from the training data. Is there a way to actually utilize this kind of data in a multi-class classifier?
\nI've read this post: https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/200562 and the tip that was given was to \"use them as hard negative examples\", what that actually means and how can I implement it in practice?
\nThanks.
", "rawMarkdown": "I've been thinking about how to use the false positive data to improve my classifier, but all I can think is removing the fp from the training data. Is there a way to actually utilize this kind of data in a multi-class classifier? \n\nI've read this post: https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/200562 and the tip that was given was to \"use them as hard negative examples\", what that actually means and how can I implement it in practice? \n\nThanks.", "votes": null }, { "id": "1135509", "postDate": "01/02/2021 10:01:30", "content": "Hard negatives are tricky samples that models generally struggle to get correct. If you were to (over) sample the false positives during training, your model might be able to learn a better decision boundary.
", "rawMarkdown": "Hard negatives are tricky samples that models generally struggle to get correct. If you were to (over) sample the false positives during training, your model might be able to learn a better decision boundary.", "votes": null }, { "id": "1160126", "postDate": "01/19/2021 17:21:56", "content": "Same question for me, in single-class classification task we can easily use fp sample as negative sample and feed to network to train. But for multi-class classification, shall we label such samples as all 24 zeros and feed to network? Hope someone can help on this.
", "rawMarkdown": "Same question for me, in single-class classification task we can easily use fp sample as negative sample and feed to network to train. But for multi-class classification, shall we label such samples as all 24 zeros and feed to network? Hope someone can help on this.", "votes": null }, { "id": "1252505", "postDate": "03/25/2021 18:18:30", "content": "For anyone interested, there is a way to include false positives in a multi-class classification task, that being, modifying the loss function to accept two different types of representations, for instance, one can represent the classes with hot-one-encodings and use \"1\" or \"0\" to represent that the class is there or not, while the other classes are marked as \"unknown”. There is a great paper in which the authors use this approach: https://www.sciencedirect.com/science/article/abs/pii/S0003682X20304795
", "rawMarkdown": "For anyone interested, there is a way to include false positives in a multi-class classification task, that being, modifying the loss function to accept two different types of representations, for instance, one can represent the classes with hot-one-encodings and use \"1\" or \"0\" to represent that the class is there or not, while the other classes are marked as \"unknown”. There is a great paper in which the authors use this approach: https://www.sciencedirect.com/science/article/abs/pii/S0003682X20304795", "votes": null } ], "comments": [ { "id": 1135509, "author_name": "reppic", "author_url": "", "post_date": "01/02/2021 10:01:30", "content": "Hard negatives are tricky samples that models generally struggle to get correct. If you were to (over) sample the false positives during training, your model might be able to learn a better decision boundary.
", "votes": null, "replies": [] }, { "id": 1160126, "author_name": "superchenhao", "author_url": "", "post_date": "01/19/2021 17:21:56", "content": "Same question for me, in single-class classification task we can easily use fp sample as negative sample and feed to network to train. But for multi-class classification, shall we label such samples as all 24 zeros and feed to network? Hope someone can help on this.
", "votes": null, "replies": [] }, { "id": 1252505, "author_name": "alefiury", "author_url": "", "post_date": "03/25/2021 18:18:30", "content": "For anyone interested, there is a way to include false positives in a multi-class classification task, that being, modifying the loss function to accept two different types of representations, for instance, one can represent the classes with hot-one-encodings and use \"1\" or \"0\" to represent that the class is there or not, while the other classes are marked as \"unknown”. There is a great paper in which the authors use this approach: https://www.sciencedirect.com/science/article/abs/pii/S0003682X20304795
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "1135266": "I've been thinking about how to use the false positive data to improve my classifier, but all I can think is removing the fp from the training data. Is there a way to actually utilize this kind of data in a multi-class classifier? \n\nI've read this post: https://www.kaggle.com/c/rfcx-species-audio-detection/discussion/200562 and the tip that was given was to \"use them as hard negative examples\", what that actually means and how can I implement it in practice? \n\nThanks.", "1135509": "Hard negatives are tricky samples that models generally struggle to get correct. If you were to (over) sample the false positives during training, your model might be able to learn a better decision boundary.", "1160126": "Same question for me, in single-class classification task we can easily use fp sample as negative sample and feed to network to train. But for multi-class classification, shall we label such samples as all 24 zeros and feed to network? Hope someone can help on this.", "1252505": "For anyone interested, there is a way to include false positives in a multi-class classification task, that being, modifying the loss function to accept two different types of representations, for instance, one can represent the classes with hot-one-encodings and use \"1\" or \"0\" to represent that the class is there or not, while the other classes are marked as \"unknown”. There is a great paper in which the authors use this approach: https://www.sciencedirect.com/science/article/abs/pii/S0003682X20304795" }, "source": "meta" }