I have managed to get over the frustration of how much time I have put into getting a good CV/LB but now am more just curious. Why is it that this data is so different from the test data? Is there some reason why it makes sense for CORNELL LAB OF ORNITHOLOGY to structure the comp in this way? I am sure that their is, but I can't quite think of a reason?
", "messages": [ { "id": "2829798", "postDate": "05/22/2024 20:03:58", "content": "I have managed to get over the frustration of how much time I have put into getting a good CV/LB but now am more just curious. Why is it that this data is so different from the test data? Is there some reason why it makes sense for CORNELL LAB OF ORNITHOLOGY to structure the comp in this way? I am sure that their is, but I can't quite think of a reason?
", "rawMarkdown": "I have managed to get over the frustration of how much time I have put into getting a good CV/LB but now am more just curious. Why is it that this data is so different from the test data? Is there some reason why it makes sense for CORNELL LAB OF ORNITHOLOGY to structure the comp in this way? I am sure that their is, but I can't quite think of a reason?", "votes": null }, { "id": "2829853", "postDate": "05/22/2024 20:47:04", "content": "I tried to explain here: https://www.kaggle.com/competitions/birdclef-2024/discussion/498404
", "rawMarkdown": "I tried to explain here: https://www.kaggle.com/competitions/birdclef-2024/discussion/498404", "votes": null }, { "id": "2829980", "postDate": "05/22/2024 23:25:06", "content": "Then why does my no call from unlabeled soundscapes not improve my LB 😪
", "rawMarkdown": "Then why does my no call from unlabeled soundscapes not improve my LB 😪", "votes": null }, { "id": "2830011", "postDate": "05/23/2024 00:38:22", "content": "Yep, I have seen that. I follow you but I am more wondering how this could benefit the hosts rather than just having the train being the same as the data in test. Because I cannot imagine a scenario where a portion of the labeled testing data does not serve as better training data than the version we have now because it is so different and the learning is on a slightly different task.
", "rawMarkdown": "Yep, I have seen that. I follow you but I am more wondering how this could benefit the hosts rather than just having the train being the same as the data in test. Because I cannot imagine a scenario where a portion of the labeled testing data does not serve as better training data than the version we have now because it is so different and the learning is on a slightly different task.", "votes": null }, { "id": "2830353", "postDate": "05/23/2024 06:46:49", "content": "For what @CPMP is referring, probably they would like to automate the process more, e.g. put many cheap omnidirectional mics (maybe even with on-edge processing if we recall 2h runtime cap) in the area and use such data instead of crowdsourced one where people could direct the mic to the bird. By doing that, we could track e.g. population of some particular species over large periods.
\nAnd on labeling, the way it is done for LBs requires a lot of effort and not easily scalable, while there is a lot of open data with weak labels, and it scales well.
", "rawMarkdown": "For what @CPMP is referring, probably they would like to automate the process more, e.g. put many cheap omnidirectional mics (maybe even with on-edge processing if we recall 2h runtime cap) in the area and use such data instead of crowdsourced one where people could direct the mic to the bird. By doing that, we could track e.g. population of some particular species over large periods.\n\nAnd on labeling, the way it is done for LBs requires a lot of effort and not easily scalable, while there is a lot of open data with weak labels, and it scales well.", "votes": null }, { "id": "2830369", "postDate": "05/23/2024 06:53:27", "content": "Thank you, well said! That is exactly the reason for the competition. We need outside help to tackle this problem because, as you can see, it's a difficult task.
", "rawMarkdown": "Thank you, well said! That is exactly the reason for the competition. We need outside help to tackle this problem because, as you can see, it's a difficult task.", "votes": null }, { "id": "2841331", "postDate": "05/28/2024 14:28:32", "content": "I believe that test data just has more noise and worse data this would imply that the solution for this competition would be very augment-driven.
", "rawMarkdown": "I believe that test data just has more noise and worse data this would imply that the solution for this competition would be very augment-driven.", "votes": null } ], "comments": [ { "id": 2829853, "author_name": "cpmpml", "author_url": "", "post_date": "05/22/2024 20:47:04", "content": "I tried to explain here: https://www.kaggle.com/competitions/birdclef-2024/discussion/498404
", "votes": null, "replies": [ { "id": 2829980, "author_name": "willrice", "author_url": "", "post_date": "05/22/2024 23:25:06", "content": "Then why does my no call from unlabeled soundscapes not improve my LB 😪
", "votes": null, "replies": [] }, { "id": 2830011, "author_name": "cody11null", "author_url": "", "post_date": "05/23/2024 00:38:22", "content": "Yep, I have seen that. I follow you but I am more wondering how this could benefit the hosts rather than just having the train being the same as the data in test. Because I cannot imagine a scenario where a portion of the labeled testing data does not serve as better training data than the version we have now because it is so different and the learning is on a slightly different task.
", "votes": null, "replies": [ { "id": 2830353, "author_name": "mkotyushev", "author_url": "", "post_date": "05/23/2024 06:46:49", "content": "For what @CPMP is referring, probably they would like to automate the process more, e.g. put many cheap omnidirectional mics (maybe even with on-edge processing if we recall 2h runtime cap) in the area and use such data instead of crowdsourced one where people could direct the mic to the bird. By doing that, we could track e.g. population of some particular species over large periods.
\nAnd on labeling, the way it is done for LBs requires a lot of effort and not easily scalable, while there is a lot of open data with weak labels, and it scales well.
", "votes": null, "replies": [ { "id": 2830369, "author_name": "stefankahl", "author_url": "", "post_date": "05/23/2024 06:53:27", "content": "Thank you, well said! That is exactly the reason for the competition. We need outside help to tackle this problem because, as you can see, it's a difficult task.
", "votes": null, "replies": [] } ] } ] } ] }, { "id": 2841331, "author_name": "max1mum", "author_url": "", "post_date": "05/28/2024 14:28:32", "content": "I believe that test data just has more noise and worse data this would imply that the solution for this competition would be very augment-driven.
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2829798": "I have managed to get over the frustration of how much time I have put into getting a good CV/LB but now am more just curious. Why is it that this data is so different from the test data? Is there some reason why it makes sense for CORNELL LAB OF ORNITHOLOGY to structure the comp in this way? I am sure that their is, but I can't quite think of a reason?", "2829853": "I tried to explain here: https://www.kaggle.com/competitions/birdclef-2024/discussion/498404", "2829980": "Then why does my no call from unlabeled soundscapes not improve my LB 😪", "2830011": "Yep, I have seen that. I follow you but I am more wondering how this could benefit the hosts rather than just having the train being the same as the data in test. Because I cannot imagine a scenario where a portion of the labeled testing data does not serve as better training data than the version we have now because it is so different and the learning is on a slightly different task.", "2830353": "For what @CPMP is referring, probably they would like to automate the process more, e.g. put many cheap omnidirectional mics (maybe even with on-edge processing if we recall 2h runtime cap) in the area and use such data instead of crowdsourced one where people could direct the mic to the bird. By doing that, we could track e.g. population of some particular species over large periods.\n\nAnd on labeling, the way it is done for LBs requires a lot of effort and not easily scalable, while there is a lot of open data with weak labels, and it scales well.", "2830369": "Thank you, well said! That is exactly the reason for the competition. We need outside help to tackle this problem because, as you can see, it's a difficult task.", "2841331": "I believe that test data just has more noise and worse data this would imply that the solution for this competition would be very augment-driven." }, "source": "meta" }