I was taking a look at some of what I think to be SED architectures, but I was wondering if anyone could link the birdclef specific variants that have been used in the past. From what I've gathered so far it seems like it's a convolutional feature extraction passed to an RNN module. Any resources would be appreciated, even those not related to birdclef.
", "messages": [ { "id": "3195370", "postDate": "05/07/2025 00:43:09", "content": "I was taking a look at some of what I think to be SED architectures, but I was wondering if anyone could link the birdclef specific variants that have been used in the past. From what I've gathered so far it seems like it's a convolutional feature extraction passed to an RNN module. Any resources would be appreciated, even those not related to birdclef.
", "rawMarkdown": "I was taking a look at some of what I think to be SED architectures, but I was wondering if anyone could link the birdclef specific variants that have been used in the past. From what I've gathered so far it seems like it's a convolutional feature extraction passed to an RNN module. Any resources would be appreciated, even those not related to birdclef.", "votes": null }, { "id": "3195377", "postDate": "05/07/2025 00:53:43", "content": "It applies attention over time-step features to highlight most important time-steps.
\nIt's weighted sum of class predictions over time, and weight is determined by the attention.
Thank you!
", "rawMarkdown": "Thank you!", "votes": null }, { "id": "3197264", "postDate": "05/08/2025 01:34:33", "content": "From what I know, this is the first kernel that introduce SED from the earlier bird competition. It's very comprehensive
\nhttps://www.kaggle.com/code/hidehisaarai1213/introduction-to-sound-event-detection
Hi, Did you use the intermediate layer features of the backbone? I'm currently a bit confused about whether a sequence length of 8×8 contains enough temporal/spatial information, even if it has a higher channel dimension.
", "rawMarkdown": "Hi, Did you use the intermediate layer features of the backbone? I'm currently a bit confused about whether a sequence length of 8×8 contains enough temporal/spatial information, even if it has a higher channel dimension.", "votes": null }, { "id": "3199648", "postDate": "05/11/2025 09:25:03", "content": "As I remember, it is an original paper - https://arxiv.org/abs/1710.02998
", "rawMarkdown": "As I remember, it is an original paper - https://arxiv.org/abs/1710.02998", "votes": null }, { "id": "3200064", "postDate": "05/12/2025 01:52:08", "content": "Sweet. Thank you!
", "rawMarkdown": "Sweet. Thank you!", "votes": null }, { "id": "3200065", "postDate": "05/12/2025 01:54:04", "content": "Thank you! I thought I remembered an RNN being involved. I think this is what I was thinking about.
", "rawMarkdown": "Thank you! I thought I remembered an RNN being involved. I think this is what I was thinking about.", "votes": null } ], "comments": [ { "id": 3195377, "author_name": "salmanahmedtamu", "author_url": "", "post_date": "05/07/2025 00:53:43", "content": "It applies attention over time-step features to highlight most important time-steps.
\nIt's weighted sum of class predictions over time, and weight is determined by the attention.
Thank you!
", "votes": null, "replies": [] }, { "id": 3197751, "author_name": "ryenhails", "author_url": "", "post_date": "05/08/2025 14:22:31", "content": "Hi, Did you use the intermediate layer features of the backbone? I'm currently a bit confused about whether a sequence length of 8×8 contains enough temporal/spatial information, even if it has a higher channel dimension.
", "votes": null, "replies": [] } ] }, { "id": 3197264, "author_name": "nyleve", "author_url": "", "post_date": "05/08/2025 01:34:33", "content": "From what I know, this is the first kernel that introduce SED from the earlier bird competition. It's very comprehensive
\nhttps://www.kaggle.com/code/hidehisaarai1213/introduction-to-sound-event-detection
Sweet. Thank you!
", "votes": null, "replies": [] } ] }, { "id": 3199648, "author_name": "vladimirsydor", "author_url": "", "post_date": "05/11/2025 09:25:03", "content": "As I remember, it is an original paper - https://arxiv.org/abs/1710.02998
", "votes": null, "replies": [ { "id": 3200065, "author_name": "willrice", "author_url": "", "post_date": "05/12/2025 01:54:04", "content": "Thank you! I thought I remembered an RNN being involved. I think this is what I was thinking about.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3195370": "I was taking a look at some of what I think to be SED architectures, but I was wondering if anyone could link the birdclef specific variants that have been used in the past. From what I've gathered so far it seems like it's a convolutional feature extraction passed to an RNN module. Any resources would be appreciated, even those not related to birdclef.", "3195377": "It applies attention over time-step features to highlight most important time-steps.\nIt's weighted sum of class predictions over time, and weight is determined by the attention.", "3197205": "Thank you!", "3197264": "From what I know, this is the first kernel that introduce SED from the earlier bird competition. It's very comprehensive\nhttps://www.kaggle.com/code/hidehisaarai1213/introduction-to-sound-event-detection", "3197751": "Hi, Did you use the intermediate layer features of the backbone? I'm currently a bit confused about whether a sequence length of 8×8 contains enough temporal/spatial information, even if it has a higher channel dimension.", "3199648": "As I remember, it is an original paper - https://arxiv.org/abs/1710.02998", "3200064": "Sweet. Thank you!", "3200065": "Thank you! I thought I remembered an RNN being involved. I think this is what I was thinking about." }, "source": "meta" }