The final submission deadline is approaching, and with the remaining attempts, I’m sure everyone is exploring various ensemble strategies.
\nI’d love to hear your thoughts on architecture—are you planning to use a hybrid approach (CNN/SED/raw signal) or a single architecture?
\nAnd for the backbone, are you leaning toward a single backbone, two backbones, or an ensemble of multiple backbones?
\nFrom our experiments, SED has significantly outperformed CNN, while the different 5 backbones we tested showed slight variations but overall similar performance on the public leaderboard (same pipe, from 867-885).
\nSo for now, we’ve decided to go with a single architecture (SED) and ensemble as many different backbones as possible.
\nHow are you approaching this decision based on your own results?
", "messages": [ { "id": "3217154", "postDate": "06/04/2025 15:22:52", "content": "The final submission deadline is approaching, and with the remaining attempts, I’m sure everyone is exploring various ensemble strategies.
\nI’d love to hear your thoughts on architecture—are you planning to use a hybrid approach (CNN/SED/raw signal) or a single architecture?
\nAnd for the backbone, are you leaning toward a single backbone, two backbones, or an ensemble of multiple backbones?
\nFrom our experiments, SED has significantly outperformed CNN, while the different 5 backbones we tested showed slight variations but overall similar performance on the public leaderboard (same pipe, from 867-885).
\nSo for now, we’ve decided to go with a single architecture (SED) and ensemble as many different backbones as possible.
\nHow are you approaching this decision based on your own results?
", "rawMarkdown": "The final submission deadline is approaching, and with the remaining attempts, I’m sure everyone is exploring various ensemble strategies.\n\nI’d love to hear your thoughts on architecture—are you planning to use a hybrid approach (CNN/SED/raw signal) or a single architecture?\n\nAnd for the backbone, are you leaning toward a single backbone, two backbones, or an ensemble of multiple backbones?\n\nFrom our experiments, SED has significantly outperformed CNN, while the different 5 backbones we tested showed slight variations but overall similar performance on the public leaderboard (same pipe, from 867-885).\n\nSo for now, we’ve decided to go with a single architecture (SED) and ensemble as many different backbones as possible.\n\nHow are you approaching this decision based on your own results?", "votes": null }, { "id": "3217449", "postDate": "06/05/2025 03:06:01", "content": "In my experiments,ensemble different train_data(select different 5s segments)improve a lot.Ensemble different architecture also help.
\nSome tests:
\n1.different train_data 3CNN(0.833/0.823/0.812)→0.862
\n2.different architecture 2SED(0.873/0.876) + 3CNN(0.862)→0.897 ,however,replace 0.873 with 0.881 single model will decrease public LB.
\nNow I have some doubts about whether I am overfitting public LB,because when i replace one of CNN model will decrease a lot.
In my experiments, there's a big difference in prediction probability for the same audio using either CNN or SED architecture. Is it the same in your case?
", "rawMarkdown": "In my experiments, there's a big difference in prediction probability for the same audio using either CNN or SED architecture. Is it the same in your case?", "votes": null }, { "id": "3217560", "postDate": "06/05/2025 06:17:12", "content": "yep.But just simple weighted ensemble can improve public LB.
\nI try rank ensemble,not good as weighted ensemble.
Great diversity! Our results show that the ensemble score is heavily influenced by the highest-performing backbone. Currently, a seven-model ensemble has only marginally improved the score—by less than 0.001—compared to the best single model. It makes me question the generalization of the training pipeline😢
", "rawMarkdown": "Great diversity! Our results show that the ensemble score is heavily influenced by the highest-performing backbone. Currently, a seven-model ensemble has only marginally improved the score—by less than 0.001—compared to the best single model. It makes me question the generalization of the training pipeline😢", "votes": null } ], "comments": [ { "id": 3217449, "author_name": "minatoyukinaxlisa", "author_url": "", "post_date": "06/05/2025 03:06:01", "content": "In my experiments,ensemble different train_data(select different 5s segments)improve a lot.Ensemble different architecture also help.
\nSome tests:
\n1.different train_data 3CNN(0.833/0.823/0.812)→0.862
\n2.different architecture 2SED(0.873/0.876) + 3CNN(0.862)→0.897 ,however,replace 0.873 with 0.881 single model will decrease public LB.
\nNow I have some doubts about whether I am overfitting public LB,because when i replace one of CNN model will decrease a lot.
In my experiments, there's a big difference in prediction probability for the same audio using either CNN or SED architecture. Is it the same in your case?
", "votes": null, "replies": [ { "id": 3217560, "author_name": "minatoyukinaxlisa", "author_url": "", "post_date": "06/05/2025 06:17:12", "content": "yep.But just simple weighted ensemble can improve public LB.
\nI try rank ensemble,not good as weighted ensemble.
Great diversity! Our results show that the ensemble score is heavily influenced by the highest-performing backbone. Currently, a seven-model ensemble has only marginally improved the score—by less than 0.001—compared to the best single model. It makes me question the generalization of the training pipeline😢
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3217154": "The final submission deadline is approaching, and with the remaining attempts, I’m sure everyone is exploring various ensemble strategies.\n\nI’d love to hear your thoughts on architecture—are you planning to use a hybrid approach (CNN/SED/raw signal) or a single architecture?\n\nAnd for the backbone, are you leaning toward a single backbone, two backbones, or an ensemble of multiple backbones?\n\nFrom our experiments, SED has significantly outperformed CNN, while the different 5 backbones we tested showed slight variations but overall similar performance on the public leaderboard (same pipe, from 867-885).\n\nSo for now, we’ve decided to go with a single architecture (SED) and ensemble as many different backbones as possible.\n\nHow are you approaching this decision based on your own results?", "3217449": "In my experiments,ensemble different train_data(select different 5s segments)improve a lot.Ensemble different architecture also help.\nSome tests: \n1.**different train_data** 3CNN(0.833/0.823/0.812)→0.862\n2.**different architecture** 2SED(0.873/0.876) + 3CNN(0.862)→0.897 ,however,replace 0.873 with 0.881 single model will decrease public LB.\nNow I have some doubts about whether I am overfitting public LB,because when i replace one of CNN model will decrease a lot.", "3217553": "In my experiments, there's a big difference in prediction probability for the same audio using either CNN or SED architecture. Is it the same in your case?", "3217560": "yep.But just simple weighted ensemble can improve public LB.\nI try rank ensemble,not good as weighted ensemble.", "3217627": "Great diversity! Our results show that the ensemble score is heavily influenced by the highest-performing backbone. Currently, a seven-model ensemble has only marginally improved the score—by less than 0.001—compared to the best single model. It makes me question the generalization of the training pipeline😢" }, "source": "meta" }