\n
![\"https://i.ibb.co/C66S7Cf/Selection-800.png\"](\"https://i.ibb.co/C66S7Cf/Selection-800.png\")
\n
![\"https://i.ibb.co/JrgF4Gw/Selection-801.png\"](\"https://i.ibb.co/JrgF4Gw/Selection-801.png\")
The excess of false rates can be decreased by combining the output of the NNs for pairs and single interferometers.
\n
\n
https://www.uv.es/igwm2021/slides/Alexis_Menendez-Vazquez.pdf
\nhttps://www.uv.es/igwm2021/slides/Elena_Cuoco.pdf
more here at:
\n11th Iberian Gravitational Waves Meeting
https://www.uv.es/igwm2021/program.html
", "messages": [ { "id": "1504285", "postDate": "09/06/2021 08:37:55", "content": "The excess of false rates can be decreased by combining the output of the NNs for pairs and single interferometers.
\n
\n
https://www.uv.es/igwm2021/slides/Alexis_Menendez-Vazquez.pdf
\nhttps://www.uv.es/igwm2021/slides/Elena_Cuoco.pdf
more here at:
\n11th Iberian Gravitational Waves Meeting
https://www.uv.es/igwm2021/program.html
", "rawMarkdown": "The excess of false rates can be decreased by combining the output of the NNs for pairs and single interferometers.\n\n\n\n\n\n\nhttps://www.uv.es/igwm2021/slides/Alexis_Menendez-Vazquez.pdf\nhttps://www.uv.es/igwm2021/slides/Elena_Cuoco.pdf\n\nmore here at:\n11th Iberian Gravitational Waves Meeting\n\nhttps://www.uv.es/igwm2021/program.html", "votes": null }, { "id": "1504296", "postDate": "09/06/2021 08:50:54", "content": "https://www.uv.es/igwm2021/slides/Osvaldo_Freitas.pdf
\n![\"https://i.ibb.co/ZxVHL6n/Selection-803.png\"](\"https://i.ibb.co/ZxVHL6n/Selection-803.png\")
\n![\"https://i.ibb.co/9cc3VnR/Selection-802.png\"](\"https://i.ibb.co/9cc3VnR/Selection-802.png\")
Where they say false-rate here, based on the distributions, I take that to mean false-positive rate (blue creeping into orange zone)? If so, I don't think we realllllly have that issue here. It's the false-negative rate that is hurting our auc in this competition's case, with very low snr bbh mergers being marked as background.
", "rawMarkdown": "Where they say false-rate here, based on the distributions, I take that to mean false-positive rate (blue creeping into orange zone)? If so, I don't think we realllllly have that issue here. It's the false-negative rate that is hurting our auc in this competition's case, with very low snr bbh mergers being marked as background.", "votes": null }, { "id": "1504476", "postDate": "09/06/2021 12:12:36", "content": "\n\nA network trained on small amplitude signals is able to detect larger amplitude signals
\n
One of the experiments I had ran was to get rid of the easy samples a model of mine was classifying at > 0.999. I reran training on all negative samples and just the remaining positive labels, which ended up being 2:1 ratio. The AuC was around 0.742.
\nMy goal in doing this was I believed I could use my main model as a stage-1, then run a fine-tuner model. No cigar. The AuC of the original model on the 'confusing' data used to train the fine tuner model was around ~0.768 (again, this is after the removal of about 50% samples that the base model classifies correctly with high confidence). The fine-tuner model trained from scratch on the resulting data couldn't surpass ~0.742 AuC. And I had iterated on the training hyperparameters as well as filtering.
\nTo me that feels like the inclusion of 'more data' even if it was easy to classify, helped the model learn more about what signal 'look like', which in turn bolstered its performance on the low snr samples.
", "rawMarkdown": "> A network trained on small amplitude signals is able to detect larger amplitude signals\n\nOne of the experiments I had ran was to get rid of the easy samples a model of mine was classifying at > 0.999. I reran training on all negative samples and just the remaining positive labels, which ended up being 2:1 ratio. The AuC was around 0.742.\n\nMy goal in doing this was I believed I could use my main model as a stage-1, then run a fine-tuner model. No cigar. The AuC of the original model on the 'confusing' data used to train the fine tuner model was around ~0.768 (again, this is after the removal of about 50% samples that the base model classifies correctly with high confidence). The fine-tuner model trained from scratch on the resulting data couldn't surpass ~0.742 AuC. And I had iterated on the training hyperparameters as well as filtering.\n\nTo me that feels like the inclusion of 'more data' even if it was easy to classify, helped the model learn more about what signal 'look like', which in turn bolstered its performance on the low snr samples.", "votes": null }, { "id": "1504552", "postDate": "09/06/2021 13:08:17", "content": "I am not able to view .png files posted by @hengck23. Is it just me ?
", "rawMarkdown": "I am not able to view .png files posted by @hengck23. Is it just me ?", "votes": null }, { "id": "1505916", "postDate": "09/07/2021 16:14:33", "content": "Nice work
", "rawMarkdown": "Nice work", "votes": null }, { "id": "1506309", "postDate": "09/08/2021 05:40:41", "content": "I cannot open the last two links\b. But your work seems amazing!
", "rawMarkdown": "I cannot open the last two links\b. But your work seems amazing!", "votes": null }, { "id": "1512956", "postDate": "09/14/2021 17:49:28", "content": "Thanks. I'm learning a lot in this competition
", "rawMarkdown": "Thanks. I'm learning a lot in this competition", "votes": null }, { "id": "1559714", "postDate": "10/27/2021 07:08:46", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris
https://www.uv.es/igwm2021/slides/Osvaldo_Freitas.pdf
\n
\n
\n\nA network trained on small amplitude signals is able to detect larger amplitude signals
\n
One of the experiments I had ran was to get rid of the easy samples a model of mine was classifying at > 0.999. I reran training on all negative samples and just the remaining positive labels, which ended up being 2:1 ratio. The AuC was around 0.742.
\nMy goal in doing this was I believed I could use my main model as a stage-1, then run a fine-tuner model. No cigar. The AuC of the original model on the 'confusing' data used to train the fine tuner model was around ~0.768 (again, this is after the removal of about 50% samples that the base model classifies correctly with high confidence). The fine-tuner model trained from scratch on the resulting data couldn't surpass ~0.742 AuC. And I had iterated on the training hyperparameters as well as filtering.
\nTo me that feels like the inclusion of 'more data' even if it was easy to classify, helped the model learn more about what signal 'look like', which in turn bolstered its performance on the low snr samples.
", "votes": null, "replies": [] } ] }, { "id": 1504465, "author_name": "authman", "author_url": "", "post_date": "09/06/2021 12:04:48", "content": "Where they say false-rate here, based on the distributions, I take that to mean false-positive rate (blue creeping into orange zone)? If so, I don't think we realllllly have that issue here. It's the false-negative rate that is hurting our auc in this competition's case, with very low snr bbh mergers being marked as background.
", "votes": null, "replies": [] }, { "id": 1504552, "author_name": "rashmibanthia", "author_url": "", "post_date": "09/06/2021 13:08:17", "content": "I am not able to view .png files posted by @hengck23. Is it just me ?
", "votes": null, "replies": [] }, { "id": 1505916, "author_name": "annjclark", "author_url": "", "post_date": "09/07/2021 16:14:33", "content": "Nice work
", "votes": null, "replies": [] }, { "id": 1506309, "author_name": "jiachengzhang6", "author_url": "", "post_date": "09/08/2021 05:40:41", "content": "I cannot open the last two links\b. But your work seems amazing!
", "votes": null, "replies": [] }, { "id": 1512956, "author_name": "robsonsan", "author_url": "", "post_date": "09/14/2021 17:49:28", "content": "Thanks. I'm learning a lot in this competition
", "votes": null, "replies": [] }, { "id": 1559714, "author_name": "zerafachris", "author_url": "", "post_date": "10/27/2021 07:08:46", "content": "Hey All,
\nThank you all for taking part in our competition. The participation has been overwhelmingly positive. We are currently conducting a survey to gauge the demographic and outreach achieved. Kindly spare 2min and fill in this survey https://forms.gle/QP9L16niPexozyhu5.
\nThank you all,
\nRegards,
\nChris