\"the [azimuth/zenith] angle in radians of the neutrino. The target columns. The direction vector represented by zenith and azimuth points to where the neutrino came from.\"
\nhttps://www.kaggle.com/competitions/icecube-neutrinos-in-deep-ice/data
\nCitation: Deep learning reconstruction in ANTARES
\nAuthors: J. García-Méndez, N. Geißelbrecht, T. Eberl, M. Ardid, and S. Ardid, on behalf of the ANTARES collaboration
\nhttps://arxiv.org/pdf/2107.13654.pdf
\"The authors explored different architectures and network parameters. To diminish over-fitting during training, they considered early stopping with a patience of 10 epochs, for a maximum of 150 training epochs. In each epoch, learning batches of 64 input elements were considered. They used the Adam algorithm as the learning optimizer because it is able to dynamically regulate the learning rate, which was initialized at 0.001.\"
\n\"For the Zenith prediction, the authors scaled the tanh activation function for 𝜇, so that its value lay in [0, 𝜋] radians. No scaling was necessary for the prediction of the Azimuth since its estimation was derived from the Cartesian XY components of the unit vector.\"
\n\"Their DCN clearly outperforms the conventional BBfit reconstruction method of ANTARES. The DCN method results in a significantly larger proportion of low errors for the Zenith angle and a first estimation for the Azimuth angle, which was previously missing for single-line events.\"
\nhttps://arxiv.org/pdf/2107.13654.pdf
\n\"BBfit single-line reconstruction is, for low energy neutrinos, the most efficient algorithm in ANTARES. This energy range is highly relevant for studying Dark Matter and neutrino oscillations. However, BBfit only provides information about the Zenith angle of the neutrino direction.\"
\nhttps://lofnadi.github.io/project/ml-based-direction-estimate/
", "messages": [ { "id": 2108893, "postDate": "2023-01-20T23:29:28.150Z", "content": "\"the [azimuth/zenith] angle in radians of the neutrino. The target columns. The direction vector represented by zenith and azimuth points to where the neutrino came from.\"
\nhttps://www.kaggle.com/competitions/icecube-neutrinos-in-deep-ice/data
\nCitation: Deep learning reconstruction in ANTARES
\nAuthors: J. García-Méndez, N. Geißelbrecht, T. Eberl, M. Ardid, and S. Ardid, on behalf of the ANTARES collaboration
\nhttps://arxiv.org/pdf/2107.13654.pdf
\"The authors explored different architectures and network parameters. To diminish over-fitting during training, they considered early stopping with a patience of 10 epochs, for a maximum of 150 training epochs. In each epoch, learning batches of 64 input elements were considered. They used the Adam algorithm as the learning optimizer because it is able to dynamically regulate the learning rate, which was initialized at 0.001.\"
\n\"For the Zenith prediction, the authors scaled the tanh activation function for 𝜇, so that its value lay in [0, 𝜋] radians. No scaling was necessary for the prediction of the Azimuth since its estimation was derived from the Cartesian XY components of the unit vector.\"
\n\"Their DCN clearly outperforms the conventional BBfit reconstruction method of ANTARES. The DCN method results in a significantly larger proportion of low errors for the Zenith angle and a first estimation for the Azimuth angle, which was previously missing for single-line events.\"
\nhttps://arxiv.org/pdf/2107.13654.pdf
\n\"BBfit single-line reconstruction is, for low energy neutrinos, the most efficient algorithm in ANTARES. This energy range is highly relevant for studying Dark Matter and neutrino oscillations. However, BBfit only provides information about the Zenith angle of the neutrino direction.\"
\nhttps://lofnadi.github.io/project/ml-based-direction-estimate/
", "rawMarkdown": "\"the [azimuth/zenith] angle in radians of the neutrino. The target columns. The direction vector represented by zenith and azimuth points to where the neutrino came from.\"\n\nhttps://www.kaggle.com/competitions/icecube-neutrinos-in-deep-ice/data\n\nCitation: Deep learning reconstruction in ANTARES\nAuthors: J. García-Méndez, N. Geißelbrecht, T. Eberl, M. Ardid, and S. Ardid, on behalf of the ANTARES collaboration\nhttps://arxiv.org/pdf/2107.13654.pdf\n\n#Epochs, Optimizer Adam algorithm, Tanh activation function\n\n\"The authors explored different architectures and network parameters. To diminish over-fitting during training, they considered early stopping with a patience of 10 epochs, for a maximum of 150 training epochs. In each epoch, learning batches of 64 input elements were considered. They used the Adam algorithm as the learning optimizer because it is able to dynamically regulate the learning rate, which was initialized at 0.001.\"\n\n\"For the Zenith prediction, the authors scaled the tanh activation function for 𝜇, so that its value lay in [0, 𝜋] radians. No scaling was necessary for the prediction of the Azimuth since its estimation was derived from the Cartesian XY components of the unit vector.\"\n\n\"Their DCN clearly outperforms the conventional BBfit reconstruction method of ANTARES. The DCN method results in a significantly larger proportion of low errors for the Zenith angle and a first estimation for the Azimuth angle, which was previously missing for single-line events.\"\n\nhttps://arxiv.org/pdf/2107.13654.pdf\n\n#BBfit algorithm\n\n\"BBfit single-line reconstruction is, for low energy neutrinos, the most efficient algorithm in ANTARES. This energy range is highly relevant for studying Dark Matter and neutrino oscillations. However, BBfit only provides information about the Zenith angle of the neutrino direction.\"\n\nhttps://lofnadi.github.io/project/ml-based-direction-estimate/", "votes": 11 }, { "id": 2109750, "postDate": "2023-01-21T16:16:23.377Z", "content": "Thanks for sharing these resources @mpwolke! I think there could be potential with this type of approach.
", "rawMarkdown": "Thanks for sharing these resources @mpwolke! I think there could be potential with this type of approach.", "votes": 1, "replies": [ { "id": 2109863, "postDate": "2023-01-21T18:19:51.157Z", "content": "Thank you Ravi. I could barely open the parquet files : )
", "rawMarkdown": "Thank you Ravi. I could barely open the parquet files : )", "votes": 1, "replies": [ { "id": 2110062, "postDate": "2023-01-21T22:47:43.477Z", "content": "Yea, I just realized the dataset size was 117.18 GB. It will definitely be a challenge to train a neural net on a dataset of this size. I am very interested to see what approach will be most common and what approach will be most effective in this competition.
", "rawMarkdown": "Yea, I just realized the dataset size was 117.18 GB. It will definitely be a challenge to train a neural net on a dataset of this size. I am very interested to see what approach will be most common and what approach will be most effective in this competition.", "votes": 1 } ] } ] }, { "id": 2108994, "postDate": "2023-01-21T03:17:46.073Z", "content": "Thanks for sharing the article…..it will help in improving the score
", "rawMarkdown": "Thanks for sharing the article.....it will help in improving the score", "votes": 1, "replies": [ { "id": 2109333, "postDate": "2023-01-21T10:30:23.677Z", "content": "I hope it helps since with that amount of data I have to check how to reduce memory with parquets.
", "rawMarkdown": "I hope it helps since with that amount of data I have to check how to reduce memory with parquets." } ] }, { "id": 2111373, "postDate": "2023-01-22T20:55:28.710Z", "rawMarkdown": "", "votes": 1, "isDeleted": true, "replies": [ { "id": 2111386, "postDate": "2023-01-22T21:10:40.743Z", "content": "Thank you Yemen I'm glad that it could help anyone.
", "rawMarkdown": "Thank you Yemen I'm glad that it could help anyone." } ] } ], "comments": [ { "id": 2109750, "author_name": "Ravi Shah", "author_url": "", "post_date": "2023-01-21T16:16:23.377000", "content": "Thanks for sharing these resources @mpwolke! I think there could be potential with this type of approach.
", "votes": 1, "replies": [ { "id": 2109863, "author_name": "Marília Prata", "author_url": "", "post_date": "2023-01-21T18:19:51.157000", "content": "Thank you Ravi. I could barely open the parquet files : )
", "votes": 1, "replies": [ { "id": 2110062, "author_name": "Ravi Shah", "author_url": "", "post_date": "2023-01-21T22:47:43.477000", "content": "Yea, I just realized the dataset size was 117.18 GB. It will definitely be a challenge to train a neural net on a dataset of this size. I am very interested to see what approach will be most common and what approach will be most effective in this competition.
", "votes": 1, "replies": [] } ] } ] }, { "id": 2108994, "author_name": "Satya", "author_url": "", "post_date": "2023-01-21T03:17:46.073000", "content": "Thanks for sharing the article…..it will help in improving the score
", "votes": 1, "replies": [ { "id": 2109333, "author_name": "Marília Prata", "author_url": "", "post_date": "2023-01-21T10:30:23.677000", "content": "I hope it helps since with that amount of data I have to check how to reduce memory with parquets.
", "votes": 0, "replies": [] } ] }, { "id": 2111373, "author_name": "", "author_url": "", "post_date": "2023-01-22T20:55:28.710000", "content": "", "votes": 1, "replies": [ { "id": 2111386, "author_name": "Marília Prata", "author_url": "", "post_date": "2023-01-22T21:10:40.743000", "content": "Thank you Yemen I'm glad that it could help anyone.
", "votes": 0, "replies": [] } ] } ], "raw_markdown_by_id": { "2108893": "\"the [azimuth/zenith] angle in radians of the neutrino. The target columns. The direction vector represented by zenith and azimuth points to where the neutrino came from.\"\n\nhttps://www.kaggle.com/competitions/icecube-neutrinos-in-deep-ice/data\n\nCitation: Deep learning reconstruction in ANTARES\nAuthors: J. García-Méndez, N. Geißelbrecht, T. Eberl, M. Ardid, and S. Ardid, on behalf of the ANTARES collaboration\nhttps://arxiv.org/pdf/2107.13654.pdf\n\n#Epochs, Optimizer Adam algorithm, Tanh activation function\n\n\"The authors explored different architectures and network parameters. To diminish over-fitting during training, they considered early stopping with a patience of 10 epochs, for a maximum of 150 training epochs. In each epoch, learning batches of 64 input elements were considered. They used the Adam algorithm as the learning optimizer because it is able to dynamically regulate the learning rate, which was initialized at 0.001.\"\n\n\"For the Zenith prediction, the authors scaled the tanh activation function for 𝜇, so that its value lay in [0, 𝜋] radians. No scaling was necessary for the prediction of the Azimuth since its estimation was derived from the Cartesian XY components of the unit vector.\"\n\n\"Their DCN clearly outperforms the conventional BBfit reconstruction method of ANTARES. The DCN method results in a significantly larger proportion of low errors for the Zenith angle and a first estimation for the Azimuth angle, which was previously missing for single-line events.\"\n\nhttps://arxiv.org/pdf/2107.13654.pdf\n\n#BBfit algorithm\n\n\"BBfit single-line reconstruction is, for low energy neutrinos, the most efficient algorithm in ANTARES. This energy range is highly relevant for studying Dark Matter and neutrino oscillations. However, BBfit only provides information about the Zenith angle of the neutrino direction.\"\n\nhttps://lofnadi.github.io/project/ml-based-direction-estimate/", "2109750": "Thanks for sharing these resources @mpwolke! I think there could be potential with this type of approach.", "2108994": "Thanks for sharing the article.....it will help in improving the score", "2111373": "" } }