The presence of \"noise\" in spectrograms appears crucial for accurate classifications. During the training of a variational auto-encoder, much of this noise is lost due to compression, as evidenced in a Kaggle notebook (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-training ). When examining the latent space visualization, no clear correlation emerges between position in the space and approximate classifications (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-analysis ).
\nExperiments involving augmentations indicate that adding noise often leads to a decrease in LB score, as noted in discussions (https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477841 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/469953 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/481436 ).
\nFrom these observations, it can be inferred that certain noise-like features contain valuable information. However, the VAE removes this noise, diminishing its ability to depict a useful classified latent space, while augmentations alter necessary noise, thus affecting expected classifications. It might be best to explore alternative models capable of effectively capturing finer details, such as experimenting with smaller kernel sizes or other methods.
", "messages": [ { "id": "2702544", "postDate": "03/17/2024 17:33:21", "content": "The presence of \"noise\" in spectrograms appears crucial for accurate classifications. During the training of a variational auto-encoder, much of this noise is lost due to compression, as evidenced in a Kaggle notebook (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-training ). When examining the latent space visualization, no clear correlation emerges between position in the space and approximate classifications (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-analysis ).
\nExperiments involving augmentations indicate that adding noise often leads to a decrease in LB score, as noted in discussions (https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477841 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/469953 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/481436 ).
\nFrom these observations, it can be inferred that certain noise-like features contain valuable information. However, the VAE removes this noise, diminishing its ability to depict a useful classified latent space, while augmentations alter necessary noise, thus affecting expected classifications. It might be best to explore alternative models capable of effectively capturing finer details, such as experimenting with smaller kernel sizes or other methods.
", "rawMarkdown": "#### Results from testing with VAE\n\nThe presence of \"noise\" in spectrograms appears crucial for accurate classifications. During the training of a variational auto-encoder, much of this noise is lost due to compression, as evidenced in a Kaggle notebook (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-training ). When examining the latent space visualization, no clear correlation emerges between position in the space and approximate classifications (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-analysis ).\n\n#### Augmentations\n\nExperiments involving augmentations indicate that adding noise often leads to a decrease in LB score, as noted in discussions (https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477841 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/469953 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/481436 ).\n\n#### Conclusions\n\nFrom these observations, it can be inferred that certain noise-like features contain valuable information. However, the VAE removes this noise, diminishing its ability to depict a useful classified latent space, while augmentations alter necessary noise, thus affecting expected classifications. It might be best to explore alternative models capable of effectively capturing finer details, such as experimenting with smaller kernel sizes or other methods.", "votes": null }, { "id": "2702562", "postDate": "03/17/2024 17:51:15", "content": "I can't see the notebooks you linked to. They are probably private; you need to make them public for us to see them.
", "rawMarkdown": "I can't see the notebooks you linked to. They are probably private; you need to make them public for us to see them.", "votes": null }, { "id": "2702566", "postDate": "03/17/2024 17:55:51", "content": "Just fixed it. It seems like the )'s were getting included in the links
", "rawMarkdown": "Just fixed it. It seems like the )'s were getting included in the links", "votes": null } ], "comments": [ { "id": 2702562, "author_name": "shlomoron", "author_url": "", "post_date": "03/17/2024 17:51:15", "content": "I can't see the notebooks you linked to. They are probably private; you need to make them public for us to see them.
", "votes": null, "replies": [ { "id": 2702566, "author_name": "ryankim17920", "author_url": "", "post_date": "03/17/2024 17:55:51", "content": "Just fixed it. It seems like the )'s were getting included in the links
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2702544": "#### Results from testing with VAE\n\nThe presence of \"noise\" in spectrograms appears crucial for accurate classifications. During the training of a variational auto-encoder, much of this noise is lost due to compression, as evidenced in a Kaggle notebook (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-training ). When examining the latent space visualization, no clear correlation emerges between position in the space and approximate classifications (https://www.kaggle.com/code/ryankim17920/variational-auto-encoder-analysis ).\n\n#### Augmentations\n\nExperiments involving augmentations indicate that adding noise often leads to a decrease in LB score, as noted in discussions (https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/477841 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/469953 , https://www.kaggle.com/competitions/hms-harmful-brain-activity-classification/discussion/481436 ).\n\n#### Conclusions\n\nFrom these observations, it can be inferred that certain noise-like features contain valuable information. However, the VAE removes this noise, diminishing its ability to depict a useful classified latent space, while augmentations alter necessary noise, thus affecting expected classifications. It might be best to explore alternative models capable of effectively capturing finer details, such as experimenting with smaller kernel sizes or other methods.", "2702562": "I can't see the notebooks you linked to. They are probably private; you need to make them public for us to see them.", "2702566": "Just fixed it. It seems like the )'s were getting included in the links" }, "source": "meta" }