![\"https://cds.cern.ch/record/2777883/files/ROC_curve_log_BBHdataset_5e-5.png\"](\"https://cds.cern.ch/record/2777883/files/ROC_curve_log_BBHdataset_5e-5.png\")
I don't have a lot experience in wave/signal processing based problems, in previous project I had tried converting the 1d signal to spectrogram and feed that to a cnn model, which seems like was what everyone else using (similarly in this comp).
\nThis time I wanted to play around with TCN (temporal convolution) that operates directly on the raw wave, the intuition is simple, basically i was hoping whatever transformation needed could be learned by the convolution kernels themselves, which then might yield better performance than using a particular transformation (e.g. fourier), but im having difficulty to get it work, the performance on my local validation set is always hovering around 0.5 AUC, even when I increased the network size (e.g. simple more stacked layers) the val performance is still 0.5, and on train set its underfitting. Just wondering if anyone has tried it, if so whats your performance like? if not, any thoughts on this
", "messages": [ { "id": "1378995", "postDate": "07/07/2021 03:00:43", "content": "I don't have a lot experience in wave/signal processing based problems, in previous project I had tried converting the 1d signal to spectrogram and feed that to a cnn model, which seems like was what everyone else using (similarly in this comp).
\nThis time I wanted to play around with TCN (temporal convolution) that operates directly on the raw wave, the intuition is simple, basically i was hoping whatever transformation needed could be learned by the convolution kernels themselves, which then might yield better performance than using a particular transformation (e.g. fourier), but im having difficulty to get it work, the performance on my local validation set is always hovering around 0.5 AUC, even when I increased the network size (e.g. simple more stacked layers) the val performance is still 0.5, and on train set its underfitting. Just wondering if anyone has tried it, if so whats your performance like? if not, any thoughts on this
", "rawMarkdown": "I don't have a lot experience in wave/signal processing based problems, in previous project I had tried converting the 1d signal to spectrogram and feed that to a cnn model, which seems like was what everyone else using (similarly in this comp). \n\nThis time I wanted to play around with TCN (temporal convolution) that operates directly on the raw wave, the intuition is simple, basically i was hoping whatever transformation needed could be learned by the convolution kernels themselves, which then might yield better performance than using a particular transformation (e.g. fourier), but im having difficulty to get it work, the performance on my local validation set is always hovering around 0.5 AUC, even when I increased the network size (e.g. simple more stacked layers) the val performance is still 0.5, and on train set its underfitting. Just wondering if anyone has tried it, if so whats your performance like? if not, any thoughts on this", "votes": null }, { "id": "1379161", "postDate": "07/07/2021 06:26:32", "content": "good idea!
", "rawMarkdown": "good idea!", "votes": null }, { "id": "1503871", "postDate": "09/05/2021 19:27:53", "content": "My guess is that spectrogram (2d image) works better because of the nature of the problem -- two objects rotating around each other in a shrinking orbit until they merge.
", "rawMarkdown": "My guess is that spectrogram (2d image) works better because of the nature of the problem -- two objects rotating around each other in a shrinking orbit until they merge.", "votes": null }, { "id": "1503885", "postDate": "09/05/2021 19:58:10", "content": "refer to :
\nhttps://kinwaicheuk.github.io/nnAudio/intro.html#trainable-kernals
the kernels trained are not smoothed if do not have enough training samples.
\nrather than learning the 1d conv filters weights, parameterised it and learned the parameter.
\ne.g. 1d conv kernel = wavelet(a,b)
back prop to find a,b
\nalso, it is the power of the frequency that matters. hence you need to whiten and remove as much noise as possible if you want to use 1d convolution.
\nthe key difference of 2dCNN and 1dCNN approaches is that 2dCNN looks at spectrogram which is essentially the power.
\nfor 1d CNN it is essentially the dot product of the 1d kernel and the input.
", "rawMarkdown": "refer to :\nhttps://kinwaicheuk.github.io/nnAudio/intro.html#trainable-kernals\n\nthe kernels trained are not smoothed if do not have enough training samples.\n\nrather than learning the 1d conv filters weights, parameterised it and learned the parameter.\ne.g. 1d conv kernel = wavelet(a,b)\n\nback prop to find a,b\n\n----\n\nalso, it is the power of the frequency that matters. hence you need to whiten and remove as much noise as possible if you want to use 1d convolution.\n\nthe key difference of 2dCNN and 1dCNN approaches is that 2dCNN looks at spectrogram which is essentially the power.\n\nfor 1d CNN it is essentially the dot product of the 1d kernel and the input.", "votes": null }, { "id": "1504001", "postDate": "09/06/2021 01:39:16", "content": "STFT is a convolution on raw waveform, therefore you have already used a TCN.
\nThe question is whether you can achieve the same accuracy with a convolution learned from scratch.
", "rawMarkdown": "STFT is a convolution on raw waveform, therefore you have already used a TCN. \n\nThe question is whether you can achieve the same accuracy with a convolution learned from scratch.", "votes": null }, { "id": "1507269", "postDate": "09/09/2021 03:38:21", "content": "Underrated comment imo.
", "rawMarkdown": "Underrated comment imo.", "votes": null }, { "id": "1507575", "postDate": "09/09/2021 10:36:32", "content": "lstm results:
\nhttps://cds.cern.ch/record/2777883/plots
results for BBH and BNS
\nSource-Agnostic Gravitational-Wave Detection with Recurrent Autoencoders - Moreno, Eric A. et al - arXiv:2107.12698
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
good idea!
", "votes": null, "replies": [] }, { "id": 1503871, "author_name": "tolgadincer", "author_url": "", "post_date": "09/05/2021 19:27:53", "content": "My guess is that spectrogram (2d image) works better because of the nature of the problem -- two objects rotating around each other in a shrinking orbit until they merge.
", "votes": null, "replies": [] }, { "id": 1503885, "author_name": "hengck23", "author_url": "", "post_date": "09/05/2021 19:58:10", "content": "refer to :
\nhttps://kinwaicheuk.github.io/nnAudio/intro.html#trainable-kernals
the kernels trained are not smoothed if do not have enough training samples.
\nrather than learning the 1d conv filters weights, parameterised it and learned the parameter.
\ne.g. 1d conv kernel = wavelet(a,b)
back prop to find a,b
\nalso, it is the power of the frequency that matters. hence you need to whiten and remove as much noise as possible if you want to use 1d convolution.
\nthe key difference of 2dCNN and 1dCNN approaches is that 2dCNN looks at spectrogram which is essentially the power.
\nfor 1d CNN it is essentially the dot product of the 1d kernel and the input.
", "votes": null, "replies": [ { "id": 1507269, "author_name": "authman", "author_url": "", "post_date": "09/09/2021 03:38:21", "content": "Underrated comment imo.
", "votes": null, "replies": [] } ] }, { "id": 1504001, "author_name": "cpmpml", "author_url": "", "post_date": "09/06/2021 01:39:16", "content": "STFT is a convolution on raw waveform, therefore you have already used a TCN.
\nThe question is whether you can achieve the same accuracy with a convolution learned from scratch.
", "votes": null, "replies": [] }, { "id": 1507575, "author_name": "hengck23", "author_url": "", "post_date": "09/09/2021 10:36:32", "content": "lstm results:
\nhttps://cds.cern.ch/record/2777883/plots
results for BBH and BNS
\nSource-Agnostic Gravitational-Wave Detection with Recurrent Autoencoders - Moreno, Eric A. et al - arXiv:2107.12698
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