Thank you very much for the great explanation, but I have a question please, is there an inverse transform for Wavelet scattering? since I need to make some processing in the features extracted by wavelet scattering then go back to same domain.
Thanks for your feedback. This topic has been discussed here dsp.stackexchange.com/questions/78514/wavelet-scattering-properties-implementation/78515#78515. An example is also available here where the authors train a convolutional network to invert the scattering transform see www.kymat.io/gallery_2d/regularized_inverse_scattering_MNIST_torch.html Read this paper if you trying to implement GANs: arxiv.org/pdf/1805.06621
Really very helpful and knowledge based video. Please make more videos based on EEG signal feature extraction with python implementations. Are you taking any course on any other platform?? I am really interested.
There are some errors and also mild misconceptions in this presentation, but also a lot of good in it. I definitely do not regret having spend the hour on it, thank you!
Thank you for your feedback, much appreciated. When you have some time, it will be great if you can share with me what went wrong and the misconceptions through an anonymous email through my website www.rami-khushaba.com/contacts.
Dr. Rami, that was really a crystal clear and crisp explanation of the topic. Sir could you please help me with the concept of multiscale pca. I have referred a lot of lectures and videos but I'm not getting that point clear. I m sure you can help me with the concept. I'm able to understand your videos very well. Thank you.
9:20 I woul add that with SFFT you lose low level frequencies for capturing of which you need the longer sample time. That nicely lays out the motivation for wavelets, which addresses thei conflict between time resolution and detection of low frequencies
This is a great resource. A few months back I had to go through several videos being from an unrelated field. These videos have everything in one place. Please let me know if I can reach out to you in some way to discuss further.
Hi, @AlaphBeth ty for your video! Could you clarify something? When decomposing the signal with wavelets, the decimation process won't make the left most portion of the spectrum (lowest frequencies) have less duration? If so, is it not contradictory? Should not the left most have the greatest duration due to larger wavelength? TY (:
Thanks for your feedback. With every decomposition step you have a filtering process and a down sampling step. The down sampling step will result in less samples being kept as you keep decomposing. This downsampling applies to left and right sides of the generated nodes. As a result, the number of samples left after each decomposition step reduces on both sides. About your question on wavelength, not sure why you assume the left most to have the greatest duration if you keep downsampling! Check the book titled Ripples in mathematics, it explains it all in an excellent way link.springer.com/book/10.1007/978-3-642-56702-5
Thank for the feedback The short answer for your question is yes, cwt or wst can be both utilised to convert a 1D signal into an image that researchers usually feed then to deep neural networks for classification (or regression) problems. You can also directly use the WST with the 1D signals for classification/regression. In one example I used the radar signals with WST to infer the different materials, see here th-cam.com/video/60eOV4tZT1o/w-d-xo.html
thank you for wonderful lecture !, I have one quick question,In Multi-channel EEG problem, why there are 6 features/channel. my understanding is that each outputs from the filter are also series of coefficients. so if we use coefficient as a feature, there will result in much more features per channel
Thanks for your feedback, much appreciated. You are right in that each output from the filter is also a series of coefficients, but In the example I extract the “energy” of each set of coefficients rather than the coefficients themselves. You can use either of the two approaches. Remember that the original version of the DWT is impacted by shifts (unless you are using a new version), so if you take the energy then you reduce the impact of the shifts while compacting the energy/variance. If you take the energy then you also won’t need PCA to reduce dimensionally (depending on how many channels/nodes you have).
Can this technique be used for feature detection in an unsupervised manner(no training dataset, just one single data sample)? One of the wonderful things that CWT offered was pointwise discontinuity analysis (Wavelet Transform Modulus Maxima) which helped to detect interesting discontinuities from noise (based on chains built by WTMM) without any training. Is there a possibility that this technique can offer something similar to WTMM ?
Hi, the scattering transform is itself using the modulus of the CWT and convolving that with low pass filters or wavelets of some order. If you want to detect discontinuity with wavelet scattering then you can look at the scalogram coefficients associated with the scattering transform (Matlab has examples), or the scattering coefficients themselves (I haven’t tested that myself). On the other hand, your question about extracting features when you have small sample size is already the focus of this article towardsdatascience.com/a-convnet-that-works-on-like-20-samples-scatter-wavelets-b2e858f8a385
The number of scattering coefficients depends on the values of the SignalLength, InvarianceScale, and OversamplingFactor properties of the scattering framework SF. Specifically, len = 1+fix((sl-1)./2^(cr-osfac)); sl = SignalLength cr = criticalResolution osfac = OversamplingFactor
The are so many ways to do that, it basically depends on what sort of ML pipeline are you developing. So if you plan to use CNN for example then you can feed the time frequency energy map as an image to CNN and train base on the images from the different observations you have. Alternatively, you might want to zoom on certain frequency range of interest in your problem and either feed that to CNN or extract further features to feed to a traditional classifier. Another very simply way, take the time frequency energy map generated by CWT and flatten (reshape) to a vector. This way you collect all the generated vectors from the different observations, run something like PCA or t-SNE or UMAP or any other dimensionally reduction and pass that to a traditional classifier. As I said, so many different methods, and you can even find more methods in some publications.
For those wanting to skip revision, Part 2 starts at 28:30
th-cam.com/play/PLOUHOQsIOpRF43dwJzVEeoFdEOlf7-fWW.html
SUBSCRIBE THIS CHANNEL FOR COMPLETE WAVELET TRANSFORM VIDEOS
Jesus😂I also endure the first half to find it out
Thank you very much amazing description of the wavelet scattering transform. please keep going.
Sir its soo helpful......you explain topics seem to be very dried in such an awesome manner....stay blessed always
Top drawer explanation, really appreciate it
Extremely excellent content, thank you for the detailed explanation!
Thanks for the excellent lecture!
Thanks, appreciate your feedback.
Thank you so much for this amazing explanation.
hello. it's a so good explanation. thanks👏
so well explained
Thank you very much for the great explanation, but I have a question please, is there an inverse transform for Wavelet scattering? since I need to make some processing in the features extracted by wavelet scattering then go back to same domain.
Thanks for your feedback. This topic has been discussed here dsp.stackexchange.com/questions/78514/wavelet-scattering-properties-implementation/78515#78515. An example is also available here where the authors train a convolutional network to invert the scattering transform see www.kymat.io/gallery_2d/regularized_inverse_scattering_MNIST_torch.html
Read this paper if you trying to implement GANs: arxiv.org/pdf/1805.06621
@@AlaphBeth Thank you very much.
Really very helpful and knowledge based video. Please make more videos based on EEG signal feature extraction with python implementations. Are you taking any course on any other platform?? I am really interested.
Many thanks dr.Rami
There are some errors and also mild misconceptions in this presentation, but also a lot of good in it. I definitely do not regret having spend the hour on it, thank you!
Thank you for your feedback, much appreciated. When you have some time, it will be great if you can share with me what went wrong and the misconceptions through an anonymous email through my website www.rami-khushaba.com/contacts.
Dr. Rami, that was really a crystal clear and crisp explanation of the topic. Sir could you please help me with the concept of multiscale pca. I have referred a lot of lectures and videos but I'm not getting that point clear. I m sure you can help me with the concept. I'm able to understand your videos very well. Thank you.
Thanks a lot for your feedback.
I will see if I can prepare another video about multiscale PCA soon and share it on TH-cam.
@@AlaphBeth thank you very much 🙏
9:20 I woul add that with SFFT you lose low level frequencies for capturing of which you need the longer sample time. That nicely lays out the motivation for wavelets, which addresses thei conflict between time resolution and detection of low frequencies
This is a great resource. A few months back I had to go through several videos being from an unrelated field. These videos have everything in one place. Please let me know if I can reach out to you in some way to discuss further.
Thanks for your feedback, please connect with me on LinkedIn if you have any questions or want to discuss anything.
Hi, @AlaphBeth ty for your video! Could you clarify something? When decomposing the signal with wavelets, the decimation process won't make the left most portion of the spectrum (lowest frequencies) have less duration? If so, is it not contradictory? Should not the left most have the greatest duration due to larger wavelength? TY (:
Thanks for your feedback. With every decomposition step you have a filtering process and a down sampling step. The down sampling step will result in less samples being kept as you keep decomposing. This downsampling applies to left and right sides of the generated nodes. As a result, the number of samples left after each decomposition step reduces on both sides. About your question on wavelength, not sure why you assume the left most to have the greatest duration if you keep downsampling! Check the book titled Ripples in mathematics, it explains it all in an excellent way link.springer.com/book/10.1007/978-3-642-56702-5
Well explained, I really wonder if you can make a video on using wavelet transform on (spectroscopy data), transferring 1D signal into an image
Thank for the feedback
The short answer for your question is yes, cwt or wst can be both utilised to convert a 1D signal into an image that researchers usually feed then to deep neural networks for classification (or regression) problems.
You can also directly use the WST with the 1D signals for classification/regression. In one example I used the radar signals with WST to infer the different materials, see here th-cam.com/video/60eOV4tZT1o/w-d-xo.html
thank you for wonderful lecture !, I have one quick question,In Multi-channel EEG problem, why there are 6 features/channel. my understanding is that each outputs from the filter are also series of coefficients. so if we use coefficient as a feature, there will result in much more features per channel
Thanks for your feedback, much appreciated. You are right in that each output from the filter is also a series of coefficients, but In the example I extract the “energy” of each set of coefficients rather than the coefficients themselves. You can use either of the two approaches. Remember that the original version of the DWT is impacted by shifts (unless you are using a new version), so if you take the energy then you reduce the impact of the shifts while compacting the energy/variance. If you take the energy then you also won’t need PCA to reduce dimensionally (depending on how many channels/nodes you have).
@@AlaphBeth thank you for your explanation Dr. Rami, I really appreciate it.
Can this technique be used for feature detection in an unsupervised manner(no training dataset, just one single data sample)? One of the wonderful things that CWT offered was pointwise discontinuity analysis (Wavelet Transform Modulus Maxima) which helped to detect interesting discontinuities from noise (based on chains built by WTMM) without any training. Is there a possibility that this technique can offer something similar to WTMM ?
Hi, the scattering transform is itself using the modulus of the CWT and convolving that with low pass filters or wavelets of some order. If you want to detect discontinuity with wavelet scattering then you can look at the scalogram coefficients associated with the scattering transform (Matlab has examples), or the scattering coefficients themselves (I haven’t tested that myself). On the other hand, your question about extracting features when you have small sample size is already the focus of this article towardsdatascience.com/a-convnet-that-works-on-like-20-samples-scatter-wavelets-b2e858f8a385
hello very good video but where can I find more information on the topic of wavelet scattering?
Thanks for the feedback, look for the wavelet scattering papers by Mallat and his group, mostly available free online.
Thank u Dr. For this explanation.. I have one doubt.. Could scattering be useful jn stegoanalysis apps?
Please make 3rd part of scattering transform.
Dr. Rami, I have a question that I would love to discuss/
Hi Saleh,
Thanks for your feedback, please reach out on LinkedIn if you have any questions.
@@AlaphBeth I tried but I could not send you a private message on linkedin
can you suggest a book to increase my knowledge more about detail?
Check this book link.springer.com/book/10.1007/978-3-642-56702-5
Check this book link.springer.com/book/10.1007/978-3-642-56702-5
In 1D example, why does the feature dimension (2nd) become 13 ?
The number of scattering coefficients depends on the values of the SignalLength, InvarianceScale, and OversamplingFactor properties of the scattering framework SF. Specifically,
len = 1+fix((sl-1)./2^(cr-osfac));
sl = SignalLength
cr = criticalResolution
osfac = OversamplingFactor
how i separate the 2 sources of the signal in one signal ? it's bout the strain ?
th-cam.com/play/PLOUHOQsIOpRF43dwJzVEeoFdEOlf7-fWW.html
SUBSCRIBE THIS CHANNEL FOR COMPLETE WAVELET TRANSFORM VIDEOS
how to extraction feature in continuous wavelet transform
The are so many ways to do that, it basically depends on what sort of ML pipeline are you developing. So if you plan to use CNN for example then you can feed the time frequency energy map as an image to CNN and train base on the images from the different observations you have. Alternatively, you might want to zoom on certain frequency range of interest in your problem and either feed that to CNN or extract further features to feed to a traditional classifier. Another very simply way, take the time frequency energy map generated by CWT and flatten (reshape) to a vector. This way you collect all the generated vectors from the different observations, run something like PCA or t-SNE or UMAP or any other dimensionally reduction and pass that to a traditional classifier. As I said, so many different methods, and you can even find more methods in some publications.
Intriguing…
Thank you so much sir .. please how i can get the ppt file
You are welcome, I will add all the ppt files to my GitHub repo github.com/RamiKhushaba
Thanks can I have the slide?