Hi Steve. Slightly different comment than Roger. Many (all?) of your sparsity/sensor/compression videos in the Linear Algebra playlist are "Unlisted". They do not appear in your RUclips "Home" tab, nor if I go thru everything in the "Videos" tab. Nor if I do a search on your RUclips page will they appear. So, the only way I even came across them (including this one) was because they are in the Linear Algebra playlist. I just thought I'd point it out in case this was an error. The view counts on them are very low because of this and it's great content so people are missing out! (specifically, it's video #s 70-77 and 86-88 in the Linear Algebra playlist) Thanks for all the amazing content, I am learning so much. Just bought your book!
Hi Prof. Steve, Thank you for sharing such an amazing content. It is always fascinating to see the amount of energy you put in these videos, and I really enjoy learning from it. I wanted to ask a question regarding the minimization condition (theta). I understand the concept of generating data to create a library of flow field (psi r ), I also get the part of QR factorisation to get the sensor location C , then measuring the values at the C location to create y. But then you stated around 14:00 time stamp that we also know the value of theta (minimization condition). My question is how do we know this value so we can find a?
The principle behind that compressing of information is very simple and brilliant. I understand it as an adaption of principal component analysis in statistics. You often uses images as examples. But what if we have a dynamical system like a production line, which depends on the evolving time. How can I use the method when xt with t=0,…,n is given and not only x?
I think it is has been beneficial to the field that targeted sensing came after generalised compressed sensing. It feels like the field has first discovered a general concept and went narrower from there, as opposed to having to generalise a narrow technique. Thoughts?
This seems like it could also be used in remote sensing, where over time pixels are lost from the satellite, and historically they were just pasted over by interpolating the pixels still working on either side. Training the data on previous images from the region being sensed would provide a better way to restore the lost components, or provide a indication of how serious the lost elements are (if they appear as pivot rows).
Hi Steve I watched your videos for some time and loved them. I wanted to ask you something, what you recommend a second degree in data science or electrical engineering?.
Good question. In this case, you can either just use the first "r" QR pivot points, or you can transpose the matrix, depending on the interpretation of the rows and columns.
Hi Professor , I'm inspired by this idea and would like to try extending it beyond 2d images. But first of all, I need to get the EigenFace example running. I went to the book link for the Eigenface dataset, but the data format are different and there is no "allFaces.mat" directly available, which is used by the code example in the book. I searched in the book, but there is no instructions on how to make this "".mat" file. I could probably try to create this "allFaces.mat" with my own understanding of how this file is structured, but it would be much more convenient if you can provide it directly for the readers. Thanks!
Since U shows measurements and principal components, V shows principal components and original features, it is not clear why/how QR of U gives the shortlist of sensor placement. Any clarification would be appreciated.
Hello, thank you for the amazing video ! The methodology seems very powerful. I just have a question about the Phi*Phi^t, used when the number of sensors is bigger than modes. Why did you use Phi*Phi^t and what are the change on the R matrix ? Because we only found that the diagonaux elements relating to the number of modes considered are bigger than those after. Thank you in advance for you return. Maurine
Hi! Can you show a simulation(video) where is trying to classify diffrent images of different classes. I have been trying this with GNU Octave and this method did not work for me.
Does that not mean that in graphics engine, they could like 5x their performance by rendering random points only, like 20% and then upscaling that with compressed sensing?
Great question! It usually takes me about 5-10 minutes to make the intro 5 second clip and align the audio, touch up picture, etc., and then it processes on my laptop for 30min.
Isn't this a really hard combinatory problem? I mean you essentially want a vector mask C that is optical for chosing k points of a big vector of size n. And you use those k number of measurements, multiply them by k eigenvalues and hope the sum is as close to n as possible. Wouldn't this just mean you have n chose k options for said mask? And your approach is to take the most weighted point from every single eigenvalue and that results in your mask? I have to see some of the older lectures or sleep a day first to understand the mathematics of those QR matrix optimization.
Hi Steve. Slightly different comment than Roger. Many (all?) of your sparsity/sensor/compression videos in the Linear Algebra playlist are "Unlisted". They do not appear in your RUclips "Home" tab, nor if I go thru everything in the "Videos" tab. Nor if I do a search on your RUclips page will they appear. So, the only way I even came across them (including this one) was because they are in the Linear Algebra playlist. I just thought I'd point it out in case this was an error. The view counts on them are very low because of this and it's great content so people are missing out! (specifically, it's video #s 70-77 and 86-88 in the Linear Algebra playlist)
Thanks for all the amazing content, I am learning so much. Just bought your book!
Brilliant video, really interesting and well explained
Very clear as crystal!
High-quality presenation...
Excellent recap summary of the vid at 13:16. I think more technical explanations should include a helicopter view like that at the end.
Amazing lecture and super interesting topic! Always a pleasure watching these videos, Steve.
Thanks so much!
Hi Prof. Steve, Thank you for sharing such an amazing content. It is always fascinating to see the amount of energy you put in these videos, and I really enjoy learning from it. I wanted to ask a question regarding the minimization condition (theta). I understand the concept of generating data to create a library of flow field (psi r ), I also get the part of QR factorisation to get the sensor location C , then measuring the values at the C location to create y. But then you stated around 14:00 time stamp that we also know the value of theta (minimization condition). My question is how do we know this value so we can find a?
The principle behind that compressing of information is very simple and brilliant. I understand it as an adaption of principal component analysis in statistics. You often uses images as examples. But what if we have a dynamical system like a production line, which depends on the evolving time. How can I use the method when xt with t=0,…,n is given and not only x?
just beautiful!
I think it is has been beneficial to the field that targeted sensing came after generalised compressed sensing. It feels like the field has first discovered a general concept and went narrower from there, as opposed to having to generalise a narrow technique. Thoughts?
Thank you! These amazing videos helped me a lot!
Glad they have been helpful!
This seems like it could also be used in remote sensing, where over time pixels are lost from the satellite, and historically they were just pasted over by interpolating the pixels still working on either side. Training the data on previous images from the region being sensed would provide a better way to restore the lost components, or provide a indication of how serious the lost elements are (if they appear as pivot rows).
Hi Steve I watched your videos for some time and loved them.
I wanted to ask you something, what you recommend a second degree in data science or electrical engineering?.
Thank you very much for your great lecture. Could you please explain a bit how Si_r matrix is calculated using the training data?
Phenomenal work
Thank you for the presentation.
What if our data matrix short and fat and instead of tall and skinny. Does QR apply in that case?
Good question. In this case, you can either just use the first "r" QR pivot points, or you can transpose the matrix, depending on the interpretation of the rows and columns.
@@Eigensteve thanks
"If you do know what type of *sister* you're measuring..." lol made me giggle
Hi Professor , I'm inspired by this idea and would like to try extending it beyond 2d images. But first of all, I need to get the EigenFace example running. I went to the book link for the Eigenface dataset, but the data format are different and there is no "allFaces.mat" directly available, which is used by the code example in the book. I searched in the book, but there is no instructions on how to make this "".mat" file. I could probably try to create this "allFaces.mat" with my own understanding of how this file is structured, but it would be much more convenient if you can provide it directly for the readers. Thanks!
Is psi U or V from the SVD? Thanks for the great content
Hi some of your videos on this playlist are on private.....
They are released on a schedule. Stay tuned!
which playlist is this from?
@@abdjahdoiahdoai Sparsity and Compression
Thank you Steve!!!
Since U shows measurements and principal components, V shows principal components and original features, it is not clear why/how QR of U gives the shortlist of sensor placement. Any clarification would be appreciated.
Hello, thank you for the amazing video ! The methodology seems very powerful. I just have a question about the Phi*Phi^t, used when the number of sensors is bigger than modes. Why did you use Phi*Phi^t and what are the change on the R matrix ? Because we only found that the diagonaux elements relating to the number of modes considered are bigger than those after. Thank you in advance for you return. Maurine
diagonal elements
Hi! Thanks for sharing this. I’m curious to know your views about connection of this to optimal experimental design?
Second this. Great video!
Hi!
Can you show a simulation(video) where is trying to classify diffrent images of different classes. I have been trying this with GNU Octave and this method did not work for me.
Does that not mean that in graphics engine, they could like 5x their performance by rendering random points only, like 20% and then upscaling that with compressed sensing?
12:20 for the short summary
How long does this take you to edit
Great question! It usually takes me about 5-10 minutes to make the intro 5 second clip and align the audio, touch up picture, etc., and then it processes on my laptop for 30min.
Aah
Could somebody point me in the direction of why the pivoted QR factorization minimizes the condition number. I can't find any source that says that.
Isn't this a really hard combinatory problem? I mean you essentially want a vector mask C that is optical for chosing k points of a big vector of size n. And you use those k number of measurements, multiply them by k eigenvalues and hope the sum is as close to n as possible.
Wouldn't this just mean you have n chose k options for said mask?
And your approach is to take the most weighted point from every single eigenvalue and that results in your mask?
I have to see some of the older lectures or sleep a day first to understand the mathematics of those QR matrix optimization.
What pen are you using in this video->ruclips.net/video/s2K1JfNR7Sc/видео.html
Can you lead me?
thank you very much
Every language you would ever program in... *drumroll* MATLAB.
Yeah.