So if you would normalize skin color and hair color etc, the performance of this algorithm would easily be improved (without extrapolating the face data to 3D). Well maybe not that easily, since this normalisation would have to be done by hand for now. Very interesting videos, you explain things very clearly.
I've been watching your videos for a couple of hours in a row, they are just amazing. This is a new topic for me, and I can say that I already have a very good idea of what SVD is, in only 1 afternoon! Thanks :)
Thank you, I’m really enjoying the course. It would be interesting to know how much any rotation/translation is relative to what humans recognize as features. What is the difference in the projected Swift/Stallone(?) experiment versus performance with {another random|carefully curated} data sets? I guess I’ll give that a try myself.
I wish you didnt use "econ" option. That i very specific to matlab, and I have a real trouble following your steps using mathematica. For example, my u matrix is square and my sigma has bunch of zeroes. I think I have trouble using your formulas because my u and sigma dont come from this weired "econ" mode, but a normal svd function... Please dump "econ"
Hi, this is not a new comment, but for any new readers, the 'econ' option specifies that the economy SVD should be taken, rather than the full SVD. This is explained in detail in the text, but in short, for matrices where n >= m (an extremely common scenario in the sort of data processing carried out in the text), the economy SVD is an exact representation of X, but saves considerable computation and memory. In short, the economy SVD is in a large majority of cases what should be used, rather than the full SVD representation.
@@andrewrittenhouse9742 i wasnt able to fully grasp fundamental meaning of svd, and i think very very few individuals do. If u think you do then try to derive hooks law from xyz data of 3 cameras looking at oscillating spring with mass. I wanted to really understand svd and pca fundamentally before sorting out computational efficiency.
@@user-ym8rz6mw5r it's great that you want to understand the SVD completely! I'm just explaining why the economy SVD was used here - and it is because economy SVD is an exact representation of X in a large majority of common use cases. You certainly can compute the full SVD for matrices where n>=m, there just isn't any point in doing so, because it doesn't yield any extra data and often comes with a considerable cost in computation and memory.
![SVD: Eigenfaces 1 [Python]](http://i.ytimg.com/vi/ofWji_wQBEE/mqdefault.jpg)
![SVD: Eigenfaces 1 [Python]](/img/tr.png)
![SVD: Optimal Truncation [Matlab]](http://i.ytimg.com/vi/u4LG9z4AAdM/mqdefault.jpg)
![Residual Networks (ResNet) [Physics Informed Machine Learning]](http://i.ytimg.com/vi/w1UsKanMatM/mqdefault.jpg)
![SVD: Image Compression [Matlab]](/img/n.gif)
Best teacher ever, this video should get more views
I wonder what a 3D plot of the PCs 4, 5, 6 would look like! Very interesting stuffs you're showing here :)
So if you would normalize skin color and hair color etc, the performance of this algorithm would easily be improved (without extrapolating the face data to 3D). Well maybe not that easily, since this normalisation would have to be done by hand for now.
Very interesting videos, you explain things very clearly.
I've been watching your videos for a couple of hours in a row, they are just amazing. This is a new topic for me, and I can say that I already have a very good idea of what SVD is, in only 1 afternoon! Thanks :)
Another "one of your favorite topics"
Caught ya! :)
Love your motivation and enthusiasm
it was so much fun !! N I learned so much..thank you!
I'm gonna have nightmares with those eigenfaces!
Thanks sir
Thank you, I’m really enjoying the course. It would be interesting to know how much any rotation/translation is relative to what humans recognize as features. What is the difference in the projected Swift/Stallone(?) experiment versus performance with {another random|carefully curated} data sets? I guess I’ll give that a try myself.
I wish you didnt use "econ" option. That i very specific to matlab, and I have a real trouble following your steps using mathematica. For example, my u matrix is square and my sigma has bunch of zeroes. I think I have trouble using your formulas because my u and sigma dont come from this weired "econ" mode, but a normal svd function... Please dump "econ"
Hi, this is not a new comment, but for any new readers, the 'econ' option specifies that the economy SVD should be taken, rather than the full SVD. This is explained in detail in the text, but in short, for matrices where n >= m (an extremely common scenario in the sort of data processing carried out in the text), the economy SVD is an exact representation of X, but saves considerable computation and memory. In short, the economy SVD is in a large majority of cases what should be used, rather than the full SVD representation.
@@andrewrittenhouse9742 i wasnt able to fully grasp fundamental meaning of svd, and i think very very few individuals do. If u think you do then try to derive hooks law from xyz data of 3 cameras looking at oscillating spring with mass. I wanted to really understand svd and pca fundamentally before sorting out computational efficiency.
@@user-ym8rz6mw5r it's great that you want to understand the SVD completely! I'm just explaining why the economy SVD was used here - and it is because economy SVD is an exact representation of X in a large majority of common use cases. You certainly can compute the full SVD for matrices where n>=m, there just isn't any point in doing so, because it doesn't yield any extra data and often comes with a considerable cost in computation and memory.
14:02 Arnold correlates closely with Taylor because Arnold is a girly man