The paper covered: J. A. Weyn, D. R. Durran, and R. Caruana, “Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere,”Journal of Advances in Model-ing Earth Systems, vol. 12, no. 9, Sep. 2020,issn: 1942-2466.doi:10.1029/2020ms002109.[Online]. Available:dx.doi.org/10.1029/2020MS002109. Code: github.com/jweyn/DLWP-CS CNN explanation video►: ua-cam.com/video/YUyec4eCEiY/v-deo.html
Thanks for the great explanation! I didn't understand why the longitude and latitude cannot be used in their original scales and why the cube has to be enforced. I'll make sure I read the paper as well. Thanks!
@@devashishd12 Assume that grid cells have 4 sides. Thus you can travel from one grid cell to 4 adjacent other grid cells. For all latitudes, you can tell the model, that the longitude 179°E connects to 178°E on the one side and 179°W on the other side. I.e. you can make the world cylindrical and travel from immediately from the left edge to right edge of the map. But at the poles you run into issues: For latitude 90°N, you cannot tell the model that all longitudes connect. I.e. you cannot tell the model that you can immediately travel from any longitude (let's say 112°E) to any other longitude (let's say 4°W) at 90°N; which you can easily do by simply rotating in reaility. That is meant by the mathematical singularity at the pole.
Let's talk about Fast Transform nets a little more. You use fixed dot products (enacted with fast transforms) and adjustable (parametric) activation functions like fi(x)=ai.x x=0, i=0 to m. The fast Walsh Hadamard transform will do very well. Then the computational cost per layer of width n is nlog2(n) add and subtract operations, n switching decisions and n multiplies. In particular on an application specific chip the transistor count would be very low. Anyway to stop the first transform from taking a spectrum you can a randomly chosen fixed pattern of sign flips to the input data and use a final transform as a sort of readout layer. Such a net then is: sign flips, transform, activation functions, transform, activation functions......transform.
Actually it could be a sub-random pattern of sign flips. And you can have low curvature initialization which is very helpful for training deep nets. Maybe will have to learn the fast Walsh Hadamard transform. Which was popular in the 1960s due to its low computational demands.
The paper covered: J. A. Weyn, D. R. Durran, and R. Caruana, “Improving data-driven global weather prediction using deep convolutional neural networks on a cubed sphere,”Journal of Advances in Model-ing Earth Systems, vol. 12, no. 9, Sep. 2020,issn: 1942-2466.doi:10.1029/2020ms002109.[Online]. Available:dx.doi.org/10.1029/2020MS002109.
Code: github.com/jweyn/DLWP-CS
CNN explanation video►: ua-cam.com/video/YUyec4eCEiY/v-deo.html
Thanks for the great explanation! I didn't understand why the longitude and latitude cannot be used in their original scales and why the cube has to be enforced. I'll make sure I read the paper as well. Thanks!
Let me know if it is still unclear, I can try to enforce this explanation. :)
@@WhatsAI Could you please explain briefly? Thanks :D
@@devashishd12 Assume that grid cells have 4 sides. Thus you can travel from one grid cell to 4 adjacent other grid cells.
For all latitudes, you can tell the model, that the longitude 179°E connects to 178°E on the one side and 179°W on the other side. I.e. you can make the world cylindrical and travel from immediately from the left edge to right edge of the map.
But at the poles you run into issues: For latitude 90°N, you cannot tell the model that all longitudes connect. I.e. you cannot tell the model that you can immediately travel from any longitude (let's say 112°E) to any other longitude (let's say 4°W) at 90°N; which you can easily do by simply rotating in reaility.
That is meant by the mathematical singularity at the pole.
I suspect that with this system you should be able to predict Climate Change outcomes into The Future as well
What do you think about Weather.com predictions? They are owned by IBM for several years and claim that improvement in calculation was made.
Let's talk about Fast Transform nets a little more. You use fixed dot products (enacted with fast transforms) and adjustable (parametric) activation functions like fi(x)=ai.x x=0, i=0 to m. The fast Walsh Hadamard transform will do very well. Then the computational cost per layer of width n is nlog2(n) add and subtract operations, n switching decisions and n multiplies. In particular on an application specific chip the transistor count would be very low. Anyway to stop the first transform from taking a spectrum you can a randomly chosen fixed pattern of sign flips to the input data and use a final transform as a sort of readout layer. Such a net then is: sign flips, transform, activation functions, transform, activation functions......transform.
Missing 'apply'.
Actually it could be a sub-random pattern of sign flips. And you can have low curvature initialization which is very helpful for training deep nets. Maybe will have to learn the fast Walsh Hadamard transform. Which was popular in the 1960s due to its low computational demands.
'you will have to learn' I already know it.😼😼😼
Bruh you are speaking language of gods