If you're here because of Fireship, be warned my English pronunciation is terrible, and I sound like a potato trying to talk on it :( Also - I made many simplifications here, as not to over-complicate the explanation to my colleagues (including not explaining the difference between inets and icombs, not going in deep in the literature of optimal reduction, etc.) - if you're looking for a proper overview of ICs, this is the video you're looking for: ruclips.net/video/sDPuQ-UjhVQ/видео.html - it is Portuguese, but the English subs are good, and I made it as pedagogical as I could. It is worth watching (unlike this one haha)
I dont understand much of it yet but I was so fascinated by Bend that I had to try that out. The performance improvements are massive in parallel computing but so far I have only scratched the surface.
Sorry, but I think you've got better luck scaling unconventional neural networks trained using backprop (esp sparse neural networks) with CUDA. Scale is not the only piece of the puzzle. It's scale + learning algorithms that scale with compute, and the only one known to do so is deep learning.
Who came from Fireship ?
This is going to be BIG. This is no small advance in runtime research.
Bravo !!!
If you're here because of Fireship, be warned my English pronunciation is terrible, and I sound like a potato trying to talk on it :( Also - I made many simplifications here, as not to over-complicate the explanation to my colleagues (including not explaining the difference between inets and icombs, not going in deep in the literature of optimal reduction, etc.) - if you're looking for a proper overview of ICs, this is the video you're looking for: ruclips.net/video/sDPuQ-UjhVQ/видео.html - it is Portuguese, but the English subs are good, and I made it as pedagogical as I could. It is worth watching (unlike this one haha)
Your english is impecable. Great job!
I dont understand much of it yet but I was so fascinated by Bend that I had to try that out. The performance improvements are massive in parallel computing but so far I have only scratched the surface.
can you imagine that finally Emacs will be multithreaded ?
100% sure that was taelin's goal
Wow, I didn't know this video existed. Great video!
Good stuff!
This is gnarly... Its like a real life Silicon Valley episode. Instead of finding the "Jerk Ratio" he used old hieroglyphic scriptures from the past.
brilliant
Really inspiring, thanks!
Russell Gardens
LOVE THIS!
Pfannerstill Burg
Sorry, but I think you've got better luck scaling unconventional neural networks trained using backprop (esp sparse neural networks) with CUDA.
Scale is not the only piece of the puzzle. It's scale + learning algorithms that scale with compute, and the only one known to do so is deep learning.