Kolmogorov-Arnold Network (KAN) - A New Deep Learning Architecture to Disrupt a $10T Industry?
HTML-код
- Опубликовано: 22 авг 2024
- Will Kolmogorov-Arnold Networks (KANs) disrupt the AI industry once again? Are they strictly superior to Multi-Layer Perceptrons (MLPs)? What are their weaknesses? Let's find out in this video!
♥️ Join my free email newsletter to stay on the right side of change:
👉 blog.finxter.c...
Also, make sure to check out the AI and prompt engineering courses on the Finxter Academy:
👉 academy.finxte...
🚀 Prompt engineers can scale their reach, success, and impact by orders of magnitude!
You can get access to all courses by becoming a channel member here:
👉 / @finxter
Very good presentation - interesting, exciting, informative, and balanced. Thank you!
Thanks, appreciate you like it. I tried to keep it simple and focused on the key insight, i.e., replacing dumb functions that make up an intelligent function with more intelligent functions that make up an even more intelligent function. ;)
That answers some questions.
When trying to implement NNs myself, I was amazed by simplicity of functions used in typical NNs. And I tried to utilize some more complex functions, at least activaton ones. In the result, I came to the same conclusion: more complex activation functions demand more compute, but their benefit is quite not obvious. So, in the end I used ReLu or something similar.
I use NN to sort some images on my pc. The goal is to rank them according to the likelihood that I'll find those images visually pleasing. Program works ok, it does the job.
That's great although I think using NN with small data and small training (e.g., on your computer) is likely to be overly complex. Their strength is mostly on HUGE data with MASSIVE compute due to the scaling laws.
@@finxter That's true. Real power is in big numbers.
However, my experience shows that some significant estimations can be made even on simple networks. For me, the network has to vagely recognize the scene and asess image quality. I suppose that can be done even on moderately sized networks.
As for training dataset, mine counts tens of thousands images. Dataset has only two categories: accepted and regected. And program has to guess likelihood that given image will be in one of those categories. On other words, NN gives a number between 0 and 1, but tries to be as close to a expected answer as possible. As I said, results are quite acceptable.
One notable obstacle is that I have no Nvidia card, so that had to be set up to work on Radeon and under Windows. Still, it works and I'm content.
If you are curious about KANs, find a different source than this video.
Fair. For instance check out this one for a more technical explanation (without memes as neurons): ruclips.net/video/CkCijaXqAOM/видео.htmlsi=90V0Vj11XlRhnKvD