1. I know there's been some work regarding the interaction between topology and deep learning. Holes tend to be more limited by the number of layers rather than the number of neurons. More neurons allow for more complex boundaries along the holes, IIRC. 2. Another approach to neural image compression would be pixel prediction; trying to predict the next pixel from a given sequence of previous pixels, perhaps a few to the left and a few above. If you want to turn that into a file format, you just include the image height/width, neuron weights, some initial starting pixels, and then just store the difference between the predicted values and the actual values, which should generally be pretty small and should be able to use very few bits for most pixels. Also, because each pixel relies on previous pixels, you get much of the power of a very deep neural network while only needing to train a very small one.
@BryanChance
Жыл бұрын
Tsoding, I think you are the most talented programmer (if that's the right word) that I've ever seen on KZitem. I've seen much of your other videos too. It covers many topics and subjects. I don't understand at least 50% (or more? I don't know what I don't know hehe) of what you do but I still learn something. :-) Thank you very much.
@easyBob100
Жыл бұрын
A lil tip for ya. Don't init your weights between -1 and 1. Try something smaller like -0.2 and 0.2.
@visintel
Жыл бұрын
very cool that you described a simplified version of NERF. instead of the function taking x and y, you should try taking the sinusoidal embedding of x and y to see how it can improve high frequency features (harder edges)
@devonaanika
Жыл бұрын
That int matrix (38m in) is awesome. This is a wonderfully done tutorial! Thank you for creating this :)
@sheevys
Жыл бұрын
For that kind of use cases people either encode the input first by taking sine and cosine with increasing frequencies (NeRF), or use sine as an activation function (SIREN). For the latter, there are some smart initialisation for the weights required, but it only matters for deep networks.
@anon_y_mousse
Жыл бұрын
That last thought about neural nets being a sort of compression, it's true, it's like a very highly lossy form of compression that is computationally expensive to process. There are some decent compression programs out there that use statistical analysis of the data to generate a model that can accurately represent it, and I'm kind of thinking of using something like that if I ever get around to making the game I've got in my head. Take all the assets and compress them in such a way that a model can just spit them out lickety split, but accurately. I'd consider it a win if I could fit a whole 3D game on a floppy. I know it's possible, and has been done before, but I've never tried it, so I figure it'll be fun. Your command line processing is a bit weird. Why don't you just use an index into argv? When you do anything more complex, like with command line switches, you should look into using getopt, as it's a very helpful "library". I'm curious about that right click bug, is that a Chromium issue or is it due to particular websites? With Firefox when a website attempts to suppress the right click menu I can hold shift and right click and it will bypass that nonsense. My alternative thought on that is that it may be the WM that you're using, which would be really weird but might explain how swapping apps makes it function again.
@gameofpj3286
Жыл бұрын
Really enjoyed this one :D Nice work
@pauleagle97
Жыл бұрын
Amazing material, I have just tested out my framework built from the previous episodes in Python (using Numpy only). It works! (though not as fast as your C implementation)
@abanoubha
Жыл бұрын
was waiting for this video 🎉
@vitorsantos6310
Жыл бұрын
Do you get project inspiration from classes you liked while at University? I say this because ever since I started following you there’s always some sort of coincidence between my current uni classes and your project themes .
@k5tggiv6ubmmb3as9
Жыл бұрын
don't think he went to university
@NjRmMTIxMjVlNDFiNDkzMDVh
Жыл бұрын
@@k5tggiv6ubmmb3as9 He did, but not for compsci.
@_.madpie._
7 ай бұрын
@@NjRmMTIxMjVlNDFiNDkzMDVh he's a bachelor of compsci, it's in his faq page
@nesvand
Жыл бұрын
Around 47:00 you discuss why you're using C and suggest you'd happily use Jai as an alternative. What are your thoughts on using Zig? By all accounts it seems like it's intending to be as simple as C while addressing the issues of the workflow, and it's easily accessible if you want to provide shared code.
@heyclore5728
Жыл бұрын
There's also programming in Zig on his list of videos.
@nesvand
Жыл бұрын
Cheers, just had a look and I'll check out the vids in full when I have a chance :D
@c.Orange
Жыл бұрын
✨️shit right here
@wakatuvia
Жыл бұрын
ChatGPT learns from you 🙇♂️
@dewijones92
6 ай бұрын
so awesome. I would love more AI vids?
@jimshtepa5423
9 ай бұрын
feel some geohotz vibes around 2:08:40
@labsendeyshent
Жыл бұрын
2 minutes ago tsodinPog
@Alguem387
Жыл бұрын
Did you ever tried odin?
@SolathPrime
Жыл бұрын
Ahoieee Azozy
@varshneydevansh
Жыл бұрын
I really loved your content. From where could I improve my programming skills/ concepts?
@EmOui-l3q
Жыл бұрын
Coffee
@calebfalcon9424
Жыл бұрын
Oolong tea
@kgra69
Жыл бұрын
From your computer. Just write code dude.
@erfanmirzazadeh1248
Жыл бұрын
You can start by doing advent of code's puzzles you learn new things when looking for the answer
@rian0xFFF
Жыл бұрын
just type ascii characters in the screen
@TheTormozz
Жыл бұрын
What are you doing in real life? What is your profession? In my mind, a person with such skills works for NASA or Pentagon, or Google
@CaptainBullzAQW
Жыл бұрын
probably a react developer who just enjoys recreational programming on the side
@andresconrado
Жыл бұрын
I would bet he works as clerk in a supermarket or something equally mundane.
@Vethane
Жыл бұрын
47:17 what about lua?
@duartelucas5746
Жыл бұрын
what is your machine?
@tragdate
Жыл бұрын
When you said Romanian language content I dreamt that you were talking about my channel 😂
@tragdate
Жыл бұрын
Shameless plug
@tragdate
Жыл бұрын
Also I think this is the wrong video, you should wear multiple shirts throughout a series 😂😂
@L1Q
Жыл бұрын
imagine "inventing" your own shitty compression "algorithm" and using it as an argument to bypass copyright xD
Пікірлер: 41