I was always into the fresh new stuff, but as it turns out, the fresh new stuff came from way back, just there was no computation power available.
- unsupervised neural networks.
- global illumination raytracing
- subsurface scattering
global illumination, is the pissweakest one, you can figure it out cause its all a pack of adds. the radiosity method is the hardest one to do, and goes the slowest.
subsurfacescattering, you sorta have to be settled for a proximity blur, instead of the real analytical version which will drive you insane trying to learn it. im yet to even work that one out.
The craziest analytical neural network, (the boltzmann machine) happens to have a little friend, which as long as you supply the sigmoid, and feedback deblurring, (the restricted boltzman machine) then your fine, itll run real time, but theres much optimization to do... to get it flying.
my favorite one is the neural network, once fully optimized and functioning, itll be able to store motion video games, that look real, and you replug the system by changing the weights, and thats all it is, just happens to be 4 bytes per weight, but actually only the first level needs detailed weights because the rest is all black and white contrast caused by the sigmoid!
I even tried to use a linear model, cause i was thinking it would end up involving more computation and be a superior system, but i tried it and it just went black.
so thats a point lost for linear.