Ai Dreams Forum

Member's Experiments & Projects => AI Programming => Topic started by: frankinstien on August 01, 2022, 02:38:41 am

Title: Interesting Machine Learning course
Post by: frankinstien on August 01, 2022, 02:38:41 am
I took this Coursers course: Machine Learning (, it uses Octave/Matlab which I never had used before. The course emphasizes the matrix and vector approaches where I'm now going to replace a ton of code with Matrix libraries.  :) Octave can use your GPU, but you have to compile it with the library to do that. But over all a good course that explains the math of machine learning clearly and how to implement solutions for large data sets. I would suggest taking this course before endeavoring on other literature, it really will help you understand what's being done on other fronts.
Title: Re: Interesting Machine Learning course
Post by: MagnusWootton on August 01, 2022, 01:22:38 pm
A neural network can store any kind of ai on it,  it stores it as a sequence of inputs to outputs and u can use backpropagation to train it into the weights of the system.

The neural network itself isnt really the brains of the situation tho,  because it can be anything, its like water it depends on what the weights are,   its what makes the I/O pairs, is the "brains" what develops what the weights of the network are going to be.   And it can be anything, it can just be an ALU,  or any IC with some binary inputs and binary outputs to go with them,    but if it was what a robot is going to do with its arms and legs, depending on what it sees with its eyes,   then its more what i would call actual A.I.  but its just an I/O machine the same.

If you just record what the screen is,  and what the person did at each screen, then u just train these I/O pairs into it,  and thats why neural networks can be clones of people,  if u went all the way to doing the whole horrible recording job required...   but maybe theres a better way to do it than just that.

Architectures of neural networks maybe help the program tween between the I/O pairs because there is an exponential gap (of missing keys) between the recording "keys",   so if there was a way you could make sure all the halfway frames were correct as a tween to the actual real data,  then that might be what ur after - as a better system?   But theres more than that too.

But the record is just a record,  its how you develop the I/O pairs is what Ai properly is,  not just a neural net recording things and smodging it all together in a collected transform of synapses.  More like what Acuitas is,  you could smodge him into a bunch of synapses as well, but hes actually making the inputs and outputs so hes more important. :)