/biz/ - Business & Finance

>>54414108
Neural networks have been the next big thing for a while now. Everything from speech to text, picture touch up, speech synthesis, art creation, and even simple things like object detection in photos have all been because we created neural networks. They've been the first significant jump in AI ever since we coined the term.

What makes them unique is that they need to be created but then trained. So for example if you have a simple NN that has to look at a number and tell you if its 0-9, you'd design the network and then start training it. You'd show it millions of photos and have it guess and then tell it if it got it right or wrong. But this is absurdly complicated. You have to quantify why its wrong and measure it, and then do some bullshit reverse matrix integration to propagate that error down the network to the beginning and tweak it along the way. It's insanely complicated mathematics.

The end result is that once the program works, you have a unique neural network that you yourself don't even understand. You didn't create it, you made the initial net of millions of nodes and then trained it and it created itself. Two people could train the same initial network and arrive at completely different final results. Because its not just the individual nodes but the endless amount of connections between those nodes that are tweaked and grow to be weak or strong from 0.00 to 1.00 in strength.

When a NN sees an image the date goes into a senory input node and just like life that signal travels down into the millions of nodes like a splash in a pond but into an artificial brain, and the end result somewhere in there is what makes the AI make an answer that is correct.

The important thing is this final result is something you can patent. A company could devote its entire existence to designing a single type of AI Neural Network that does one thing. That new age is basically beginning.