Poll of the Day > Is there an AI framework like this

Topic List
Page List: 1
Yellow
07/28/20 1:03:15 AM
#1:


I worked with a couple for fun and they basically went like this.

Set up an array of neurons using the most popular library, run them overnight, (spank it if it does bad, and reward it if it gets it right or closer to right) and in the morning hope it works. There's no way to understand what it's doing to get it right, and little way to correct mistakes aside from tweaking a setting or two.

The most challenging part is figuring out how to gather and format your data. The networks are extremely inefficient at accomplishing anything, and are not deployable at all, due to being mostly Python environments with specialized GPU drivers. They accomplish nothing if a human cannot do it as well.

Anyway my idea of a network would be to create a library of networks that would index, look up, and apply logic based on context, with a reward incentive for doing as little calculations as possible to get the correct outcome. A network that would be trained for one task would reference another if it found it beneficial. It would create a new network if it saw that no existing ones would fit. It would not fire off neurons that were unused, it would train while running. There would be no limit to the size of this network as it would optimize itself and not become slower as it got larger. It might be able to identify thought processes as objects to be referenced, like how we can reference and be aware of ideas, unlike the brute force 30 by 3 neurons brute force Frankenstein microscopic slice of simulated brain matter solution we commonly use. (Which is an incredibly unoptimized attempt of emulating something akin to molecular simulation, like simulating a batch of water molecules instead of doing the math and figuring out what tempurature they evaporate at)

I've been into emulation for a while, I consider modern NN to be a LLE of something we haven't got the technology to ever fully emulate, when we don't even need to be using HLE, but instead making our own solution using the benefits of computer hardware.

I only ask because I want to work on my custom framework and I have a bad habit of doing things before looking up if other people have done them.

---
... Copied to Clipboard!
Topic List
Page List: 1