Ai Dreams Forum

Member's Experiments & Projects => General Project Discussion => Topic started by: HS on May 12, 2020, 01:52:51 am

Title: Mind Molecules: Creating a Homogeneous Substrate
Post by: HS on May 12, 2020, 01:52:51 am
My new physical neural net idea consists of five enmeshed neural nets made of fitted together molecule like pieces consisting of: 




(https://i.ibb.co/h20kYDf/Screenshot-1406.png) (https://ibb.co/9yPJZst)


These are like the five senses of the brain. First, the outside world is mirrored or reproduced inside the brain, then you need to become aware of this internal world, systems which don’t require constant attention are delegated to the autonomic net, you feel a certain way about the things which are in your awareness thanks to the limbic net, your reasoning net lets you do something about it.

It's probably inefficient to have separate sensation/emotion/reasoning centers. They need to communicate anyways, why bottleneck the information flow and waste space with separate modules? Better if they were totally enmeshed, which would allow for all these separate systems to coexist symbiotically from the get go.

This would have to be organised somehow to be efficiently reproducible. It would require some sort of crystal lattice like structure. A cube has six sides, so I should be able to enmesh five nets… if I can figure out the right pattern. Like any homogeneous mixture, you only need to study a small part of it to know how the rest should go. So, this brain substrate could be scalable, depending on the requirements.

If I could discover all the necessary systems (possibly nets), and come up with a pattern of enmeshment, these should be able to parallel process to the upmost degree.
Title: Re: Mind Molecules: Creating a Homogeneous Substrate
Post by: infurl on May 12, 2020, 02:09:40 am
I really hope that you are simulating all your designs in software to see if they actually work. Until you have enough money to design your own silicon chips and have them manufactured, there is nothing you can do with electronic components that would compare to what you could do in software. Admittedly though, making things is fun. I like making robots and the electronics that go with them too, even though it's useless compared to what I can accomplish with software.
Title: Re: Mind Molecules: Creating a Homogeneous Substrate
Post by: WriterOfMinds on May 12, 2020, 06:54:40 am
I have to admit I'm also worried about whether you'll be able to scale your custom hardware up to the required processing speed, given that you aren't ready to miniaturize it to microscopic sizes yet.  But I like the way you're thinking through the different systems.

What does the awareness net do?  Is it like an attention/focus selector?

If you want a way to prototype the lattice network, but software isn't parallel enough for you ... have you thought about an FPGA?
Title: Re: Mind Molecules: Creating a Homogeneous Substrate
Post by: HS on May 12, 2020, 06:16:12 pm
I really hope that you are simulating all your designs in software to see if they actually work.

That sounds like good advice. I don't want to... but I should do that.

Admittedly though, making things is fun. I like making robots and the electronics that go with them too, even though it's useless compared to what I can accomplish with software.

True. I don't want to lose that and end up with a neural net tailored to the strengths of software and simulation, because there are so many of those. I want to invent/explore something else.

I have to admit I'm also worried about whether you'll be able to scale your custom hardware up to the required processing speed, given that you aren't ready to miniaturize it to microscopic sizes yet.

Speed could be orthogonal to quality, like ability is orthogonal to morality. I'd rather focus on the precise quality of the AGI, and speed should be a simple matter later on.

What does the awareness net do?  Is it like an attention/focus selector?

Yes, as far as I understand how awareness functions, that's a good way to put it.

If you want a way to prototype the lattice network, but software isn't parallel enough for you ... have you thought about an FPGA?

I’ve had FPGA'S suggested before. Back then I was less certain about what my ideas would become and require, but now might be the time to codify these vague concepts. I suspect a repeating pattern of neurons will lend itself well to the benefits of virtual space, with virtually identical cost, and similar amounts of work needed to model various scales.

EDIT: I guess it's not really virtual space. It's already an array of programmable gates lol. If they made 3D ones, that would be perfect, but then they might have problems with temperature control. Anyways, thanks, I'll look at the options for virtual, and semi-virtual neural net modeling.