Ai Dreams Forum

Software => General Software Talk => Topic started by: Hopefully Something on January 18, 2019, 07:57:21 am

Title: Modified Plan For Making AI
Post by: Hopefully Something on January 18, 2019, 07:57:21 am
Wireless artificial neurons are my dreamy goal, but not currently a productive pursuit. So the first step in the baby steps plan is:

Put artificial neurons on a Christmas light arrangement. Two wires, positive & negative, powered by one large battery, charged by one large solar panel. A few hundred neurons will be soldered in parallel to the wires, like steps on a ladder. Then some insulation will be applied (aerosol? dip?), then the arrangement will be twisted up like DNA and squeezed into an appropriately sized cranial cavity; with an input sensor and an output, so I can do experiments (evil laugh). The neurons themselves will still communicate with light. But each will only have three parts, 1) a PV sensor, 2) a micro-controller, 3) an LED. Signal identification could be done with a "secret knock" flash pattern before a transmission, or with wave forms if such things are optically detectable. Then the hard part, coding the MIC with something useful, I'm open to suggestions on anything, but that part especially.
Title: Re: Modified Plan For Making AI
Post by: LOCKSUIT on January 18, 2019, 08:49:23 am
We're both 23   : D

You sound like me when I was 18 lol.

If you want millions of neuron nodes, you use a computer to compute around all the 'virtual' nodes. You can even have a viz flyby in 3D of the simulated virtual brain and look around.

And it is duplicateable/cloneable and can upload to the cloud and run on even more processors/memory.

Google Cloud is like a hot momma just asking for power hungry fiends.
Title: Re: Modified Plan For Making AI
Post by: Hopefully Something on February 08, 2019, 07:22:57 am
Still want to go the physical route. I'm sure it will be faster than virtual once it reaches a certain complexity. But you're right, the single wire was crazy. The answer is plates! muah haha!! I will still rule the world after all. Once I visit the hardware store then I'll have easy parallel wiring! My enemies will be doomed! I've had multiple cups of coffee tonight...

(https://i.ibb.co/GHbSht9/20190207-224646-2.jpg) (https://ibb.co/8YSZhdb)
Title: Re: Modified Plan For Making AI
Post by: Korrelan on February 08, 2019, 09:30:04 am
Quote
I've had multiple cups of coffee tonight...

Haha... we can tell. 

Whilst I applaud your enthusiasm and ingenuity for creating a physical neural net I feel you still have quite a few technical difficulties to overcome.  The principle of a NN requires that the properties of both the connections and neurons can be altered on the fly.

Any ideas how you are going to achieve this?

:)

https://www.youtube.com/watch?v=5v1U7cCdWng
Title: Re: Modified Plan For Making AI
Post by: LOCKSUIT on February 08, 2019, 09:57:57 am
Omg this so cool lol. He thinks of a crazy idea that won't be good, like i used to or still do...

Korr, he has an answer for that :D We alwayyys have a backup/answer!

Idk if i got time but dude, no, ... Maybe using real neural stem cells but then you lose a lot of control over actually figuring out how a brain works, and are just creating more slow, brains, the natural way. I want fast movement, to get immortal.
Title: Re: Modified Plan For Making AI
Post by: Hopefully Something on February 08, 2019, 10:13:53 am
It will have to run in tandem with a traditional computer to reprogram individual nodes by some method. Maybe a genetic algorithm, or a well thought out plain old algorithm. Groups of neurons could be wired to a number of wireless transmitters. Then those could communicate with a desktop computer, or multiple laptops. This arrangement could also provide the effect of opposite forces spurring each other on and creating some sort of cycle of mutual improvement.
Title: Re: Modified Plan For Making AI
Post by: Hopefully Something on February 08, 2019, 10:26:51 am


It's gonna work man! Just wait and see, ye of little faith... gotta go sleep for 12 hour now.
Title: Re: Modified Plan For Making AI
Post by: Korrelan on February 08, 2019, 10:49:29 am
Quote
It's gonna work man! Just wait and see, ye of little faith...

I don’t doubt it for a second lol… but… why go to all the trouble and expense of building the matrix when you can just create it with one line of code, even in BASIC.

Dim a(100,100,100, 2) as Single

That’s it… so to set the brightness of a node on row 10, column 12, depth 20…

a(10,12,20,1)=10

It would look extremely cool once built; I'm just trying to save you time and expense.

Perhaps prototype/ simulate it first in code?

 :)
Title: Re: Modified Plan For Making AI
Post by: LOCKSUIT on February 08, 2019, 11:05:25 am
It's all the machine learners faults. We can't explain well, but neither, can yous! Else me and him wouldn't be stuck on GOFAI-idiot-ideas for years on end!

Why can't korr explain why ANNs are the way to go? Not just say it, but say why, and prove those features. For example, LSTMs keep propagating weight propagations and do some tricks in linguistics etc, but, why ANNs, and not my new researched way that does lots more? They are hard to teach, limited, have real issues, ex. not white box and no online learning usually and forgets nodes and blurs the data etc etc.

Why should I use a ANN to be a generalizing deductive reasoner like Don's AI that doesn't use ANNs?

If you could prove the reasons, the key features found in RNNs, VAEs, etc, how, why, what they work, like the weight, 3D word2vec coordinates, where the data is, etc, then it may need to be explained to us like this:

ANNs are x
Y is x.
B is x
Y and B are good, therefore ANNs are good.
ANNs use weights to do coordinate clustering in Dimensional space.
External inputs do too, therefore...
And more support on those key fact claims.
Prove it.

ANNs suck

Now time to rant.

I don't even know totally how they work and I read so many places.

All I know from top of mind is they made of input and output nodes, and the hidden unit nodes make up a hierarchy connected by weights that propagate mathematical signals around like LSTMs do. They run on GPUs typically for parellal speedup. They try to get taught and give desired outputs to the user. They always give messy sloppy results. The high dimensionality curses the backprop clustering and the generalization in the weights could help make discoveries in text. What annoys me most is they think LSTMs know what comes next by magic when it not.
Title: Re: Modified Plan For Making AI
Post by: LOCKSUIT on February 08, 2019, 03:04:22 pm
Remember folks: as korr linked us, I'm surprised it says many key words:
https://en.m.wikipedia.org/wiki/Scientific_method
We gotta do internal Don Korr Deductive hypothesis consequence predictive reasoning generalization in our knowledgebase cluster using logic, and do external tests of the sequence predictions too as a final narrow down method more scrutinately of the informings.

As you know, Lock finds a way to get away with not having to do too many actual testings externally.
Title: Re: Modified Plan For Making AI
Post by: 8pla.net on February 08, 2019, 04:41:45 pm
For discussion purposes only, this C Language source code listing is prototypical.
The variable names are: M for matrix, R for row, C for column, D for depth, and T for time.

It may roughly be like one little "ice cube" corner,  in the video Korrelan posted.
Otherwise the code would get long, which is OK, but not needed for this demo.

Scroll down to this line: if(R==0){M[R][C][D][T]=0x8888;}
To see where the row is being filled in...

Code: [Select]
#include <stdio.h>

int main()


int R, C, D, T, M[2][2][2][2]={
{
  {
    {{0},{1}},
    {{2},{3}},
  },
  {
    {{4},{5}},
    {{6},{7}},
  },
},
{
  {
    {{8},{9}},
    {{0xA},{0xB}},
  },
  { {{0xC},{0xD}},
    {{0xE},{0xF}},
  },
}
};
 
for(R = 0; R < 2; R++) {
    for(C = 0; C < 2; C++) {
        for(D = 0; D < 2; D++) {
    for(T = 0; T < 2; T++) {
if(R==0){M[R][C][D][T]=0x8888;}
printf("MATRIX[%d][%d][%d][%d] = %04X\n",R, C, D, T, M[R][C][D][T]);
}
    }
        }
   }
return 0;
}


Compiles with: gcc matrix.c -o matrix.o
Runs with: ./matrix.o

Program Output:

Quote
MATRIX[0][0][0][0] = 8888
MATRIX[0][0][0][1] = 8888
MATRIX[0][0][1][0] = 8888
MATRIX[0][0][1][1] = 8888
MATRIX[0][1][0][0] = 8888
MATRIX[0][1][0][1] = 8888
MATRIX[0][1][1][0] = 8888
MATRIX[0][1][1][1] = 8888
MATRIX[1][0][0][0] = 0008
MATRIX[1][0][0][1] = 0009
MATRIX[1][0][1][0] = 000A
MATRIX[1][0][1][1] = 000B
MATRIX[1][1][0][0] = 000C
MATRIX[1][1][0][1] = 000D
MATRIX[1][1][1][0] = 000E
MATRIX[1][1][1][1] = 000F


Title: Re: Modified Plan For Making AI
Post by: Hopefully Something on February 10, 2019, 07:46:37 am
Perhaps prototype/ simulate it first in code?

Here's an attempt using excel. What I'm trying to do here is analyze an input in a variety of ways and combine those different "perspectives" into an output.

(https://i.ibb.co/DKMQzsH/Screenshot-805.png) (https://ibb.co/bdvXQqM)
Title: Re: Modified Plan For Making AI
Post by: Hopefully Something on February 10, 2019, 08:44:32 am
When I stimulated it from the side, adding a value (1) then returning the state of the cell to (0), waves of "energy" "or "numbers", traveled away from the spot, splitting off from the main energy mass. It's cool but I'm not sure what to do with it yet.

(https://i.ibb.co/3m2qWCP/Screenshot-764.png) (https://ibb.co/N25PpyG)

(https://i.ibb.co/cgTTzM8/Screenshot-765.png) (https://ibb.co/vwzz2gd)
Title: Re: Modified Plan For Making AI
Post by: Korrelan on February 10, 2019, 10:01:42 am
Excellent, early prototyping is a very important stage.

Besides giving you a better idea of the problem space it will give you insights into the logic required if you do decide to build it physically.

https://en.wikipedia.org/wiki/Liquid_state_machine

ED. I believe you’re only young so I can’t resist giving old man advice.

If you idea works then cool… if it doesn’t, do not get disheartened.  Sometimes just knowing you have an idea can be comforting concept, if through research you realise it’s not going to work it can leave a hole/ gap in your enthusiasm.

The project is important but secondary, the most important thing is developing the best tool you have… your brain. 

 :)
Title: Re: Modified Plan For Making AI
Post by: LOCKSUIT on February 10, 2019, 11:41:21 am
Looks like something of mine I made, but his is a simulation by the looks of it.

However I suppose mine was a simulation too. When I was 13, I thought wacky stuff:

https://www.youtube.com/watch?v=o4hlHVmDAko&feature=youtu.be

And I believed it was a valid plan all the way up to age 17 until talking to people online and learning/intaking more knowledggge/infO on the internet.

I kept wanting to meet with professors and get into the university microscopes.

In fact I still give it a thought of honor even now, who knows what we can do with photons.