Maker Pro
Maker Pro

analogue perceptron

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
Just got my first backpropper (neural network dataset based trainer) to work! congratulations me!!!

Most people will implement a perceptron with a computer, microcontroller.
But if u dont do that, and u put it all down in solid hard connections, you could get really good performance out of one, it pops out the other side just with electrical delay, so its barely anything at all.

The kind without feedback is what I'm talking about, just a direct pathway from I to O, not even a register or oscillator is required.

I also have this nifty idea where I can avoid all the over crossing of the wires to do the full all onto all connection between the layers, by putting it only an ordinary grid! it seems its possible to do it on a 2d grid, where each grid unit centrepoint is a synapse.

So u could actually make it a potentiometre! so its just an array of potentiometres, which adjust the multiplication strength, then they just have to add together somehow, I wonder how I could fix it.

Anyone seen a guy with one of these before?
 

kellys_eye

Jun 25, 2010
6,514
Joined
Jun 25, 2010
Messages
6,514
Anyone seen a guy with one of these before?
Seem to recall seeing one in......fantasy land.

If your aim is to 'impress' then try doing it with something you've really done, not just picked off t'internet and exaggerated....
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
Heres some old stuff on it, its been around for ages, maybe everyone who knew anything about is dead now.



They probably work pretty good, probably not too complicated as well, I could go draw something up, I was just wondering if anyone here knew anything about it. Maybe Hevans? I'm sure hes got an old story for me. :)
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
perceptroncircuit.png
ok so ive got a circuit, I have the stuff in my house so I might put it together, see if it works!

so its a lowpixel ldr screen. (and u have to make the occluders for each of its eye pixels) connected up to potentiometres, which you synch up to the output.

you have to synch it whilst its looking at the pictures, and then do it until youve gone over the pots 2-3 times, and then it should respond to the pictures by itself, after its set.
 

kellys_eye

Jun 25, 2010
6,514
Joined
Jun 25, 2010
Messages
6,514
What?

or, more specifically, W.T.F?

(add-on) - are you talking about?
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
Hey Kelly!!!
So this little circuit, should be able to recognize and tell apart different objects you show it, and it does so by turning an analogue mechanical dial, (you can use a servo for it, you input into the signal wire into the servo to change the angle.)
and it just points at the different things in a little half circle, when u put the thing in front of the view window, which is comprised of Light dependant resistors.

If you want to show it a 3d thing, you have to teach it the different sides separately, because it sees it in 2d, but it can still recognize it.


perc2.pngheres the circuit fix, just need to add a little negative bias at the motor, so the dial can go left and right, otherwise u cant recorrect backwards.

you need to make it a separate battery, note the batteries will go in series and its a bit of a waste of power, but it should still work, just put a resistor there so it doesnt waste the battery tooquickly.
Also the potentiometres need to be set so the dial isn't turning.

And maybe instead of powering off one battery, maybe u make all the lines separate powered maybe with capacitors with resisors segregating it, so you dont get load hogging error on the pots.
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
ok so theres a prob with the fix, I need the dial to go back and forth with the potetiometre opening and closing
I couldnt think of how to do it just with powering it, but if u add 2 diodes like how I show, and run it on ac instead,
then it should work.
perc22.png
 

Martaine2005

May 12, 2015
4,932
Joined
May 12, 2015
Messages
4,932
You lost me in the first post.
And now I’m even more lost-er..
 

Martaine2005

May 12, 2015
4,932
Joined
May 12, 2015
Messages
4,932
All it’s missing is a 270 milliHenry capacitor on the biased negative side and 2 x 47uF resistors across each LDR. Should work a treat.
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
So the potentiometres are the synapse weights, and they can be negative weights or positive weights, clamp the input. (Put the object to be classified/recognized in its view, and do it from 3- 4 different angles) and then just pick a pot that tweaks the output dial to the right position.

I just simmed it in Falstad, seems to be correct, but I couldnt fully test it properly because of interface blues.

Add an led to the potentiometres, cause u only should adjust a pot which has alot of signal going through it, to tweak the output.


One layer of a neural network looks like it makes a simple passive circuit, but putting the multilayer in is what poses all the hardware difficulty. (where one would be then reaching for his transistors.)


Heres a link to the falstad sim, note the camera is just switches, and its not fully linked cause it took too long, so thats why I couldnt fully test it, but it did work just the top pixels or so! So should I build it I can't decide.

https://www.falstad.com/circuit/cir...ktMtJFJxVVcZW1FIsFWUMZatXVGtQpMxXVOtEmvhYQQAA
 

hevans1944

Hop - AC8NS
Jun 21, 2012
4,878
Joined
Jun 21, 2012
Messages
4,878
heres a Girl with one on the INTERNET!!!! EDIT: Here's a girl with one on the Internet!

The above quote (edited) from your post #12, and the attached YouTube video in that post, provides a good example for you to follow, @dragon. Note in particular her use of real parts and real test equipment to test out her breadboard demonstration. Not a simulator anywhere in sight!

Maybe it's a Generation X, Y, or Z thing, along with the presence of texting on mobile devices, that leads people (peeps) or you (u) to butcher English text while attempting to communicate here in the MakerPro forums. But it seems to me that @dragon does not possess sufficient English Language skills: spelling, grammar, sentence structure and the proper use of abbreviations (to name just a few)
to participate meaningfully in any discussion related to machine learning and its corollary artificial intelligence.

I will take the bait hinted at in post #3: "Maybe Hevans? I'm sure hes got an old story for me. :)" Nope, this is your show. What I know (or pretend to know) about neural networks and AI is so obscure, mundane, and (ask Elon Musk) so frightening that I don't even want to think about it, much less discuss it in an open forum for anyone to read.

BTW I am (by definition) an artificial intelligence, since before I existed I wasn't conscious of any intelligence. I was assembled in a complex, self-directed, process that took place in my mother's womb. Same-o, same-o with my brother, too, except he was assembled using a different set of parameters and came out looking different than me. It wasn't our fault! Our parents did it! Brother and I, being human, just had to live with what abilities we were born with. Future humans, if still around, may have different options with the aid of nanotechnology and AI.

I have witnessed, either in person or by reading about it in books, virtually a miraculous transformation in human civilization. Perhaps my personal recollections would be of some value to someone, which is why I sometimes share stories here.

Hop
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
Hevans1944, you are right about the "secretiveness" of it, for all sorts of lame ordinary reasons, giving away the secrets can be end in disaster for u.

Its a battle between the worth of sharing, and then all the bad things that happen with it. Then we find out when we were going to uni, we thought we were having an education, but its more complicated than that. :)

and when we just share with each other on the internet, sometimes your not getting the help you thought you were getting from your so called "friends", for all sorts of complicated reasons.


One more thing, is if a simulator is seemingly outputing perfectly every result for all the circuits you tried on it, maybe its time to admit maybe electronics is fully simulatable today, and its probably better to sim things first instead of doing trial builds, by a long shot.

but of course, to learn from scratch both is required.
 
Last edited:

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
I just finished the gpu version in direct compute, and I have 1000 18 bit pictures stored in about 72k synapses!!!! (so extremely successful.)

Thats 72k synapses, 160 scalar inputs (which are put into combined features for the rest of the input) and 18 binary outputs. (but the outputs can also be scalar if you want to.)

But I dont know for sure if its going to "blend the results" properly for the near in the middle responses, Its only tested for the exact ones.

So I'm just putting together my framework thats going to be 4 robots doing a wrestling match to fully test it properly. But really the hardware for this thing doesn't need a single transistor to get what I need out of it!

Quite amazing. But it might not come through with the blend results, just have to wait and see, but theoretically its at least close, so maybe theres something I have to do to it to get it better, it works very similar to k nearest neighbour, but theres a reason why its actually doing a little more than that! and a very interesting thing too.
 

hevans1944

Hop - AC8NS
Jun 21, 2012
4,878
Joined
Jun 21, 2012
Messages
4,878
its probably better to sim things first instead of doing trial builds, by a long shot.
Those clever folks in Silicon Valley, California, and Texas came to the same conclusion. So they created SPICE (Simulation Program with Integrated Circuit Emphasis) software to better understand their integrated circuitry before performing the expensive tape-out required to manufacture the real circuit for sale to the public.

Back then designers used a sharp "swivel knife" to manually cut out exposure openings in rubylith masks used for photo-lithographic exposure of sensitized silicon wafers. Their simulations didn't guarantee the circuit would work after it was fabricated on the silicon, but it sure saved a lot of time and usually caught certain design errors before those errors were committed to zillions of chips. Technology has much improved since then, but errors still manage to occur. I remember a famous instance of a computer chip that gave "wrong" but nearly "right" answers to simple mathematical operations. So, if you want to simulate first and then build a trial based on the results of your simulation, that is exactly what SPICE tools are for. But the build part is always necessary IMHO.

Here is a link to a pretty good overview of how SPICE works, as well as some descriptions of its limitations.
 
Last edited:

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
Np Hevans1944. If you don't build it then it'll never hit reality for sure.

Heres the proof vid for the 1000 different records stored in 4000 synapses, (72k for 18 outs per record.)


As long as u get the resistances right, it should be the same result for real wires in the real world!!! :)
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
So if you pretrain it, the transform is just this simple and works for a motor generator for a robot, it goes straight to the inputs of the H-bridge transistors.

Theres one of these for each hinge of the robot.
perceptron1.png
 

roughshawd

Jul 13, 2020
465
Joined
Jul 13, 2020
Messages
465
Perception... Ah yes, the days when all there was on the internet to surf was a magazine index called...tada infotrak.!!
I used to spend many hours hoping I would be able to go to the library and study something...anything that required the computer! Perception was put on a chip by Intel...yes that's right the famous powerhouse that's brained everything from history to 'beep'. They had this cockamamie idea that with that kind of thinking process power they could do well anything faster.... Good things sometimes don't last very long, and as Intel backloaded all the binary stacks ISA(industrial standard application) came tumbling down. Today you can't get an ISA system that's network compatible... I have tried. I could only come up with one thing to say...PCI on you!
 

dragon

Oct 31, 2022
248
Joined
Oct 31, 2022
Messages
248
i stored 2000 records on my gpu version, with 10,000 synapses just one big braincell with one output. (10,000 inputs) these things work,and in hardware they can take a program and put it into overdrive, 1000000 fps, out of one these babies.
also one more thing, the potential of the weights isnt met by the average training system, there is a set of weights that increases the amount of patterns it can store, but its impossible to find it without a quantum computer. (intractable problem.)
 
Top