/sci/ - Science & Math » Thread #12348148

3MiB, 930x628, 340e5f8e86885cfa73d332fd534de36a.gif

View Same Google iqdb SauceNAO

Anonymous Mon 16 Nov 02:24:27 2020 No.12348148 View Reply Original Report

Quoted By:

Yo, what's up. Can we get a general discussion of neural nets going? I've built one from scratch in JavaScript that can pick the highest (or lowest if you flip the pressures) number from three numbers.

The red lines are inhibitory connections, and the green lines are activation connections. The width of the lines is the weight of the connection. It has an evolutionary algorithm attached to help refine the tensors.

you can mess with the live version here.
http://ditzbitz.com/perceptronRGBg.html

Right now I'm converting it to pixel data interpretation.

Anonymous

View Same Google iqdb SauceNAO Screen Shot 2020-11-15 at 2.08.5 (...).png, 666KiB, 1320x1316

Anonymous Mon 16 Nov 2020 02:28:01 No.12348167 Report

Quoted By: >>12348168 >>12348214

Here's the visual output of one that was able to handle the largest data set with three convolution nodes with 100% accuracy for several thousand cycles.

Anonymous

View Same Google iqdb SauceNAO Screen Shot 2020-11-15 at 2.02.3 (...).png, 347KiB, 1870x1129

Anonymous Mon 16 Nov 2020 02:29:06 No.12348168 Report

Quoted By:

>>12348167
And these are the weights (always between -1 and 1)

Anonymous

Anonymous Mon 16 Nov 2020 02:29:42 No.12348170 Report

Quoted By: >>12348190

So uh, what does it do exactly?

Anonymous

Anonymous Mon 16 Nov 2020 02:32:44 No.12348180 Report

Quoted By: >>12348197

Just use gradient decent

Anonymous

Anonymous Mon 16 Nov 2020 02:33:36 No.12348190 Report

Quoted By: >>12348245

>>12348170
It multiplies the input magnitudes (normalized to between 0 and 1) by the connection weights, sums them into the magnitude of the convolution node, and multiplies them into the outputs, which are then normalized as proportions out of 1 to display the confidence of the estimation.

It basically just classifies a data set and returns its confidence that the max is what it thinks it is.

the genetic algorithm just selects the best answer every cycle and removes any that were outright wrong, then it duplicates and mutates the best one and runs again.

Anonymous

Anonymous Mon 16 Nov 2020 02:35:54 No.12348197 Report

Quoted By:

>>12348180
I'm trying to learn the basics of NNs, I'm sure I will get to that point. My overall goal right now is to build one that takes player position and momentum as 4 input nodes for instance (x,y,x-momentum,y-momentum) and learns to play the games I have made.

Anonymous

Anonymous Mon 16 Nov 2020 02:43:46 No.12348214 Report

Quoted By: >>12348323

>>12348167
Very neat
I hope to see basic nn/ml in the repertoire of any good programmer in the future, no different than understanding automata or graph theory.
Somehow have their building blocks standardized, like a regex or etc.

Anonymous

Anonymous Mon 16 Nov 2020 02:51:20 No.12348245 Report

Quoted By: >>12348323

>>12348190
That's not what I mean anon.
what does it do?

Anonymous

View Same Google iqdb SauceNAO fdbedd9b08b2bca643288fe65d9e6837.gif, 706KiB, 630x638

Anonymous Mon 16 Nov 2020 03:14:23 No.12348323 Report

Quoted By: >>12348365 >>12348494

>>12348214
Thank you, I'm tying the output to moving a dot, and feeding the coordinates back in as inputs now.
>>12348245
I'm not sure I understand, it picks the highest number using weighted multiplication.

Anonymous

Anonymous Mon 16 Nov 2020 03:24:56 No.12348365 Report

Quoted By: >>12348384

>>12348323
>I'm not sure I understand, it picks the highest number using weighted multiplication.
So it pick random numbers?

Anonymous

Anonymous Mon 16 Nov 2020 03:28:14 No.12348384 Report

Quoted By: >>12348770

>>12348365
The weights are random, yeah, but the evolutionary algorithm hones them.

Anonymous

View Same Google iqdb SauceNAO 4b208d91b791092f90ab7a08141fe541.gif, 106KiB, 218x216

Anonymous Mon 16 Nov 2020 03:57:54 No.12348494 Report

Quoted By:

>>12348323
I set the red dot up as the fitness goal, they're starting to seek it.

Anonymous

View Same Google iqdb SauceNAO 72848c6d25667d2b2d2b0406ccc8b3d5.gif, 2MiB, 638x638

Anonymous Mon 16 Nov 2020 05:07:44 No.12348750 Report

Quoted By:

A few thousand generations and wham! this one can find the dot where I put it.

Anonymous

Anonymous Mon 16 Nov 2020 05:12:46 No.12348770 Report

Quoted By: >>12348777

>>12348384
So it guesses the number of balls?
based on random numbers?

Anonymous

Anonymous Mon 16 Nov 2020 05:14:53 No.12348777 Report

Quoted By: >>12348833

>>12348770
It guesses which color ball is most common based on random numbers yes. But they stop being 'random' really quick, as the evolution of the population hones it toward a proper answer.

Anonymous

Anonymous Mon 16 Nov 2020 05:31:57 No.12348833 Report

Quoted By: >>12348848

>>12348777
So it compares the random numbers to the correct numbers then resifts the weights until the number is a certain degree of accuracy?

Anonymous

Anonymous Mon 16 Nov 2020 05:35:57 No.12348848 Report

Quoted By: >>12348928

>>12348833
Not quite, it multiplies the inputs by random numbers in what is effectively a matrix, the net result is a collapsed 1x3 matrix of values that sum to 1. The highest of these values is its 'guess'. When it is wrong, a mutated version of one that was right takes its place until the population can be right 100% of the time.

Anonymous

Anonymous Mon 16 Nov 2020 05:56:26 No.12348928 Report

Quoted By: >>12348952

>>12348848
So it's a guessing machine?
Couldn't you get the same result with a simpler algorithm?
If I'm guess a number and I'm told a percent error I can find that number no matter what extremely trivially and extremely fast.
say theres a number lie from zero to infinity, the correct guess is somewhere there.
each time I find the error.
a rational way to measure is in counts of 100.
say I pick 100, it's wrong, I can go left or right. If I do a step to the left and it's closer to being correct then I know it's between the bounds of 0-100, otherwise I've eliminated 0-100.
Then I can just do some bijections to find the right values.
because that will close my bounds very quickly. I can go to 50 and if it's not 50 then it's between 0-50 or 50-100, I can test which direction brings me closer and eliminate 50 choices. then I can continue until I've gotten a very accurate answer.
Maybe there's a better searching algorithm but this will take much less than 100 steps.
So I'm curious what makes what you did better?

Anonymous

View Same Google iqdb SauceNAO 30b118f3ddd233f0f4bdb165e8a663db.gif, 2MiB, 700x698

Anonymous Mon 16 Nov 2020 06:06:45 No.12348952 Report

Quoted By: >>12349000

>>12348928
>So I'm curious what makes what you did better?
Well what you described is so different from what it does its hard to compare. I'm not trying to specifically do what it does, if I wanted to know the answer I would just use greater than operators. What I consider interesting about it is that it can solve the problem without doing any of that.

Moving past the number classification makes things clearer I think. The outputs in the new version are proportions of a vector of acceleration. It takes the x and y position of the target (the red dot) and itself every frame, and using that data it creates the proportional response to get as close as it can to the target in the time limit. Red lighting up means 'go more left', green is 'go more right', blue is 'go more up', and yellow is 'go more down'.

Anonymous

Anonymous Mon 16 Nov 2020 06:27:09 No.12349000 Report

Quoted By: >>12349066

>>12348952
This is probably your worst explanation so far.
So if it's not guessing numbers then what is it doing.
how are you not getting this.
what does it fucking do?

Anonymous

View Same Google iqdb SauceNAO NN code.png, 849KiB, 2160x3408

Anonymous Mon 16 Nov 2020 06:47:57 No.12349066 Report

Quoted By:

>>12349000
I'm sorry I'm so dense lol. Nice get also. Here is a truncated version of the code without the draw/clone/mutate functions. I think this will hopefully clear up what it is doing.

In the function named 'compute' it receives proportional data, at this point a simple comparison would give the correct answer. It doesn't have access to that. Instead it must compare the signals by weighted multiplication. The weighted multiplication returns signals. It doesn't try to count the dots, it doesn't try to say how many there are, it just tries to match its outputs to the inputs, or in the case that it is selecting the lowest number, it tries to match its outputs to the inverse of the inputs.

Anonymous

View Same Google iqdb SauceNAO 6bbe79f69d92f001910de0c1455219e7.gif, 2MiB, 700x698

Anonymous Mon 16 Nov 2020 10:08:41 No.12349492 Report

Quoted By:

The version where it learns to steer a dot to another dot is now live.

http://ditzbitz.com/perceptronDriverComplex.html

Anonymous

Anonymous Mon 16 Nov 2020 10:25:55 No.12349513 Report

Quoted By: >>12349518

I do not get which problem you are trying to solve
or why a neural network is a good approach for solving it

Anonymous

Anonymous Mon 16 Nov 2020 10:27:20 No.12349518 Report

Quoted By: >>12349522 >>12349657

>>12349513
I'm trying to do this:
https://www.youtube.com/watch?v=qv6UVOQ0F44
but with games I made. So the first step is solving really simple problems with a neural net, then moving onto gameplay.

Anonymous

Anonymous Mon 16 Nov 2020 10:28:19 No.12349522 Report

Quoted By: >>12349529

>>12349518
so basically
>draw two circles
>draw the rest of the owl
?

Anonymous

Anonymous Mon 16 Nov 2020 10:30:23 No.12349529 Report

Quoted By:

>>12349522
The first game I was planning to start with is very simple. I don't think it will be too extreme of a leap.
http://ditzbitz.com/blastyballs.html

Anonymous

Anonymous Mon 16 Nov 2020 12:16:10 No.12349657 Report

Quoted By: >>12349698

>>12349518
Evolution is blind, but it generates a brain that isn't. This approach doesn't generate such a brain, it generates an expert system that takes time as its input. It isn't very useful in the real world.

Anonymous

View Same Google iqdb SauceNAO d25eee8dc74b6011c91c0bbd47ea2bf1.gif, 86KiB, 222x214

Anonymous Mon 16 Nov 2020 12:38:12 No.12349698 Report

Quoted By:

>>12349657
>It isn't very useful in the real world.
Granted, sure. It's a lot of fun to work on though, and I have my uses for it.

I got it to go around an obstacle.

Capcode	All Only User Posts Only Moderator Posts Only Admin Posts Only Developer Posts
Show Posts	All Only With Images Only Without Images
Deleted Posts	All Only Deleted Posts Only Non-Deleted Posts
Ghost Posts	All Only Ghost Posts Only Non-Ghost Posts
Post Type	All Only Sticky Threads Only Opening Posts Only Reply Posts
Results	All Grouped By Threads
Order	Latest Posts First Oldest Posts First

Your latest searches