/sci/ - Science & Math » Thread #12395026

729KiB, 902x396, 409f7311597f03266e6395d0d487dff5.gif

View Same Google iqdb SauceNAO

Anonymous Sat 28 Nov 08:09:52 2020 No.12395026 View Reply Original Report

Quoted By: >>12395041 >>12395050 >>12395373 >>12396401 >>12396419

I made a neural network that can (sometimes) read handwritten numbers.
I made a video demo'ing it too.
https://youtu.be/MWEIU0HAfQE

You can try it here
http://ditzbitz.com/numberguess.html

Anonymous

Anonymous Sat 28 Nov 2020 08:10:55 No.12395029 Report

Quoted By: >>12395036

Neat cool stuff

Anonymous

Anonymous Sat 28 Nov 2020 08:16:08 No.12395036 Report

Quoted By: >>12395345

>>12395029
Thank you! I got filtered hard at backpropagation, but by weighting it manually I was able to achieve a 84% accuracy on the MNIST hand written numbers data.

Anonymous

Anonymous Sat 28 Nov 2020 08:17:10 No.12395039 Report

Quoted By: >>12395048

Nice work, but for what purpose is that? I've also seen application like that but never seen a program that would fully transform a written word to a digital text.
https://youtu.be/ziXJYCO37ZA

I'm old. I have a small grasp for technology.

Anonymous

View Same Google iqdb SauceNAO file.png, 532KiB, 500x450

Anonymous Sat 28 Nov 2020 08:19:09 No.12395041 Report

Quoted By: >>12395048

>>12395026
literally professor layton games do this when u input numbers u can write them in

Anonymous

View Same Google iqdb SauceNAO 16ae3ab30c4875b79e3977c7b61180a6 (...).gif, 4MiB, 702x690

Anonymous Sat 28 Nov 2020 08:23:37 No.12395048 Report

Quoted By: >>12395053

>>12395041
Well sure, it's considered the 'hello world' of neural networks. There are a lot of resources for doing this specific type of thing.
>>12395039
Just for fun, I might make a sudoku game or something out of it. I wanted to get a good feel for how neural networks work, and by making this I feel pretty confident that I could explain it to a kid or something.

Here's the visualization of one that evolved naturally when I was using a genetic algorithm to generate them, this one got to 70% accuracy, it has 16 filters and 16 hidden nodes, so this visualization is pretty much impossible to interpret.

Anonymous

View Same Google iqdb SauceNAO Numberwriting.jpg, 47KiB, 1128x408

Anonymous Sat 28 Nov 2020 08:25:38 No.12395050 Report

Quoted By: >>12395052

>>12395026
Might be helpfull to automatically center the drawing. I can get it to guess wrong or right by slightly moving my drawing left or right.

Anonymous

Anonymous Sat 28 Nov 2020 08:27:54 No.12395052 Report

Quoted By:

>>12395050
I did consider that, but the pixel data is a one-dimensional array, so getting the centered version of the image is a bit trickier than I expected. It's a good idea, just hard to implement with the way it was built in the first place, and mostly an oversight of how it would end up working.

Anonymous

Anonymous Sat 28 Nov 2020 08:28:17 No.12395053 Report

Quoted By:

>>12395048
Well goodluck with that OP, good to know that you're having fun.

Anonymous

Anonymous Sat 28 Nov 2020 10:59:06 No.12395345 Report

Quoted By: >>12395454

>>12395036
>84% accuracy on the MNIST
I wrote a dead simple neural network for a class with no regularization or adaptive learning rates using SOFTMAX and ReLU that scored 97% on MNIST.

Your method might be neat, but it is worthless against Neural nets. If you're having trouble understanding backprop, i'd reccomend Nielsens notes (http://neuralnetworksanddeeplearning.com/chap2.html). Also, translating this "simple" approach into a vectorized version that can run multiple minibatches at once using matrix operations really helps with drilling it in.

Anonymous

Anonymous Sat 28 Nov 2020 11:28:11 No.12395373 Report

Quoted By: >>12395454

>>12395026
>Neural network
>Without back propagation

Why would you even

Anonymous

Anonymous Sat 28 Nov 2020 12:13:45 No.12395454 Report

Quoted By:

>>12395373
>>12395345
I had something close to an implementation of backprop, but I just haven't got it working quite right. Thanks for the link.

Anonymous

View Same Google iqdb SauceNAO KueckenGelb.jpg, 105KiB, 1680x1050

Anonymous Sat 28 Nov 2020 13:00:58 No.12395543 Report

Quoted By: >>12396361

I also wrote a network for MNIST, in standard library python, roughly following the book by iirc Nelson. In fact I coded it up live (gist in the description box)

https://youtu.be/z2aq21lMw40

Anonymous

Anonymous Sat 28 Nov 2020 19:01:19 No.12396361 Report

Quoted By:

>>12395543
wow nice! I also like your shirt.

Anonymous

View Same Google iqdb SauceNAO basd.png, 39KiB, 1023x394

Anonymous Sat 28 Nov 2020 19:20:41 No.12396401 Report

Quoted By:

>>12395026
based

OP Sat 28 Nov 2020 19:25:23 No.12396419 Report

Quoted By:

>>12395026
nice fren, have a bump

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