[cosmie]

The demo given is also pretty interesting: http://harthur.github.com/brain/

[Kyek]

That's bad ass . About 60 seconds and I had a neural net that could determine a better font color for a given background than the traditional algorithm. It's just so damn fast! I'd love to see how quick it is to train the net for image recognition.

[soulcyon]

Neural nets make me *gnoz*

I'm working on a project that requires generating humongous trees for cascading failures, perhaps neural nets will speed up this process. I'll have to do some reading

[cosmie]

Kyek, on 06 July 2012 - 11:19 AM, said:

That's bad ass . About 60 seconds and I had a neural net that could determine a better font color for a given background than the traditional algorithm. It's just so damn fast! I'd love to see how quick it is to train the net for image recognition.

Have you seen his kitty radar library, that identifies cat heads in pics? https://github.com/harthur/kittydar
I'm not sure if it's using the same code as brain.js for the training, but I'm assuming it's at least really similar, as it's also using a neural net and is only a few weeks old.

That only took 2GB worth of annotated cat pics to train (a few thousand images, in total). He doesn't mention how slow (or fast) the training was, but with something like AWS you could probably get it done in no time at all for a relatively low cost. The curating of the imageset was probably a killer, time wise. But that was done by a Microsoft research group, anyway.

Throw in a few dozen GBs of e-commerce page-tracking data and you could probably make one creepy backend brain to curate targeted content to show to consumers.

[callumacrae]

It's like 500 lines long in total 0_o

[Kyek]

I so should not be taking the time to do this.



For some reason I'm having a rough time getting training to work... and I'm afraid the library wasn't built to handle 625 inputs. I may scale my grid down a bit :-/

Edit: Ahahahahaha scratch that, I got it working. Outputs have to be weighted -- so you can't just say an array of pixels is an "eye", you have to tell it it's

{eye: 1}

. This is a little on the hairy side, though, because for results, it spits out all the possible results and assigns a match quality to each one -- a number from 0-1. If your neural net knows what a lot of different things are, that means it's spitting out the entire dictionary every time you get results. That's a little ugly.

But this still rocks xD

Edit 2: Maybe I spoke too soon... I trained 'fish' and 'eye', then went back and drew a fish. It told me there was a 0.01% chance it was an eye. That's all well and good, but it didn't tell me what the chances are of it being a fish! If this can only report the weight of a single value.. boo.

Edit 3: Should've RTFM:

Quote

Use train() to train the network with an array of training data. The network has to be trained with all the data in bulk in one call to train().

That's incredibly shitty. That means a net can't be improved after it's initial training... making a neural net that gets smarter over time basically impossible without a lot of horrible re-training. Not a fan :-/

Edit 4: Ok, one last update before I go to bed. It now recognizes my drawings fairly well ;-). I tweaked it so it stores all the training data over time, and re-trains all of it every time it's given a new image. Not optimal, but it works and it's fun to play with.









Not bad for a couple hours of hacking ;-)

[ianonavy]

@Kyek

GitHub linky link? D: Thinking about porting the library to Java for use on my Minecraft server. I have no idea what it'd be for though...

[TheMaster]

ianonavy, on 06 July 2012 - 11:53 PM, said:

@Kyek

GitHub linky link? D: Thinking about porting the library to Java for use on my Minecraft server. I have no idea what it'd be for though...

Oh my. That could be epic. You could like...build something (say...a small house), it would recognise its a house - and then improve it for you - based on its understanding of a house!! (No idea why you'd want it to do that ).

-----------

This is really cool! I wish there was a PHP version that stored the neural net in a database or something.

[Quinn]

That is a dog.

[callumacrae]

Something like this wouldn't work too well in PHP, it would use far too much (many?) resources and be pretty slow.

[TheMaster]

callumacrae, on 07 July 2012 - 05:24 AM, said:

Something like this wouldn't work too well in PHP, it would use far too much (many?) resources and be pretty slow.

Is Node.js faster/less resource hungry than Node.js? PHP? Didn't know that....

EDIT: Ooops. Meant PHP

[Quinn]

TheMaster, on 07 July 2012 - 07:10 AM, said:

Is Node faster/less resource hungry than Node.js? Didn't know that....

Node is Node.JS

[TheMaster]

Quinn, on 07 July 2012 - 07:21 AM, said:

Node is Node.JS

See correction

[Kyek]

ianonavy, on 06 July 2012 - 11:53 PM, said:

@Kyek

GitHub linky link? D: Thinking about porting the library to Java for use on my Minecraft server. I have no idea what it'd be for though...

https://github.com/TomFrost/MindEye -- Although the neural net library itself is linked in my first post

I'm not sure how much help this would be in Minecraft though -- play with my sample a bit so you get the feeling for how much this has to be trained in order to be accurate, and how much a little variation can throw it off. Not very Mindcraft-friendly I'm afraid.

TheMaster, on 07 July 2012 - 05:05 AM, said:

Oh my. That could be epic. You could like...build something (say...a small house), it would recognise its a house - and then improve it for you - based on its understanding of a house!! (No idea why you'd want it to do that ).

Neural nets don't really work that way It could maybe tell you it's a house based on the shape, but suggesting additions would require a whole new algorithm.

Quote

This is really cool! I wish there was a PHP version that stored the neural net in a database or something.

There are quite a few libraries and guides for PHP.

Quinn, on 07 July 2012 - 05:24 AM, said:

That is a dog.

That's why it's only 42% correct

callumacrae, on 07 July 2012 - 05:24 AM, said:

Something like this wouldn't work too well in PHP, it would use far too much (many?) resources and be pretty slow.

Slower for big ones maybe, but it's more about your implementation. Mine is fully blocking and gets slower with every training, but if I were launching this in production I'd just spin off a child process to queue up training jobs, and a small cluster of child processes to run neural net queries. While PHP is downright painful for dealing with child processes and such, it could technically be done and still be fast. It's just a very poor language choice for something like this.

TheMaster, on 07 July 2012 - 07:10 AM, said:

Is Node.js faster/less resource hungry than Node.js? PHP? Didn't know that....

EDIT: Ooops. Meant PHP

Node compiles down to machine code at run time, and uses a single thread. PHP, Java, Python, Ruby, etc etc etc all compile down to opcode which has to be run through a VM to parse it every time something needs to be executed, and generally there's either a bunch of processes clogging up the CPU or a bunch of threads. They can run fast, and maybe in a few rare cases faster than Node.js, but if you're writing an app that Node is suited for then it tends to be very snappy.

[dida]

Kyek, on 07 July 2012 - 05:24 AM, said:

Uh, wtf?
0,42% not 42% . 0,42% is 0,42/100 ... report it as a bug or something.

[soulcyon]

Haha yeah, it says 0.42% when its supposed to be 42% or 0.42 or 42/100

[Quinn]

dida, on 07 July 2012 - 02:15 PM, said:

Uh, wtf?
0,42% not 42% . 0,42% is 0,42/100 ... report it as a bug or something.

Report it as a bug to..himself?

[Kyek]

Ha, yeah, I noticed that right after I took the screenshots, but it wasn't worth fixing. Just forgot the * 100 is all. I'm about to push a version to github with that fix, as well as non-blocking training so it can be launched on a server with multiple users and still function. It saves the training data every time now too, so restarts don't blow your knowledge base away.

[Kyek]

Annnnd pushed. Enjoy, if anyone's playing with it

Edit: Ok, one more bump, but with good reason -- I launched this on the wdR server for fun ;-). I'll be taking it down when I'm not nearby to monitor it (ie when I'm asleep) but for now it's rolling.

http://webdevrefinery.com:8888

Edit 2: I just looked at our word list and was very impressed that the word 'penis' was not there. Good on you, random wdR people ;-)