Turning Neural Networks Upside Down

[Chillheimer]

I just wished this was finally put into a nice little programm.
with resolutions higher than 640*480.

[kram1032]

There is https://deepart.io/ which takes this
https://deepart-io.s3.amazonaws.com/content/TWcZpFSXlwJszVdiKtaRyIjuC.jpg (large image)

and this (HUGE IMAGE of the famous "pillars of creation")
https://deepart-io.s3.amazonaws.com/style/MYLTKWmuHOSJZcdNfFqQljIRU.jpg

and turns it into this

but for higher resolutions you gotta pay

And if the servers aren't stalled by huge qeues, it's actually a matter of minutes to get the images.

I believe there are some open source implementations on GitHub though? Just gotta compile them... Just...

That leopard looks great! In experimenting with deepart.io I noticed very much that pictures of already somewhat similar structure or ones where you use a "style" with a lot of specific texture work best. The leopard was bound to be a great match with already similar colors. The result is quite astonishing.

[Chillheimer]

ah, good to know that there is progress.
but paying 149€ for a 3000*3000 image is seriously overpriced.
20€ for full hd would be ok, but for 1300*1300... not my kind of format.

I guess prices will fall with increasing cpu speeds and more competition.

[kram1032]

It's certainly on the pricey end

[kram1032]

Google Brain now has launched Magenta, dedicated to using deep neural networks to create art
http://magenta.tensorflow.org/
Everything will be open source and available here
https://github.com/tensorflow/magenta
and there is an official discussion site here
https://groups.google.com/a/tensorflow.org/forum/#!forum/magenta-discuss
though it's currently rather empty. Must have just been opened.
Generally, https://tensorflow.org/ could be interesting

[Tglad]

paying 149€ for a 3000*3000 image is seriously overpriced

They charge what people are willing to pay, like any business.
To me its an incredible algorithm, and it has no competition... but I'm happy to just look at the gallery.

[kram1032]

Actually I'm pretty sure almost every single step of what they are doing is available Open Source. You'd just have to put it all together and try it yourself. The largest problem you'll run into is computing power but even then, if you use some cloud service, you'll probably pay a lot less. And honestly I highly doubt their current price-point is optimal. There would probably be a whole lot more interest if they just dropped the prices and it would probably end up meaning lots more customers and overally more profit.
That being said, with each request taking a couple minutes to fulfill, they actually don't even really want that big an influx of people: It'd drive up the queue length a LOT. (During a recent competition to find a great themed combination, the waiting times went from minutes all the way to a week of waiting time) So maybe that's part of the reason.

[kram1032]

In other news, Google Magenta has composed its first piece of music:
http://thenextweb.com/google/2016/06/01/lets-talk-song-google-ai-made/
(I wasn't able to quickly find it on one of Google's pages, oddly enough)
It's not amazing, in fact the beginning is kinda boring, but it does get going after a while.

[Chillheimer]

thx for keeping us up to date kram.. all this development is mindblowing.

[Max Sinister]

Yes, suddenly things seem to go so fast. For years we only had tiny advances (feels like it anyway), but suddenly the development with Deep Learning...

[kram1032]

The technology was being worked out for a while now but it has reached a milestone of usability recently. And it will keep expanding quickly before running into its current limitations. Then there will be some stagnation again until those limits are overcome, leading into yet another expansion. That's how technology works

[Chillheimer]

And it will keep expanding quickly before running into its current limitations. Then there will be some stagnation again until those limits are overcome, leading into yet another expansion. That's how technology works

when the technique of neural learning will be used to learn how to improve the technique of neural learning... 

it's all very fascintating - but there's also quite some danger there.
for the first time in human history, we really have no way of understanding 100% of the technology we created. you can't 'reverse engineer' why the go-computer learned completely new ways that humans never thought of.
technology starts to live and evolve. and that evolution is in part beyond our influence and knowledge.

[kram1032]

I'd say that's largely a question of time: People already are working on the necessary tools to dissect exactly why NNs are so effective.
That's actually not really unusual, though it's less of a tech thing and more one of math. In math we often find lose connections we don't really get, then somebody takes a stab at it and BAM a deeply rooted connection is found and spawns an entire new field of research or three at the same time.

[Max Sinister]

In the past, NNs had the problem of being black boxes. They work, but nobody knows why. (Pretty much like the natural NNs, our brains.)

[kram1032]

They still have that but we are working on windows