Turning Neural Networks Upside Down

[kram1032]

Youhn that's a problem this AI isn't equipped to solve, I'm pretty sure. Gotta have an AI that actually understands composition. Those are in the works too (and I even linked to an example of one such effort earlier in this thread)

KRAFTWERK I guess if we really want it to know more images, we'd have to train our own. The full AI is available, right? You *could* technically have it learn what ever imagery you like.
But that's insanely much work. Google uses huge datasets and trains it for 1000 words. More words would require more outputs or, perhaps, a different approach altogether.

[youhn]

Just train the AI with a game. Guess the picture. Show a details, and it has to guess the context. I admit it can be pretty hard, even for humans. But seeing all kinds of things in a piece of bush is a kinda fail. Or the AI has indeed taken some psychoactive drugs;



Source: http://psychic-vr-lab.com/deepdream/pic.php?d_serial=3318

[3dickulus]

seems more "A" than "I" as in...
"this is the closest thing in my database that resembles this bit of the picture so just replace and blend that bit"
I know it's not that simple, that statement is an assumption based on the above result.

KRAFTWERK I guess if we really want it to know more images, we'd have to train our own. The full AI is available, right? You *could* technically have it learn what ever imagery you like.

Q: how many images of the mandelbrot set are represented here on FF, include all stills and every frame from every zoom/pan/morph video?
me thinks this is a reasonably large number

[cKleinhuis]

lol, if you want to know: more than 18.000 pictures in the gallery

[3dickulus]

...and how many minutes of video? I know, it's a bit of a rhetorical question but it's also an interesting piece of trivia.
...add this to the youtube vids and it should be quite substantial.

would this be enough to "train" a net?

[KRAFTWERK]

Just train the AI with a game. Guess the picture. Show a details, and it has to guess the context. I admit it can be pretty hard, even for humans. But seeing all kinds of things in a piece of bush is a kinda fail. Or the AI has indeed taken some psychoactive drugs;

LOL it is not sane, that's for sure  

I guess it has to do with this:

"We aren't actually asking the system what it thinks the image is, we're extracting the image from somewhere inside the network. From any one of the layers. Since different layers store different levels of abstraction and detail, picking different layers to generate the 'internal picture' hi-lights different features."

(taken from the description of the video here: vimeo.com/132462576 )

[kram1032]

The problem with training may not be the data set but the amount of time it takes on normal hardware. As far as I know Google does this stuff on a Google-sized computing farm.
Though even with the data set you gotta be careful to do it right.

And for context you'll have to:
- either focus on a small number of possible contexts
- or really increase the network size and data set so it actually learns relationships between objects and perhaps even pop culture.

There's a reason why this kind of task is even hard for humans and why humans in different situations will give vastly different answers.
A "guess the context" game is more likely to tell you about the person you ask than about the context in the image. It's more of a psychological state of mind test than an intelligence one.

And the AI sees all those things in the hedge because that's literally all it knows. It doesn't know hedges. It knows dogs. (In fact a rather large variety of dog races. The training data happened to emphasize that)

What this AI also doesn't have is a hierarchy of categories. For instance, while it does know a dalmatian from a pug, It'd have no clue at all that both of those things happen to be dogs.

[eiffie]

If you are really asking the AI to create art on its own then you would need LOTS of training on objects and their relationships but as a tool - like fractals - that an artist uses then it seems a small set of training images (with the usual rotations, scaling) would suffice.

[cKleinhuis]

<a href="https://www.youtube.com/v/oyxSerkkP4o&rel=1&fs=1&hd=1" target="_blank">https://www.youtube.com/v/oyxSerkkP4o&rel=1&fs=1&hd=1</a>

[kram1032]

Nice. This guy did other videos as well: https://www.youtube.com/channel/UCOfLZwshIblObgtAYOCD7BA
There's one of 2001: A Space Odyssey
<a href="https://www.youtube.com/v/tbTJH8aPl60&rel=1&fs=1&hd=1" target="_blank">https://www.youtube.com/v/tbTJH8aPl60&rel=1&fs=1&hd=1</a>
and another one which just shows off the various layers of a different network trained on landscape, using The Scream as a case study.
<a href="https://www.youtube.com/v/6IgbMiEaFRY&rel=1&fs=1&hd=1" target="_blank">https://www.youtube.com/v/6IgbMiEaFRY&rel=1&fs=1&hd=1</a>
And probably more to come.

[Syntopia]

Got CUDA working, which means generating new images new is much faster. Unfortunately, 2GB of ram on my GPU is not enough for bigger images.

If you want to ponder on the personality of these nets, here are some Rorschach-tests:

[Syntopia]

And:

[3dickulus]

@Syntopia brilliant! Rorschach-tests hmmm how about some of those weird color blindness test images ?

[M Benesi]

Check out this page.  Couple of good links, along with a nice explanation of some realtime computer tripping.

  Actually, here is a direct link to the explanation video.  I like it.

<a href="https://www.youtube.com/v/yTvOoMCGlAc&rel=1&fs=1&hd=1" target="_blank">https://www.youtube.com/v/yTvOoMCGlAc&rel=1&fs=1&hd=1</a>

http://artofericwayne.com/2015/07/08/google-deep-dream-getting-too-good/

[Syntopia]

Color blindness tests proved surprisingly robust to inceptionism. Here is an example, but I had to crank up the step size: