Google's Dreaming Artificial Neural Net

Expand the album for full descriptions

Expand the album for full descriptions

Google has an artificial neural net that dreams... Let's just let that fact settle in for a minute. Roll it around in your mind and let it find a home. A computer network is having dreams.

It's been presumed that higher animals have dreams, so really, why not computers? Google's artificial neural network is trained to learn what images of a given item look like, the example in the article is a banana. The surprising fact is that not only can the neural net discern what a banana looks like, but because of training the net to understand what a banana looks like, it can also create an image of a banana if it looks at random noise, and tries to 'find' a banana.

http://goo.gl/0stTGE

*edit* something odd is happening with this album hope this fixes it, thanks for sticking with it while I run around with my hair on fire. https://www.youtube.com/watch?v=t3otBjVZzT0

**edit #2** after frigging around with this for about an hour, I'm just adding this image and the description from the first picture. Hopefully that fixes it, if not, help me, Obi @sarah Kenobi, you're my only hope.

**edit 3** I've had to find, edit and restore this description too damned often. https://i.imgur.com/u36ZyJI.gif

Expand the album for full descriptions

The technique is called visual priming and recall, and it can have some very strange results.

The article shows specifically this example of a dumb-bell, the dumb-bell noted really looks nothing like a dumb-bell and has an odd flesh-coloured appendage. The easiest inference is that when the computer attempted to recall what a dumb-bell looked like, it had never seen one without an arm curling it.

This is likely the best way to understand that most of these techniques require a lot of refinement, and is proof of unintended consequences. You can prime a neural net with bad information, and it will act on that information.

There's a lot of discussion as to how this works, and if it can apply to humans, but that's a whole other can of worms for another day.

Expand the album for full descriptions

Inceptionism process of convolutions of Seurat's 'A Sunday Afternoon on the Island of La Grande Jatte.'

An interesting parallel of this is actually Seurat's technique (pointillism) of using dots to create a multi-point raster using varying colours. The neural net may be expounding on this technique unwittingly.

Note how it loses and gains fidelity to the original image.

Comparison of exemplar and convolution of impalas

In some cases, though, Google researchers let the neural net off the chain and gave it an exemplar, and just requested similar items back.

From the article: "This creates a feedback loop: if a cloud looks a little bit like a bird, the network will make it look more like a bird. This in turn will make the network recognize the bird even more strongly on the next pass and so forth, until a highly detailed bird appears, seemingly out of nowhere."

The animals (I'm guessing at impalas) are a useful visualization of the feedback loop, or convolution. ( Convolution: http://arxiv.org/pdf/1409.4842.pdf )

The overall process is called inceptionism because it can create other things that don't look like this, seemingly out of the ether.

Convolution - Seattle Space Needle

You're probably wondering about the correlation between dreams and the convolutions/inceptionism process...

Interestingly, this follows some speculation on how pattern and memory processing in the human brain works. The idea is that the brain uses REM/NREM sleep to do post-processing of information.

It's anecdotal evidence, but if you need something to prove to you that it's important to not miss sleep, well... Sorry... where was I?

Convolution of Munch's 'The Scream'

The exemplar was Munch 'The Scream' - https://goo.gl/Fxd8uR

Again, we're seeing primary convolutions. The neural net starts by trying to filter the image, it looks to me like it's trying to set eyes in the exemplar because it recognized the shape of a humanoid...

Further convolution of Munch's 'The Scream'

And now, it's making things up as it goes. I guess the neural net has an affinity for dogs. This is proof that the neural net is NOT in fact Ultron, as it doesn't have any pictures of cats that I've seen.

Also, dog-tax.

Convolution of random noise picture.

The rest of the really strange images that don't bear much of a relationship to anything in the world are the result of the neural net's process of inceptionism using exemplars of random noise and predefined pictures.

This is basically Google engineers letting the neural net loose with all it's learned and formed priming to go and return more images that look like the random noise, and then retrieving what pattern it thinks the engineers are looking for.

They vary from very, very early convolutions (like this one, it's on about the 3rd/4th convolution) to very late in the process.

It's not noted if these all used the same random noise picture as exemplar, or if different ones were used. Ideally though, it shouldn't matter.

Convolution

Exemplar wasn't noted.

Also, dog tax.

Exemplar

Convolution

Exemplar

Early Convolution

Further convolution

Exemplar

Convolution

Convolution of random noise picture.

More images: https://goo.gl/VSwBBH

Convolution of random noise picture.

Convolution of random noise picture.

Convolution of random noise picture.

Convolution of random noise picture.

Convolution of random noise picture.

Convolution of random noise picture.

Convolution of random noise picture.