It reminds me of that parrot image that was said to crash human brains, only even more intense. I certainly experienced some effect, as while looking at it and trying to figure out what exactly it was, I felt my head heating up --- probably increased blood flow.
That image was so striking and appeared to come out of nowhere. Was it some kind of marketing ploy do you think? I'm glad to have found the source anyway.
>If we choose higher-level layers, which identify more sophisticated features in images, complex features or even whole objects tend to emerge. Again, we just start with an existing image and give it to our neural net. We ask the network: “Whatever you see there, I want more of it!” 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.
I would say that's a pretty good description of how many psychedelic hallucinations unfold. They start off as noise in your vision which turn into loose forms which turn into geometric patterns which turn into etc...
Maybe there's something to do with how our brains interpret information differently when under the influence of psychoactive drugs.
I've been looking at Aldous Huxley's "Doors of Perception" and other psychonautic works recently and he hypothesizes that these sorts of drugs filter out the usual signals from the CNS that shut out the parts of perception that are not important for you to receive for survival.
It might be some great leap of armchair psychology, but I think we're due for another psychedelic revival, especially considering the new advances in synthetic psychedelics, legalization of more harmless recreational drugs, new tests in medical research using MDMA/LSD/Psilocybin, and the cultural shift away from the 'War on drugs'.
Had me sitting down. Felt mesmerizing, like a weird resonnance with my mind. This is how I imagined my brain working, patching bits of stimulus to recreate complex shapes fractally... Seeing it in pictures is ... just amazing.
That's a really worthwhile question, actually, considering the degree to which the human form and stylished portrayals of sexuality underpin so much of art and design.
Am I the only person who is not entirely happy about the overuse of the pop-culture term 'inception' for everything that is remotely nested, recursive or strange-loop-like?
In this paper, we will focus on an efficient deep neural network
architecture for computer vision, codenamed Inception, which derives
its name from the Network in network paper by Lin et al [12]
in conjunction with the famous “we need to go deeper” internet meme [1]
I'll repeat what I posted on facebook because I thought it was clever: "Yes, but only if we tell them to dream about electric sheep."
So, tell the machine to think about bananas, and it will conjure up a mental image of bananas. Tell it to imagine a fish-dog and it'll do its best. What happens if/when we have enough storage to supply it a 24/7 video feed (aka eyes), give a robot some navigational logic (or strap it to someone's head), and give it the ability to ask questions, say, below some confidence interval (and us the ability to supply it answers)? What would this represent? What would come out on the other side? A fraction of a human being? Or perhaps just an artificial representation of "the human experience".
Neural networks are a relatively simple mathematical model. They don't actually "think" or have a conscience. Neural networks are also regularly fed books, in order to model some properties of natural language.
Neurons are also relatively simple, at least in comparison to the mind. I don't think the simplicity or complexity of the underlying model has much bearing on the higher-level properties of the network.
Now, this isn't to say that the kinds of neural networks we build today are conscious, but I don't think that's because they're based on a simple mathematical model; I think that's because they don't have the network-level properties that conscious humans do, for example, a self-representation.
It would have some kind of intelligence, at least able to recall information and form associations between things. But there's no reason to think that it would come out looking human. I mean you can show a dog lots and lots of images and it doesn't turn human.
I never had much luck with generative networks. I did some work putting RBMs on a GPU partly because I'd seen Hinton talk showing starting with a low level description and feeding it forwards, but always ended up with highly unstable networks myself.
Neural networks are notoriously difficult to train due to the large number of hyper-parameters that need to be tuned. If your network never converged, it's possible your learning rate was too high, so it kept overshooting the minima of the loss function.
Quite possibly. The classification results were great, just wasn't good when trying to run things back through the network repeatedly. I did have issues that the learning rates reported in some of the original papers didn't match the ones in the released code.
Really cool. You could generate all kinds of interesting art with this.
I can't help but think of people who report seeing faces in their toast. Humans are biased towards seeing faces in randomness. A neural network trained on millions of puppy pictures will see dogs in clouds.
That's essentially precisely what's happening here. You can see in the different pictures where different sets of training data were used---buildings, faces, animals.
Give the machine millions of reference images to work from and then tell it to find those images in noise, and it will succeed (because it literally can't "imagine" anything else for the noise to be).
I'm blown away by this "guided hallucination" technique. It's not a big oversimplification to describe to the layperson as: enter images into neural network; receive as output artwork representing the essence of the images.
This may sound ridiculous, but I think this has the potential to be a development as foundation-shaking as Modernism itself. There has been plenty of algorithmically-derived art over the past 30 years, but generative pieces inevitably look like math – they are interesting curiosities, sometimes quite beautiful, but they don’t challenge the mind like any of the major movements of the past 150 years.
This is different because, while still just math, it’s modeled on the processes of human perception. And when successfully executed, it plays on human perception in ways that were formerly the exclusive domain of humans – Chagall, DiChirico, Picasso – gifted with some sort of insight into that perception.
Future iterations of this kind of processing, with even higher-order symbol management could get really weird, really fast.
I felt the same. I think the main aspect about these images that makes me like them is how everything feels connected, which, is what the AI is trying to find: connections. Honestly, can anyone tell me where I could order large prints of some of these?
Agreed. These are just amazing. Someone linked above to the source images on Google Photos, but even those aren't especially high-res. Would be awesome if Google released the originals.
The reason they look so 'fractal-like' (e.g. trippy!) is because they actually are fractals!
In the same way a normal fractal is a recursive application of some drawing function, this is a recursive application of different generation or "recognition -> generation" drawing functions built on top of the CNN.
So I believe that, given a random noise image, these networks don't generate the crazy trippy fractal patterns directly. Instead, that happens by feeding the generated image back to the network over and over again (with e.g. zooming in between).
Think of it a bit like a Rorschach test. But instead of ink blots, we'd use random noise and an artificial neural network. And instead of switching to the next Rorschach card after someone thinks they see a pattern, you continuously move the ink blot around until it looks more and more like the image the person thinks they see.
But because we're dealing with ink, and we're just randomly scattering it around, you'd start to see more and more of your original guess, or other recognized patterns, throughout the different parts of the scattered ink. Repeat this over and over again and you have these amazing fractals!
The trippiness is further compounded by the rainbow-ish colour effect produced by the recursive function, which mimics the "shimmering" rainbow effect you commonly get around lights when tripping on LSD.
And also, when under the influence of various drugs you tend to see patterns, particularly faces, where there aren't any.
I believe they do (in the sense that if you take one of these images, zoom it in, and run it through the algorithm again, it'll take the micro-features of the animals it hallucinated and hallucinate more animals on top of them).
Functional iteration is actually a fun way to draw fractal images in the plane. It goes like this (for anyone interested):
1) Pick a function f: R^2 ==> R^2
2) Pick a region of R^2 (this could be the unit square for instance).
3) For each point in the region do the following:
a) Plug the point into f. Then plug f(x) into f. Then plug f(f(x)) into f, etc....
b) The norm of f(f(...f(x)...)) will either run off to infinite or stay bounded.
c) Record for the original point, x, how many iterations it took the process to run off to infinite (or the maximum if the sequence stayed bounded).
4) Paint by number after assigning a unique color to each possible number of iterations.
Here's the result of this process for the function:
> The reason they look so 'fractal-like' (e.g. trippy!) is because they actually are fractals!
While I agree with your idea about fractals (though you're a bit vague on the math details to know for sure), I also believe that a large reason the images look so "trippy" is because there is some local contrasting effect at work, generating high-saturation rainbow fringes at the edges of details and features. You get loads of that on psychedelics as well.
I bet there's a pretty straightforward reason to explain these rainbow fringes, if one were to dig into it, though.
Another (unrelated) observation I had was the feeling that the neural net seemed to be reproducing JPEG-artifact type fringes in the images? Though it could be that I was just looking at scaled versions of already JPEG-compressed output images, the article doesn't provide details (if only they had been PNGs ...).
Really nice. I'd be interested in seeing a more in-depth scientific description of how these images were actually generated. Are there any other publications related to this work?
The code for the 1st is available in a Gist linked from its comments; the creator of the 2nd has a few other videos animating grid 'fantasies' of digit-recognition neural-nets.
These images are remarkably similar to chemically-enhanced mammalian neural processing in both form and content. I feel comfortable saying that this is the Real Deal and Google has made a scientifically and historically significant discovery here. I'm also getting an intense burst of nostalgia.
One from Vincent Vanhoucke: "This is the most fun we've had in the office in a while. We've even made some of those 'Inceptionistic' art pieces into giant posters. Beyond the eye candy, there is actually something deeply interesting in this line of work: neural networks have a bad reputation for being strange black boxes that that are opaque to inspection. I have never understood those charges: any other model (GMM, SVM, Random Forests) of any sufficient complexity for a real task is completely opaque for very fundamental reasons: their non-linear structure makes it hard to project back the function they represent into their input space and make sense of it. Not so with backprop, as this blog post shows eloquently: you can query the model and ask what it believes it is seeing or 'wants' to see simply by following gradients. This 'guided hallucination' technique is very powerful and the gorgeous visualizations it generates are very evocative of what's really going on in the network."
That's not really fair though, since any deterministic function can be "back-propagated" using the chain rule (or even automatic differentiation), even though it's not really necessary for simpler models such as GMM and SVM since there are much easier ways of inspecting them. Also, I don't feel single input/output pairs really describe the function itself -- knowing cos(0) = 1 doesn't reveal much about the cosine function, even though it's a local maximum. Maybe one could extend the technique to show transitions (morphing) between classes as video?
Looks like they didn't intend to release it to the public. Should have saved it at the time. It was a reaction gif where each frame they applied that algorithm. It was super creepy, with eyeballs appearing and disappearing everywhere.
Perhaps they want to keep it private because they plan to announce more later. So we might see it again.
Perhaps the argument should be steelmanned in that we should generally avoid using algorithms which are so complex that they aren't glass boxes. I doubt the idea to "simply follow gradients" can prove neural networks to be glass boxes because the output of that is still too complex. And we are clearly onto something here. If we can generate artificially hallucinated pictures today, it is not unreasonable to assume that computers will be able to hallucinate entire action sequences (including motor programs and all kinds of modalities) in a decade or two. Combining such a hallucination technique with reinforcement learning might be a key to general intelligence. I think it is highly unethical that there is almost no democratic control over what is being developed at Google, Facebook et al. in secrecy. The most recent XKCD comic is quite relevant: http://xkcd.com/1539/
> I think it is highly unethical that there is almost no democratic control over what is being developed at Google, Facebook et al. in secrecy. The most recent XKCD comic is quite relevant: http://xkcd.com/1539/
I consider myself to be very Left of center, but, I can't imagine what form of 'democratic control' you think is necessary over the research that Google and Facebook does.
I do not fault Google or Facebook for planning on time-scales longer than most governments. Governments ought to be doing this level of long-term planning, but are not (at least publicly)
It's a tricky ethical area. The Google post cites several research papers that seemt o provide more than enough information to replicate these results or get similar ones, which is good, because I think everyone should be able to explore these tools - I stick by my view from yesterday that this may be a scientific breakthrough.
At the same time, I can see the basis for some anxiety, because it's not hard to imagine proprietary research going a few steps farther and developing some sort of general intelligence or even a limited but extremely high-powered intelligence that would confer an overwhelming commercial advantage, and/or a political one. Suppose, as an exercise, that one developed an algorithm to maximize persuasiveness by first leading readers/listeners into a quiescent, semi-hypnotic state and then making your commercial or political pitch. There's certainly a potential for abuse.
In Europe this sort of thing tends to bring up the precautionary principle, the idea that you shouldn't do something without oversight and demonstrated minimization of risk. I think that's highly limiting, but expect some pushback against Google over this. Of course, I don't think democracy is all that wonderful either but then I'm a bit of a misanthrope.
> The Google post cites several research papers that seem to provide more than enough information to replicate these results or get similar ones, which is good
I agree that it is good, but even though the scientific theories and algorithms seem to be "open", having access to both the computing power and data-sets of Google, is not.
So one could replicate these experiments, but not quite on the scale that Google does. I'm not at all sure if it's practically possible for a single (really clever) person with a high-end CPU/GPU machine (and possibly some $$$ for Cloud Computing instances), to replicate something similar to the results in this blogpost.
The recognition nets used in the blogpost seem to be trained on a tremendously high number of training examples, to give the ability to "hallucinate" (or classify) such a great variety of animal species, for instance.
I'm not at all sure if it's practically possible for a single (really clever) person with a high-end CPU/GPU machine (and possibly some $$$ for Cloud Computing instances), to replicate something similar to the results in this blogpost.
It's very possible.
GoogLeNet[1] is an example in Caffe: BVLC GoogLeNet in models/bvlc_googlenet: GoogLeNet trained on ILSVRC 2012, almost exactly as described in Going Deeper with Convolutions by Szegedy et al. in ILSVRC 2014. (Trained by Sergio Guadarrama @sguada)
I'm just questioning whether an autopilot with a profit maximization heuristic is the best tool to guide technological progress. With democratic control I don't necessarily mean our current democratic systems but any kind of decentralization of decision making by voting. Yeah, I know that's vague, but given what appears be at stake it seems unreasonable not to consider alternatives.
Fine, my suggestion to solve this problem democratically was an applause light. It was an unfinished thought and a call for action. I agree that this didn't convey any new information, I just wanted to express my distrust towards these kinds of appeasement statements from people who are working on these technologies. Being able to peak at different layers of a CNN doesn't recover NNs from the fact that they are in many regards opaque to us (and will possibly always be due to their complexity). Statements of the sort "I have never understood those charges" makes it sound like they are pretty much ignorant of the potential risks associated with not knowing exactly what your program does (they can perhaps be with regards to current technology, but I could imagine that more advanced systems can potentially arrive sooner than overall anticipated).
> Combining such a hallucination technique with reinforcement learning might be a key to general intelligence.
Knowing that the most common parallel effect of induced hallucination via psychotropics is ego-loss (complete loss of subjective self-identity) [0], maybe they need to try completely inverse processes in order to create a sense of ego in a machine... Because what's real intelligence but one's sense of self?
I would argue that a sense of experience is a necessary precursor for that, and also that intelligence and consciousness are two different things, although (if I read you right) the latter certainly informs the former. Barry Sanders' A is for OX has many well-sourced musings on the emergence of consciousness as a product of literary capability vs. a purely oral tradition which you might find interesting, and of course I think everyone needs to read Jaynes, Dennett, Hofstatder on these topics.
There are plenty of reasons to be concerned about AI. Stop dismissing arguments because you don't like their conclusions. There is no law of the universe that the future can't suck.
Well, the problem with "avoid using algorithms which are so complex that they aren't glass boxes" is that for quite a few problems, the simple choice is to either use a machine learning solution that will essentially give you a black-box or to use an understandable algorithm that doesn't get nowhere near state-of-art accuracy and is practically useless. Speech recognition, for example.
My wife found them somewhat too intense to take in rapidly. If viewing these makes you uncomfortable you should probably steer clear of psychedelic drugs, which tend to induce this sort of imagery for hours on end; as you can imagine this would be mentally tiring at the best of times.
I'm only mildly trypophobic but those images did have a minor effect - and I have a possible hypothesis for why it happens: what these images and trypophobia-triggering ones all have in common is a huge number of edges of various shapes and sizes, and it's this "edge overload" stimulating many more neurons than usual that's causing the disturbance. I find that the repetitive, but not-quite-the-same patterns like (organic) holes or other curvy shapes have the greatest effect; in contrast, straight lines don't do much. This makes sense since straight lines probably only trigger neurons that detect one direction, but curves have many "directions" to them.
I've always wondered about this! Back when I was a teenager I remember playing around with one of the Kai's Power Tools to generate fractal-like images (I forget exactly which one it was, but it had a bunch of presets that would generate cell-like textures).
Thing was, I found them MASSIVELY anxiety-provoking and have never been able to figure out why. They'd literally make me panicky.
These images are doing the same. Even now, just thinking about them, my stomach is fluttering. It's something about the way they're organic, but I don't know what it is. It's definitely nothing rational.
I've never heard of trypophobia; this does seem like it. I wonder if it's closely related to the feeling of disgust; somehow an evolved response to keep us away from rotten food perhaps? Things like bacteria growing on bread, or beehives. Or any food that's started decomposing.
There's a cool documentary I remember seeing called "How Art Made The World". One of the things it talks about is how we're driven to make more-than-perfect representations of things in our art. Say we find something in the real world aesthetically pleasing. With art, we can take that aesthetically pleasing stimulus and exaggerate it, resulting in the art being more pleasing than anything in the real world.
I wonder if that's what happens here, with the almost-organic images somehow being 'hyper-disgusting' as they coincidentally line up with a hyper-exaggerated version of a stimulus that on the scale of disgusting might be 'slightly unappealing' in the form that we'd encounter it in the real world.
I found an image which stimulates the edge detectors in people's brains, far more than a natural image. It tends to cause people to feel weird and not want to look at it, in a way they can't quite describe. And making the image flash and rotate rapidly made it far worse.
This is fascinating. And important. We need better ways to see what neural nets are doing. At least for visual processing, we now have some.
This might be usable on music. Train a net to recognize a type of music, then run it backwards to see what comes out.
Run on the neural nets that do face popout (face/non face, not face recognition), some generic face should emerge. Run on nets for text recognition, letter forms should emerge. Run on financial data vs results ... optimal business strategies?
But calling it "inceptionism" is silly. (Could be worse, as with "polyfill", though.)
While I think this is beautiful, conceptually, I really am a bit terrified of the potential of this in reverse (the neural network for processing/understanding an image). With Google releasing their 'Photos' app, this network is about to get a direct pipeline for machine learning imagery to accelerate everything – my main fear would be the potential for this technology to be employed by weaponized drones able to scan a scene (with, eventually, incredibly high resolution cameras and microphones that far surpass human capability) and identify every single object/person in realtime (also at a rate that humans are incapable of).
Of course, there is great utility to be had as well, it just scares me to think about what could be done with this technology, in a mature form, if used for violent purposes.
This will happen for sure. Such super-perceptive computers will oversee our every movement.
Computers can already understand our emotions in writing, voice and from the expression on our faces, they can also estimate pose and understand your movements. They can label thousands of kinds of objects. And they're just starting.
They can also build neural nets 10x smaller by compressing a larger neural net while maintaining most of accuracy. That means once a problem such as vision or speech has been solved with a huge net, it can be transferred in a smaller, more efficient net.
> They can also build neural nets 10x smaller by compressing a larger neural net while maintaining most of accuracy. That means once a problem such as vision or speech has been solved with a huge net, it can be transferred in a smaller, more efficient net.
I'm starting to come around to sama's way of thinking on AI. This stuff is going to be scary powerful in 5-10 years. And it will continue to get more powerful at an exponential rate.
165 comments
[ 2.9 ms ] story [ 198 ms ] threadhttps://www.reddit.com/r/MachineLearning/comments/3a1ebc/ima...
http://ansible.uk/writing/c-b-faq.html
http://www.lightspeedmagazine.com/fiction/different-kinds-of...
And What Happened at Cambridge IV, which I can't find online.
https://books.google.co.uk/books?id=5d9hHvD-T7gC&lpg=PA264&o...
This last story makes the ML images even more disturbing!
Highly recommend these stories. They'd make a great black mirror episode.
https://photos.google.com/share/AF1QipPX0SCl7OzWilt9LnuQliat...
I get creeps from fractals, but now my whole body is itching. Humans are weird.
It can be found in the cover art of "10000 Days", an album from American metal band Tool. The original box comes with two magnifying lenses like this:
http://s21.photobucket.com/user/Stonergrunge/media/Mis%20cos...
This (and others in the cover) look stunning through these lenses.
I would say that's a pretty good description of how many psychedelic hallucinations unfold. They start off as noise in your vision which turn into loose forms which turn into geometric patterns which turn into etc...
Also, personally, these pictures looked 'trippy' due to color palette that reminds hippy raves, rather than fractals.
I've been looking at Aldous Huxley's "Doors of Perception" and other psychonautic works recently and he hypothesizes that these sorts of drugs filter out the usual signals from the CNS that shut out the parts of perception that are not important for you to receive for survival.
It might be some great leap of armchair psychology, but I think we're due for another psychedelic revival, especially considering the new advances in synthetic psychedelics, legalization of more harmless recreational drugs, new tests in medical research using MDMA/LSD/Psilocybin, and the cultural shift away from the 'War on drugs'.
Ibis: http://3.bp.blogspot.com/-4Uj3hPFupok/VYIT6s_c9OI/AAAAAAAAAl... Seurat: http://4.bp.blogspot.com/-PK_bEYY91cw/VYIVBYw63uI/AAAAAAAAAl... Clouds: http://4.bp.blogspot.com/-FPDgxlc-WPU/VYIV1bK50HI/AAAAAAAAAl... Buildings: http://1.bp.blogspot.com/-XZ0i0zXOhQk/VYIXdyIL9kI/AAAAAAAAAm...
I'd love to experiment with this and video. I predict a nerdy music video soon, and a pop video appropriation soon after.
https://photos.google.com/share/AF1QipPX0SCl7OzWilt9LnuQliat...
So, tell the machine to think about bananas, and it will conjure up a mental image of bananas. Tell it to imagine a fish-dog and it'll do its best. What happens if/when we have enough storage to supply it a 24/7 video feed (aka eyes), give a robot some navigational logic (or strap it to someone's head), and give it the ability to ask questions, say, below some confidence interval (and us the ability to supply it answers)? What would this represent? What would come out on the other side? A fraction of a human being? Or perhaps just an artificial representation of "the human experience".
...what if we fed it books?
Here's a good introduction: http://colah.github.io/posts/2014-07-NLP-RNNs-Representation...
Now, this isn't to say that the kinds of neural networks we build today are conscious, but I don't think that's because they're based on a simple mathematical model; I think that's because they don't have the network-level properties that conscious humans do, for example, a self-representation.
http://googleresearch.blogspot.co.uk/2015/06/inceptionism-go...
Reminds me very heavily of The Starry Night https://www.google.com/culturalinstitute/asset-viewer/the-st...
Lovely imagery.
I never had much luck with generative networks. I did some work putting RBMs on a GPU partly because I'd seen Hinton talk showing starting with a low level description and feeding it forwards, but always ended up with highly unstable networks myself.
full resolution image
I can't help but think of people who report seeing faces in their toast. Humans are biased towards seeing faces in randomness. A neural network trained on millions of puppy pictures will see dogs in clouds.
Give the machine millions of reference images to work from and then tell it to find those images in noise, and it will succeed (because it literally can't "imagine" anything else for the noise to be).
Which makes me wonder, are these sophisticated neural nets mentally ill, and what would a course of therapy for them be like?
This is different because, while still just math, it’s modeled on the processes of human perception. And when successfully executed, it plays on human perception in ways that were formerly the exclusive domain of humans – Chagall, DiChirico, Picasso – gifted with some sort of insight into that perception.
Future iterations of this kind of processing, with even higher-order symbol management could get really weird, really fast.
Makes me wonder what passes as good art nowadays. But yeah some of the renderings were particularly aesthetic.
In the same way a normal fractal is a recursive application of some drawing function, this is a recursive application of different generation or "recognition -> generation" drawing functions built on top of the CNN.
So I believe that, given a random noise image, these networks don't generate the crazy trippy fractal patterns directly. Instead, that happens by feeding the generated image back to the network over and over again (with e.g. zooming in between).
Think of it a bit like a Rorschach test. But instead of ink blots, we'd use random noise and an artificial neural network. And instead of switching to the next Rorschach card after someone thinks they see a pattern, you continuously move the ink blot around until it looks more and more like the image the person thinks they see.
But because we're dealing with ink, and we're just randomly scattering it around, you'd start to see more and more of your original guess, or other recognized patterns, throughout the different parts of the scattered ink. Repeat this over and over again and you have these amazing fractals!
The trippiness is further compounded by the rainbow-ish colour effect produced by the recursive function, which mimics the "shimmering" rainbow effect you commonly get around lights when tripping on LSD.
And also, when under the influence of various drugs you tend to see patterns, particularly faces, where there aren't any.
Do they exhibit self-similarity at different zoom levels?
1) Pick a function f: R^2 ==> R^2
2) Pick a region of R^2 (this could be the unit square for instance).
3) For each point in the region do the following:
4) Paint by number after assigning a unique color to each possible number of iterations.Here's the result of this process for the function:
f(x,y) = ( exp(x) * cos(y), exp(x) * sin(y) )
http://i.imgur.com/LZKavio.png
While I agree with your idea about fractals (though you're a bit vague on the math details to know for sure), I also believe that a large reason the images look so "trippy" is because there is some local contrasting effect at work, generating high-saturation rainbow fringes at the edges of details and features. You get loads of that on psychedelics as well.
I bet there's a pretty straightforward reason to explain these rainbow fringes, if one were to dig into it, though.
Another (unrelated) observation I had was the feeling that the neural net seemed to be reproducing JPEG-artifact type fringes in the images? Though it could be that I was just looking at scaled versions of already JPEG-compressed output images, the article doesn't provide details (if only they had been PNGs ...).
http://arxiv.org/pdf/1412.0035v1.pdf
http://arxiv.org/pdf/1506.02753.pdf
http://arxiv.org/pdf/1312.6034v2.pdf
https://www.youtube.com/watch?v=XNZIN7Jh3Sg
https://www.youtube.com/watch?v=ogBPFG6qGLM
(Or if you want to put them full-screen on infinite loop in a darkened room: http://www.infinitelooper.com/?v=XNZIN7Jh3Sg&p=n http://www.infinitelooper.com/?v=ogBPFG6qGLM&p=n )
The code for the 1st is available in a Gist linked from its comments; the creator of the 2nd has a few other videos animating grid 'fantasies' of digit-recognition neural-nets.
One from Vincent Vanhoucke: "This is the most fun we've had in the office in a while. We've even made some of those 'Inceptionistic' art pieces into giant posters. Beyond the eye candy, there is actually something deeply interesting in this line of work: neural networks have a bad reputation for being strange black boxes that that are opaque to inspection. I have never understood those charges: any other model (GMM, SVM, Random Forests) of any sufficient complexity for a real task is completely opaque for very fundamental reasons: their non-linear structure makes it hard to project back the function they represent into their input space and make sense of it. Not so with backprop, as this blog post shows eloquently: you can query the model and ask what it believes it is seeing or 'wants' to see simply by following gradients. This 'guided hallucination' technique is very powerful and the gorgeous visualizations it generates are very evocative of what's really going on in the network."
Perhaps they want to keep it private because they plan to announce more later. So we might see it again.
I consider myself to be very Left of center, but, I can't imagine what form of 'democratic control' you think is necessary over the research that Google and Facebook does.
I do not fault Google or Facebook for planning on time-scales longer than most governments. Governments ought to be doing this level of long-term planning, but are not (at least publicly)
At the same time, I can see the basis for some anxiety, because it's not hard to imagine proprietary research going a few steps farther and developing some sort of general intelligence or even a limited but extremely high-powered intelligence that would confer an overwhelming commercial advantage, and/or a political one. Suppose, as an exercise, that one developed an algorithm to maximize persuasiveness by first leading readers/listeners into a quiescent, semi-hypnotic state and then making your commercial or political pitch. There's certainly a potential for abuse.
In Europe this sort of thing tends to bring up the precautionary principle, the idea that you shouldn't do something without oversight and demonstrated minimization of risk. I think that's highly limiting, but expect some pushback against Google over this. Of course, I don't think democracy is all that wonderful either but then I'm a bit of a misanthrope.
I agree that it is good, but even though the scientific theories and algorithms seem to be "open", having access to both the computing power and data-sets of Google, is not.
So one could replicate these experiments, but not quite on the scale that Google does. I'm not at all sure if it's practically possible for a single (really clever) person with a high-end CPU/GPU machine (and possibly some $$$ for Cloud Computing instances), to replicate something similar to the results in this blogpost.
The recognition nets used in the blogpost seem to be trained on a tremendously high number of training examples, to give the ability to "hallucinate" (or classify) such a great variety of animal species, for instance.
It's very possible.
GoogLeNet[1] is an example in Caffe: BVLC GoogLeNet in models/bvlc_googlenet: GoogLeNet trained on ILSVRC 2012, almost exactly as described in Going Deeper with Convolutions by Szegedy et al. in ILSVRC 2014. (Trained by Sergio Guadarrama @sguada)
[1] http://caffe.berkeleyvision.org/model_zoo.html
http://lesswrong.com/lw/jb/applause_lights/
Knowing that the most common parallel effect of induced hallucination via psychotropics is ego-loss (complete loss of subjective self-identity) [0], maybe they need to try completely inverse processes in order to create a sense of ego in a machine... Because what's real intelligence but one's sense of self?
[0] https://en.wikipedia.org/wiki/Ego_death
Thing was, I found them MASSIVELY anxiety-provoking and have never been able to figure out why. They'd literally make me panicky.
These images are doing the same. Even now, just thinking about them, my stomach is fluttering. It's something about the way they're organic, but I don't know what it is. It's definitely nothing rational.
I've never heard of trypophobia; this does seem like it. I wonder if it's closely related to the feeling of disgust; somehow an evolved response to keep us away from rotten food perhaps? Things like bacteria growing on bread, or beehives. Or any food that's started decomposing.
There's a cool documentary I remember seeing called "How Art Made The World". One of the things it talks about is how we're driven to make more-than-perfect representations of things in our art. Say we find something in the real world aesthetically pleasing. With art, we can take that aesthetically pleasing stimulus and exaggerate it, resulting in the art being more pleasing than anything in the real world.
I wonder if that's what happens here, with the almost-organic images somehow being 'hyper-disgusting' as they coincidentally line up with a hyper-exaggerated version of a stimulus that on the scale of disgusting might be 'slightly unappealing' in the form that we'd encounter it in the real world.
The image: https://i.imgur.com/Zihujue.gif
And it flashing: http://makeagif.com/i/hHzAiq
This might be usable on music. Train a net to recognize a type of music, then run it backwards to see what comes out.
Run on the neural nets that do face popout (face/non face, not face recognition), some generic face should emerge. Run on nets for text recognition, letter forms should emerge. Run on financial data vs results ... optimal business strategies?
But calling it "inceptionism" is silly. (Could be worse, as with "polyfill", though.)
https://en.wikipedia.org/wiki/Eigenface
E.g. SaaS that takes your images and use neural network transformations. Can you make a portrait of my that I look like king.
Of course, there is great utility to be had as well, it just scares me to think about what could be done with this technology, in a mature form, if used for violent purposes.
Computers can already understand our emotions in writing, voice and from the expression on our faces, they can also estimate pose and understand your movements. They can label thousands of kinds of objects. And they're just starting.
They can also build neural nets 10x smaller by compressing a larger neural net while maintaining most of accuracy. That means once a problem such as vision or speech has been solved with a huge net, it can be transferred in a smaller, more efficient net.
This is known as "dark knowledge". Slides from Geoff Hinton: http://www.ttic.edu/dl/dark14.pdf