I was also fascinated by the results of Google's algorithm over(?)-analyzing white noise patterns. It not only generated singular images in different parts of the pictures, but overall themes (I assume due to its multi-level structure). I am interested in the question of whether human dreams may similarily derive from somatosensory cortices' white noise, with higher consistent story-like dreams as the results of further abstractions?

Yes this is indeed very exciting - it looks to me like a very good frontrunner for a computational explanation for the 'form constants' experienced under functionally selective 5-HT agonists, or psychedelics. It also does put a fairly convincing refresh on the case for gestalt psychology, and Helmholtzian 'Unconscious inference' even more so.

The only substantial work I have encountered/understand in any depth in this area is some of Allan Hobson's, which might be worth a consideration. In his theoretical perspective, these computational systems form a substrate from which all conscious visual perception arise, and are ontogenically developed in utero to ensure that newborn humans can rapidly adapt to the task of waking perception with minimal delay.

According to this theory, this same system of unconscious inference is revisited nightly, for goal-directed maintenance and evolution of its behaviour - and we experience this phenomenally as REM sleep. Thus dreaming can be conceptualised as the process by which such adaptations are tested and played out in the absence of veridical retinal input, in a kind of integrated rehearsal/failsafe mechanism.

What does this have to say about the imagination? Presuming that the visual imagination and the dreaming experience of REM sleep depend upon a common physiological structure or mechanism, which does stand to reason, it potentially demonstrates that very small degrees of top down influence over the visual cortex may be sufficient to produce vivid internally generated mental imagery, provided such automated inference can be similarly activated too.

Likewise, it suggests that since such automated/unconscious inference is integral to normal waking perception - its role in imagination could imply that what we assume to perceive veridical may be modulated and influenced by our higher cortical systems to a considerable extent, in contradiction to our intuitions on the veridicity of visual perception in general.

For specific psychobiological/phenomenal explanations for how this theory works, you might want to read "The Dream Drugstore", linked below. For a computational explanation you might want to check out some of Karl Friston's work, also linked.

Hope that's all useful :)

https://mitpress.mit.edu/index.php?q=books/dream-drugstore

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3314993/

Jarrod,

thank you for mentioning Helmholtzian 'Unconscious inference' . I have his book but the idea did not stick with me. Also thankyou for the  useful references to Hobson's and Friston's work.

"Presuming that the visual imagination and the dreaming experience of REM sleep depend upon a common physiological structure or mechanism, which does stand to reason," To me, anyway :) It is consistent, as mentioned, with similar architectures proposed by Lakoff and Gallese.

If I understand correctly, Helmholtz's theory was deemed logically incoherent because inference is reason and reason is conscious. Now it seems that refutation was itself incoherent, based on fallacious assumptions about how the mind works. Lakoff and Johnson's notion of  the 'cognitive unconscious' comes to mind. The idea that 'thinking' is entirely conscious and those activities undertaken by the brain which are non-conscious are trivial neural housekeeping seems quaintly anachronistic.

Dominik

" (I assume due to its multi-level structure)."

The devil is in the details, no doubt. I want to know more about just how specific the training in these specifiic layers is. ie. how much hard coding goes on for it to find 'biulldings' as opposed to 'faces'. It would be  nicve if we could get some more technical information

Simon,

I find it exciting primarily as an artistist tool to explore and make us of like other computer graphic tools that already exist.  The capacity to generate pareidolia is interesting but remember that simply artificially folding images like in ink blots is very effective at generating pareidolia.  It is because the visual system is tuned to bilateral sysmmetry, take it as significative of an object and then start the mental painging from there.  

Inceptionism is very similar to painting.  I am not a painter but when I draw or paint I constantly react to what I put on the cavas and this observation sudgest pareidolia that I enhances , put more paint , observe and enhanced the new pareidolia and so on.  Very similar to the positive feedback of the inceptionistic generation of images. It generate from an initial constraint a close by stable optimal image, one for which any departure is not increasing the match.  It is a fixed point of the neural network transformation.

I have difficulty to see the link you are making with the implication of this inceptionism technique and cognitive theories about the nervous systems.  I do not see anything new here that change anything to the way we see neural nets.

Hello Louis

"The capacity to generate pareidolia is interesting but remember that simply artificially folding images like in ink blots is very effective at generating pareidolia.."

Louis, my forehead is wrinkling with the sense that you miss the point. Inceptionism isn't simply making the (random) patterns, its 'seeing' and identifying the forms that it 'imagines' emerge. Its doing what you do when you draw out pareidolia while painting!

"It is because the visual system is tuned to bilateral sysmmetry,"

Oversimplification, I think. Though Rorsarch blots are bilateral, Pareidolia doesn't work only on  bilaterally symmetrical things. In any case, a man crouching is not visually bilaterally symmetrical except from two views. And seeing that already implies figure-ground separation.

Why do we have pareidolia? One evolutionary argument is that being highly attuned to subtle (visual) cues - a jaguar hiding in dappled forest cover - is good for survival. Pareidolia is (like?)  predictive coding. The visual-cognitive 'pump' is always 'primed'. False positives (which is what pariedolia is) are less dangerous than false negatives.

"I have difficulty to see the link you are making with the implication of this inceptionism technique and cognitive theories about the nervous systems. I do not see anything new here that change anything to the way we see neural nets."

Again, the point is not about neural nets, its about what Inceptionism can suggest about human vision and visual creativity.

Simon,

That I do not get the point is very likely.  Take my comments as attempts on my part to get the point.  I am trying.

From the point of view of Hauffman, Koenderink and many others vision is a controled hallucination and false positive pareidolia are hallucinations that failed which most often occur when the visual stimulus underconstraint the hallucination, when the hallucination part is revealed while so-called non-illusionary are in fact successfull or perfect hallucination where the hallucination is not reveal and stay hidden and where the observer is under the perfect illusion of seeing reality as it is when the mental painting achieves perfect illusion.  I will take a concrete example.  You are eating and see to potatoes on your plate.  You grab them with your fork and eat them and they taste good and you feel good.  It was a perfect potato hallucination that fit the your scenario of feeding and you go on surviving.  The hallucinations of seeing , of grabing and tasting where effectively successfull and in an utilitarism way to assess truth, a thruth is a productive belief, it is a perfect/rewarding hallucination. Biological evolution built you for you to live this kind of narratives, this myth we live by.

''Pareidolia doesn't work only on bilaterally symmetrical things.'' I did not say that.  What I am saying is: take any radom image, meaningless image and fold it and then it is very likely that a lot of pareidolia will be generated.  Why?  Because bilateral symmetry detection is very very important because so many natural surfaces has local bilateral symmetries and vision systems have evolved to detect these and to hallucinate from there.  This is a simplification as everything express in a few words.

''a man crouching is not visually bilaterally symmetrical except from two views.''

These two views are the most visually important ones.  It is well understood that human do not remember the animals view on a equal basis.  The studied of cave painting has reveal this.  

''And seeing that already implies figure-ground separation.'' I agree but all I have said do not contradict this.  In fact this realisation was a departure for my Ph.D. research.

I will read further on the topic.  I am not getting it yet.

You are right Simon: The notion that "activities undertaken by the brain which are non-conscious are trivial neural housekeeping" is indeed proving quite anachronistic. This is more or less the idea that the study of Complex Systems has blown wide open; with countless demonstrations for proto-intelligent and lifelike systems to essentially arise from trivially simple subsytems, when the initial conditions are favorable.

So why the original fallacy? I suppose we have Descartes and his philosophy of mind "dualism" to thank for all this. On the one hand: it did allow a scientific explanation for the reducibility of mental processes to physical matter in a way that was palatable to the political reality of religious sentiment at the time - however the price seemingly paid for this was that forever after, we have been stuck with the habit of framing these questions with the "centrality of the soul" as an indispensable component of our reasoning - sometimes without even realising, and often without even believing in a soul in the first place. This is not to say one can't believe in a soul, it's just quite damaging to presume it occupies the seat of all higher mental function and to then declare it scientifically out of bounds.

If you can trace back these arguments far enough in time and in the history of ideas, and see the fallacy of some of the premises established there; there is no fundamental reason to assume that high cognition or even self-awareness need be conscious at all, and in practice large aspects of our own conscious mind might even be compartmentalised and hidden from us, but still in operation and highly influential in and over our behaviour.

I suppose the sum conclusion you would arrive at in updating the understanding of the mind with this theory would be that our own imagination, and large aspects of the systems that underpin the way our visual perception is generated - may be in their own way display high degrees of intelligence, and yet exist fundamentally outside of our direct mental influence.

It's a hard idea to get around, if you frame it solely from the philosophies of the Greek tradition but eastern philosophy and to some extent continental philosophy are a lot friendlier to this way of seeing things. I sincerely hope that our capacity and ability to understand the mind is not to be come limited by the philosophical tradition we opt to swear our allegiance to: In my mind, they all have something valuable to say.

Louis,

I supposes the significance of this research for visual perception, following on from the ideas in my previous post, is that it practically demonstrates how the very systems that must painstakingly train themselves to facilitate visual perception, when unconstrained from veridical retinal input, may produce complex information from essentially nothing, ie white noise.

This further lends credence to the idea that the human imagination, or dreaming process is a kind of reverse-engineering of the human visual system, that has itself become so sophisticated it can demonstrate degrees of creative intelligence encapsulated within its own computational structure.

To date, the ideas expressed above have been argued extensively by Hobson and Friston, but I believe never demonstrated to this extent.

This of course all feeds into a plethora of other research, re the human imagination, creativity and intelligence. For now: it's just a very good example of how intelligence potentially emerges from supposedly unconscious, computational systems.

 Hi Jarrod,

I briefly read about Hobson and Friston theory of dreams and they say what I have been saying for almost 20 years but I am sure with a lot more details and evidences.  But I had articulated a self-enactment theory of dreams based on traces of activations of the sensory-motor systems in my Ph.D. thesis (first version).  The committeee forced me to cut it out.  The senses are not virtual reality machines they are reality machines.  There are the only reality we have access.  Since the early pre-socratic days when mathematics was created from a separation from the engineering practices and when the first naturalist narrative about the cosmos were proposed a battle of reality began.  Philosophers and later scientists tried to displace the daily normal reality of people based on their senses as an illusion and replace it by a theoretical discourse , a reality of ideas, a platonic one.  Plato casted the artists out of his ideal city.  Beleiving in the reality of the senses because of all these illusions and all we do not even perceive was and is still consider as naive.  I think this is the belief that we live in this shawdow theoretical world is the illusion and the perceptual illusions shows us more reality than the normal perception because the former lets us peak at the reality engine.  

"That I do not get the point is very likely. Take my comments as attempts on my part to get the point. I am trying."

Dear Louis, your humility puts me to shame.My tone was more excited than authoritative :) And much credit to you, and everyone, who takes part in these conversations in English when its not your mother tongue.

"Because bilateral symmetry detection is very very important"

granted.

"The studied of cave painting has reveal this."

 yes, Lambros Malafouris has a nice paper on this.

SP

Jarrod

our positions on (the danger of ) Cartesianism are similar - no argument there :)

"

our own imagination, and large aspects of the systems that underpin the way our visual perception is generated - may be in their own way display high degrees of intelligence, and yet exist fundamentally outside of our direct mental influence."

This is indeed a provocative idea. Reminsicent of multiple selves an Minsky's Society of Mind.

Yes indeed it is provocative! however I have eventually come to the conclusion that is is not so much the science that provides the greatest barrier to acceptance of this idea, but our intuitive resistance to having our concept of self so radically redefined. No amount of scientific training can change this; people will recoil from ideas that scare them and that's just another scientific obstacle to overcome. Something of a 'conflict of interest' that rarely gets discussed in the philosophy of science, I would say.

There is some very good stuff coming out of the labs of György Buzsákiand and Ed Bullmore, that details how multiple intelligences might actually look from a mathematical/graph theoretical perspective. From an empirical account for how this works in actual neural populations, there is a scientist called Walter J Freeman who has published some fascinating work too.

Obviously this has practical relevance for real life, as just about every psychopathology that I can think of has the capacity to be framed as a dysfunction in the emergence of the self  from its substrates. Unless of course we want to go back a few hundred years and attribute the phenomenon of psychosis to invading spiritual entities; we probably have to ask the question of whether mental illnesses creates phenomenon of multiple agents, or merely disrupts the illusion of a unified one.

Simon,

'' Specifically that image and pattern recognition mechanisms can be/are deployed for visual creativity.''

Athough I am very skeptical between the assumed similarity between these artificial neural nets and what actually take place in our visual system,  inceptionist demonstrate in this simple case that a network optimized through learning for recognition can be used for generation of ideal images.  This I think is a good insight that this artificial inceptionist algorithm suggests. 

I do not know much about the state of the art in neural net for visual recognition but I think that the major limitation of such approach is its limitation to a very small numbers of different objects.  And the reason for this limitation is they do not have a internal structure which reflect the hiearchical structure of what they classify.  We do not simply discriminate one objects from another ones, we are aware of similarities and put objects in large categories.  We seem to naturally put everything into hiearchies.  And when we invented languages we created words in such hierarchical relationships.

Jarrod

"No amount of scientific training can change this; people will recoil from ideas that scare them and that's just another scientific obstacle to overcome. Something of a 'conflict of interest' that rarely gets discussed in the philosophy of science, I would say."

I'm with you there.

"this has practical relevance for real life, as just about every psychopathology that I can think of has the capacity to be framed as a dysfunction in the emergence of the self from its substrates. "

right. say goodbye to "the self" , just hope your selves are getting a long. Play nice, selves! :)

Louis

"We do not simply discriminate one objects from another ones, we are aware of similarities and put objects in large categories. We seem to naturally put everything into hiearchies. And when we invented languages we created words in such hierarchical relationships."

We believe we do this on a conscious level, but we have no ability to introspect our actual neural processes.

All, here a link to an artwork Giver of Names, by the very clever David Rokeby. Begun in the late 90s, it also 'dreams' based on pattern recognition of imagery. I hope you find it interesting.

https://www.youtube.com/watch?v=sO9RggYz24Q

https://vimeo.com/17187792

http://www.fondation-langlois.org/html/e/page.php?NumPage=2121

Simon,

Thank for pointing out this cool stuff.  Regarding the hiearchical classification I am totally biased here by my own research.  Did you ever check the image surface of an image.  You first transform the image into a grey scale image and then you plot it as a topographical surface where terrain height correspond to luminance.  Check the image at various spatial scale where spatial scales is the convolution of the image by a Gaussian with the variance corresponding to the spatial scale.  Take an image of a nude human body. Make a movie of the different spatial scale of the image starting with the very very coarse spatial scale.  What you will see will amaze you. Compare what you see, the growth of the surface of the human body with its growth in the embryonic process.  Check at the sequence of development.  One movie is worth one thousand words.  So from this observation on all kind of plants, trees, humans, all the living organisms that grow, you see that the surface structure of the image is close to the object surface structure and that image surface ontogeny do correspond to object surface ontogeny.  And this ontogeny is generated at a sequence of symmetry breaking points allowing to classify the whole set of all possible image surface form into a gigantic phylogenitic tree. All image structure are a leaf of this tree and their similarity is their common ancestry.  This was my Ph.d. thesis.

I was particularly struck by the image generation exercise of the "dumbbell", and The Guardian's interpretation of it.

"For instance, asking the network to discover dumbbells in a picture of random noise reveals it thinks that a dumbbell has to have a muscular arm gripping it...The solution might be to feed it more images of dumbbells sitting on the ground, until it understands that the arm isn’t an intrinsic part of the dumbbell."[1]

Are we all certain that the arm isn't an intrinsic part of "the dumbbell"? Maybe not the arm itself, but the category of 'dumbbell' is that of an object and its function and purpose. Maybe what the author at The Guardian sees as a "problem" needing a "solution" is actually evidence for certain linguistic theories of categorization.

Also, to Simon, why the allergy to "to computational analogies for mind and brain"? It is, after all, our minds and brains that built these computers, and therefore would eventually somehow represent itself in these machines. Yes, archaic and misaligned terms like 'processors' and 'memory' and 'disk space' do not come close to equating our minds with machines, but this may prove more true as technology progresses and more 'natural' categories can be formed. Artificial neural networks are further evidence for such an argument, in my opinion, especially after reading Vinyals and Le's "A neural conversational model."[2] 

http://www.theguardian.com/technology/2015/jun/18/google-image-recognition-neural-network-androids-dream-electric-sheep

http://arxiv.org/pdf/1506.05869v2.pdf

Philip

Vinyals and Le's "A neural conversational model.

thanks for the ref, its interesting.

"It is, after all, our minds and brains that built these computers, and therefore would eventually somehow represent itself in these machines."

This is a subtle point, more subtle ,than the usual clubfooted - "the brain is a computer". I have a couple of responses off the top of my head. First, other technological advances built by minds and brains such as the "iron horse" and the "horseless carriage "  were not understood in the kind of autobiographical way that the computer is "as a brain". So we, culturally, have a lot invested in the brain-computer equivalence/analogy. Mind you, the camera has obvious physiological analogies.

Neural nets are *loosely* based on what we understand of neural structures, in the same way that genetic algorithms are "loosely* based on the mechanisms of evolution.

"as technology progresses and more 'natural' categories can be formed"

The problem I see with this idea is that *we* are continually reforming our categories about ourselves. If you could inhabit the mind of confucius or plato or alhazen or albertus magnus  or bacon or descartes, I think you would find their conceptions of self/other, mind/body, life/death and all sort of other structuring ideas quite alien or confusing. Its hard to know what  *natural* category would be.

SP

Philip.

The computer -brain metaphor is a pernicious metaphor because it carries with it many implicit assumptions which are not fruitfull for understanding biological organism:

1. the structure of the computer has nothing to do with its function because a complete separation between hardware and software.

2. it is a totally deterministic machine whose structure has everything to do with efficient algorithmic computation but nothing to do with relating with the world.  So it is the summit of the separation of the brain from the body and from the body to its environment.  It is the incarnation of an old conception Laplacian conception of the universe, a universe governed by a small number of deterministic laws.

3. Its memory is not in its structure but in data in memory.  Its structure does not change with learning, only the stored data.  Everything is explicit.

I can go on.  Lets say that this metaphor carries with it a large numbers of totally unsustainable assumptions hidden into it.  It has been one of the biggest hurdle in the development of artificial intelligence and computer science.  That metaphor pushed all these researcher two hundred years back of the philosophical reflection on this topic.  People like Heidegger had half a century before the computer metaphor came to the necessity of being-in-the-world which this metaphor is the complete negation in spite of the appareance that we can supplies the computer with enough sensors and actuator to be in the world.  But it is a alien to the world that is so connected.