The sensory centers of our brain are largely activated by anticipatory computations
Have you ever noticed how the regular drip of a faucet completely fades out after a while when you are resting comfortably? It is, however, the sudden absence of this acoustic stimulus that unmistakably pops up in your mind. This everyday experience bears witness to a foresighted mechanism that, according to new research, governs the workings of perception: "Predictive Coding".
How does the world get into our head? The question seems pretty unproductive at first glance, as we obviously must only open our eyes to obtain a picture of reality. According to common sense, perception functions a bit like a submarine commander who peeps through a periscope and scans the horizon for suspicious activity. Another popular metaphor is the camera that films the course of events and forwards the input to a monitor in our head, in front of which sits an imaginary "homunculus", satisfying it curiosity. In reality, however, all comparisons with optical devices do our sensory apparatus a disservice. When we consciously perceive, we are more like viewers in a cinema who catch sight of a well-styled work of art, tinkered behind our backs by a script, directing, editing, censorship and other brownies.
Between the first draft on the retina (or in the other sense organs) and perception sits a hidden recognition service that constantly sets up and discards hypotheses, guesses missing locations, whitewashes inconsistencies and applies complicated mathematical formulas that would have overwhelmed us at school. The work of this "ratiomorphic system" remains hidden to our consciousness, which only marvels at the finished results and takes them for granted.
The most controversial issue that has long been plaguing cognitive psychology is this: Is that which we perceive only a passive and mindless registration of stimulus information crackling onto our sense organs? Or is the picture in front of us decisively and from the outset affected by stored experiences, expectations and knowledge? In the first case, that of "direct perception", information processing is "bottom up" or "data driven", that is, from below, from the smallest sensory bits of information, up to the higher cognitive centers. Our intellect can then only interpret the final image, like an intelligence official quibbling over a satellite photo.
On the other hand, in case of "indirect perception", expectations and previous experiences serve as foundation, controlling the act of perception. The process is here "top down" or "conceptually driven", that is, knowledge constantly slips into perceptions and generates hypotheses about the expected stimulus material. With every modification of the hypothesis, incoming sensory data obtain a new structure.
According to the current paradigm, perception of the outside world is not a passive process in which the "receiver" is passively fed sensory impressions. Rather, the organism at any time produces a "concurrent world model", which includes hypotheses about the expected stimuli. These expected values are stored in long-term memory as a comprehensive simulation of external reality. During an ongoing act of perception, the retrieved hypotheses are checked against the incoming sensory data; perception is therefore an interactive process, which is taking shape through a gradual testing and refinement of predictions.
This new perspective skews the whole picture: Our expectations control what we perceive; memory and perception are inextricably linked. The world outside answers questions which our brain poses. The best evidence primarily originates from studies on the architecture and the activity patterns of the brain. The different areas of the brain are never connected to each other in only one direction; there are always feedback connections leading from the higher centers back to the lower centers. Even more importantly: In the sensory system they even make up the majority.
A few years ago, a seminal meta analysis evaluated numerous studies looking at the primary visual cortex - the drop-in center for optical input - using functional magnetic resonance imaging. The inspection showed that this area is much more busy trying to process the feedback signals from the higher-level brain regions than to analyze information from the visual system. In other words: The activity of the primary visual cortex is surprisingly independent of external stimuli. More than 90 percent of the impulses that arrive there do not originate from the visual pathway but from "higher" areas of the cerebral cortex.
Already two decades earlier, researchers had began to wonder what these feedback loops in the brain are all about. These considerations lead to the theory of the so-called "non-classical" effects of neurons of the visual cortex. Until then, it was believed that nerve cells in this area are all responsible for the representation of visual information. However, the functioning of some neurons can be explained more elegantly assuming that they compare the incoming signals with expectations. The researchers therefore called these neurons "error finders" and described their activity as "predictive coding".
The concept is traceable to telephone technology and data processing. Instead of transmitting the entire signal, it is often sufficient to consider only the deviation from the previous signal. When dubbing an image file, it doesn't make sense to indicate the color of each individual pixel separately. Only when the color changes from one to the next point, this information needs to be transmitted. By merely coding deviations from the expected, this method (which led to mp3 and the demise of the music industry) reduces transmission overhead and increases processing speed. The theory of the hypotheses testing brain assumes that a similar principle governs most brain functions.
The fact that knowledge determines vision can be easily demonstrated via the so-called "degraded images" (see the pictures below). At first glance, most people have great difficulties to recognize something meaningful. Having read the annotations, the identification runs smoothly. Even more: With some, if not all pictures it now becomes nearly impossible to carry back one's mind to the naïve state and to forget the identities of the images.
The interplay between base and superstructure in vision has also become increasingly likely since computer science has taken up the problem of perception. When you are developing artificial intelligence, you inevitably face the question of the architecture of perception. The advantage of a program lies in the fact that it forces the researchers to formulate any assumptions explicitly. And here there is a clear trend: All programs that have been developed for machine perception are based on a form of "cognitive mediation". The computer is always fed with assumptions about the nature of the expected segment of the outside world, and this knowledge paves the way for the processing of the visual patterns captured by the camera.
Ask yourself how any artificial intelligence uninstructed by assumptions and expectations could navigate the physical world. The problem would be that such a program could not know which visual data are important in the current situation. The software would be constantly on the verge of a crash, even missing only a tiny bit of information. It could not fill in gaps through the context or through foreknowledge - just as humans can in the case of degraded images. We have no idea how often our perception completes degraded images in everyday life. It is even thought that most visual stimuli are underdetermined, meaning not sufficient to uniquely identify the object by themselves. For a computer that has to work its way up from the bottom of the sensory facts without "higher" support, degraded images would eternally remain eternally meaningless doodles.
Literature: Jakob Hohwy, The Predictive Mind
Full article at http://neurosciencenews.com/neural-networks-evolution-brain-2203/.
A new study by Northeastern physicist Dmitri Krioukov and his colleagues suggests an answer: to expedite the transfer of information from one brain region to another, enabling us to operate at peak capacity.
The research is in Nature Communications. (full open access)
Research: "Navigable networks as Nash equilibria of navigation games" by András Gulyás, József J. Bíró, Attila Kőrösi, Gábor Rétvári and Dmitri Krioukov in Nature Communications doi:10.1038/ncomms8651
Image: Krioukov and his colleagues discovered that the structure of the human brain has an almost ideal network of connections (magenta), enabling optimal transmission of information from one part of the brain to another. Image credit: Krioukov.
#neuroscience #evolution #neuralnetworks
// omg yessss
This shows a lot of tiny particles moving around. If you were one of these particles, it would be hard to predict where you'd go. See why? It's because each time you approach the crossing, it's hard to tell whether you'll go into the left loop or the right one.
You can predict which way you'll go: it's not random. But to predict it, you need to know your position quite accurately. And each time you go around, it gets worse. You'd need to know your position extremely accurately to predict which way you go — left or right — after a dozen round trips.
This effect is called deterministic chaos. Deterministic chaos happens when something is so sensitive to small changes in conditions that its motion is very hard to predict in practice, even though it's not actually random.
This particular example of deterministic chaos is one of the first and most famous. It's the Lorenz attractor, invented by Edward Lorenz as a very simplified model of the weather in 1963.
The equations for the Lorentz attractor are not very complicated if you know calculus. They say how the x, y and z coordinates of a point change with time:
dx/dt = 10(x-y)
dy/dt = x(28-z) - y
dz/dt = xy - 8z/3
You are not supposed to be able to look at these equations and say "Ah yes! I see why these give chaos!" Don't worry: if you get nothing out of these equations, it doesn't mean you're "not a math person" — just as not being able to easily paint the Mona Lisa after you see it doesn't mean you're "not an art person". Lorenz had to solve them using a computer to discover chaos. I personally have no intuition as to why these equations work... though I could get such intuition if I spent a week reading about it.
The weird numbers here are adjustable, but these choices are the ones Lorenz originally used. I don't know what choices David Szakaly used in his animation. Can you find out?
If you imagine a tiny drop of water flowing around as shown in this picture, each time it goes around it will get stretched in one direction. It will get squashed in another direction, and be neither squashed nor stretched in a third direction.
The stretching is what causes the unpredictability: small changes in the initial position will get amplified. I believe the squashing is what keeps the two loops in this picture quite flat. Particles moving around these loops are strongly attracted to move along a flat 'conveyor belt'. That's why it's called the Lorentz attractor.
With the particular equations I wrote down, the drop will get stretched in one direction by a factor of about 2.47... but squashed in another direction by a factor of about 2 million! At least that's what this physicist at the University of Wisconsin says:
• J. C. Sprott, Lyapunov exponent and dimension of the Lorenz attractor, http://sprott.physics.wisc.edu/chaos/lorenzle.htm
He has software for calculating these numbers - or more precisely their logarithms, which are called Lyapunov exponents. He gets 0.906, 0, and -14.572 for the Lyapunov exponents.
The region that attracts particles — roughly the glowing region in this picture — is a kind of fractal. Its dimension is slightly more than 2, which means it's very flat but slightly 'fuzzed out'. Actually there are different ways to define the dimension, and Sprott computes a few of them. If you want to understand what's going on, try this:
• Edward Ott, Attractor dimensions, http://www.scholarpedia.org/article/Attractor_dimensions
For more nice animations of the Lorentz attractor, see:
David Szakaly has a blog called dvdp full of astounding images:
and presumably this one of the Lorenz attractor is buried in there somewhere, though I'm feeling too lazy to do an image search and find it.
With sufficient experience comes maturity, where one's data set has become robust against such errors. Racism in adults is offensive in small part because it reveals a mind that has not matured beyond such rudimentary ways of carving the world. An immature adult mind is almost impossible to change.
But that's not what's going on with Google. The result is a mistake to be sure, a sign of immaturity, underdevelopment, and a poor training set. But there's no reason whatsoever to believe that these dispositions have congealed in Google to the point of no return. Quite the contrary, these small stumbles are just peaking at the vast possibilities artificial intelligence is starting to make available. It shows every indication of continued, rapid improvement; it's handlers show every recognition that this is a mistake that requires correction. No one would yell at a baby for stumbling after its first few steps; Google deserves the same support from us here.
Ouch! It appears that, for the second time in far too short a time, a photo app designed to use AI to label people and things has systematically mistaken black people for gorillas. I wish I could say this was satire, or even a cruel joke, but it isn't.
While is right to say this is absolutely NOT OK, I believe that, to leave it at that, as an 'oopsie', would be to overlook a frightening possibility: that such mistakes are not coincidences that just happen to sound like racial stereotypes.
In other words, these AI's may be able to teach us about the same sorts of perceptual errors and biases in humans. Note the eerie correspondences: black people mistaken for gorillas, a blink feature that thinks all Asian people are blinking, etc... These don't just sound like our stereotypes: they are the same stereotypes.
Zunger's idea for a permanent solution, as well, points to the best remedy for these sorts of mistakes: to focus, not on differences (bigger or smaller lips, hips, eyes, etc...), but rather to focus on shared human features. In no way, in reality, are blacks mistakable for gorillas or Asian eyes for blinking eyes... unless one focuses entirely on the differences between people. This, in reality, is how bigotry operates: by getting you to notice and exaggerate the differences, and ignore the commonalities.
Let us therefore not dismiss these things as 'bugs', but rather as potentially valuable insights into both the origin of stereotypes, and hopefully for a solution to these issues as well.
Burn, media, burn! Why we destroy comics, disco records, and TVs
Americans love their media, but they also love to bash it—and not just figuratively. Inside the modern history of disco demolition nights, c
Using Smiles (and Frowns) to Teach Robots How to Behave - IEEE Spectrum
Japanese researchers are using a wireless headband that detects smiles and frowns to coach robots how to do tasks
DVICE: The Internet weighs as much as a largish strawberry
Dvice, Powered by Syfy. The Syfy Online Network. Top Stories • Nov 02 2011. Trending topics: cold fusion • halloween • microsoft. Japan want
philosophy bites: Adina Roskies on Neuroscience and Free Will
Recent research in neuroscience following on from the pioneering work of Benjamin Libet seems to point to the disconcerting conclusion that
Kickstarter Expects To Provide More Funding To The Arts Than NEA
NEW YORK — Kickstarter is having an amazing year, even by the standards of other white hot Web startup companies, and more is yet to come. O
How IBM's Deep Thunder delivers "hyper-local" forecasts 3-1/2 days out
IBM's "hyperlocal" weather forecasting system aims to give government agencies and companies an 84-hour view into the future o
NYT: Google to sell Android-based heads-up display glasses this year
It's not the first time that rumors have surfaced of Google working on some heads-up display glasses (9 to 5 Google first raised the