Five healthy volunteers were presented with a series of random black and white grid patterns. Each grid square was either black or white, and this was randomly determined on each trial.
There was no pattern to the images, they were completely random. But the subjects were told that half of the patterns contained a hidden face, and that their job was to work out which ones did. Each subject saw over 10,000 random images and they took about 1 second to judge each one.
The volunteers "detected" a face in 44% of the images. Somehow, all five of them convinced themselves that they were seeing faces in many of the grids. The authors say that
Upon completion of the experiment we debriefed observers, and all expressed shock that no face was ever presented.That's strange enough in itself, but here's the really clever bit. The authors compared the patterns which were declared to contain a face, to the ones that were reported as empty. The image below shows the average "face" grid, minus the average "non face" grid, for each individual subject:
As you can see, this reveals...a face! Kind of. The top half shows the raw average; the bottom half shows the statistically significant differences from random noise.
In Subjects 1 and 2, the face is pretty clear, with eyes, a nose and a mouth. For 3 and 4, it's less coherent, but you might be able to see it if you look hard enough. For Subject 5, not really.
What this means is that people (at least, most of them) were not just seeing faces in any noise. They tended to see faces when the random patterns happened to resemble a kind of primitive face, but it was a different face for each person. The authors say that these strange faces correspond to the individual's internal representations, or models, of "a face", that each subject was "seeing" in the noise.
Finally, the whole experiment was conducted while EEG data was being recorded from the participant's brains. The EEG results revealed that there was a clear difference in the neural activity associated with "face" compared to "nonface" stimuli - except in Subject 5, who you'll remember had the least coherent "internal face".
What's exciting about this approach is that it investigates perception in a purely "top down" way. Normally, when we look at anything, what we end up perceiving is a product of "bottom up" influences - the raw data - and "top down" ones - what we expect to see. In this experiment, there was no real "bottom up" data; it was all "top down".
This is a form of pareidolia - perceiving familiar things in random stimuli. Seeing the face of Jesus in your sock, that kind of thing. It works for sounds too: in the famous White Christmas Experiment, people report "hearing" music in pure white noise - when told to expect it. Real-life examples of this include the "Islam Is The Light" doll, and my personal favorite, the singing paedophile Christmas mouse.
Finally, I wonder what embodied cognition theorists make of this paper. Because this paper claims to be "Measuring Internal Representations from Behavioral and Brain Data"; embodied cognition (at least the radical kind) is the theory that "internal representations" either don't exist, or at least don't explain anything about human cognition.
Smith, M., Gosselin, F., and Schyns, P. (2012). Measuring Internal Representations from Behavioral and Brain Data Current Biology DOI: 10.1016/j.cub.2011.11.061
30 comments:
Actually, pareidolia is specifically the act of seeing faces in random stimuli, which would naturally mean imagery. For a broader range of stimuli, including things that are not faces and backgrounds such as audio, the term is apophenia.
My initial reaction;
The task is 'view image, judge face/no face' when you're expecting half the stimuli to have a face in there somewhere. The average face analysis suggests to me that Subs 1-2 were picking, on average, the one that looked more like a face from each pair, while subs 3-5 were closer to chance. So two subjects were finding something actually in the stimuli to latch onto, and and three weren't. It's possible the last 3 subjects saw images that just didn't have the lucky structure in it; the grids were generated at random and everyone saw a different set. There's no analysis in the paper of the actual stimuli presented, I don't think, and if you don't know what you pumped into the system you can't draw conclusions about what came out.
By the by, if the average image is the person's internal representation of a face, then subs 3-5 don't have one and shouldn't be able to perceive faces (from a cognitive perspective).
If I get a chance I'll look at this in more detail, but I'm swamped with grading and more just now.
Homfrog: Thanks for the comment. But according to Wikipedia pareidolia is "a vague and random stimulus (often an image or sound) being perceived as significant" and although most of the examples they give are faces, they also cite other objects and auditory stimuli.
Of course it's quite possible Wikipedia are wrong!
Andrew: Thanks for the comment; I need to read the paper again, I'll get back to you...
But your point about Subjects 3-5 not being able to see faces is a good one (although I still think 3 and 4 look quite facelike, that may just be my pareidolia...)
I've done these quizzes before in fashion magazines but their criterias were based on personality types. If you see faces you were apparently neurotic. So I just knew straight away there were no faces thanks to my tetris acumen.
Very interesting.
I don't really agree that it's purely "top-down". If it was, there would be no average difference between face and non-face patterns.
Easy to suppose this would be one of those illusions that autistics aren't prey to.
Right at the beginning, upfront, I am asked to find a face. Why would I disbelieve you? So I try to find a FACE. Like that famous "The Economist" Ad a few years ago, it is a "leading" question.
If the opening question is framed "Is there a face in it or are they random dots (Richard Gregory's puzzles)?", the response (and the patterns in the brain) would have been different, I guess.
Thanks neurosketic for that post since fRMI and other researchers have not often neurological quality symptoms in schizophrenia to work on and "top down" cognition control over images our brain protects us from seing in order to "keep sane" is a priority to study.
If you want to make a difference in easing schizophrenai suffering that is.
PSQ: My take about the hollow mask-just an educated thought but not a passing one - is that for humans a face with a "hollow nose- is frightening like a skull or a demeanor or whatever. We suppress it when schizophrenia prevents you to do that or you become schizophrenic because you cann't fool yorself into a rreassuring -false - image of the world.
Nice post!
ramesam: Right - but that's the whole point of the study. You're asked to find "a face" - you believe it because you have no reason not to - and you start looking for one.
I was quite relieved to hear that the images were random, as I thought momentarily that I was lacking some innate ability to detect faces. Then I started wondering if I had Asperger's. Now I realize I probably just need an increase in my anti-anxiety meds. At any rate, cool post!
What occurs to me is that the "average" of the subject's choices could be skewed by the fact that everyone might be looking for eyes/nose/mouth, but S1 and S2 tended to find those marked in black -- while S3-5 might have occasionally seen "faces" where the ENM were white against a black background. If this were the case, the average would look like a blob.
Likewise, the eyes and mouth could be in different places at different times. So don't say that S3-5 didn't see faces! Their variety of faces might just have been more diverse in shape and color.
Andrew Wilson, you make a very good point. However, it seems that the participants who were able to consistently pick (or recognize) face-like stimuli, did so quite differently, and I must say I find the personal differences fascinating!
I don't see how this experiment is at all in conflict even with the radical embodiment theories.
1. As the blog post you link to states: "The hypothesis of embodied cognition is not necessarily anti-representational". I like the idea of 'representationally hungry' problems and this could be one of them.
2. But also we should keep in mind that as the post explains: "One reason psychologists persist with representations is that they simply can't see what else cognition could possibly be, if not representational and computational."
I'm not sure you actually need a very strong commitment to a theory of mental representations to interpret these results. Face recognition is not a simple pattern matching exercise against a passive storage of images. It is a creative (richly embodied) process. What actually happens when you look at a stain and see a face (prompted or unprompted)? It doesn't seem plausible that you're matching the stain against some representation of a face. I obviously haven't thought this completely through - but I can certainly see a much weaker representational or completely non-representational route to accounting for the same data. (But maybe that's just because I'm primed to see such routes by my non-representational linguistic background.)
to me this exactly like the sound/voice experiment. The one where they let people hear a random sequence of sound, then tell it says 'blablabla' and after that it's impossible for a certain %tage to hear anything else in it then the given phrase.
Here it is: see face, so you see a face.
and that's easily explained by how consciousness works. It's a storyteller that makes a somewhat logical sequence of events out of a mess of various datainputs. It has to make sense, thats' what the storyteller does. So sense it makes, if there is any or not is irrelevant as long as the story holds together.
People can go on testing till they are blue in the face but in the end, that what the talking part puts out is what the storyteller made up.
And then once is surprised that most storytellers on average tell the same story whilst sharing the same 'reality'
Psychology is really more like astrology. Whatever you look for you'll find
I'm still not convinced this is top-down "internal representations". What if they have bad eyes with stronger visual completion mechanisms? size constancy limits? Vary the scatter sizes? I see faces in sand or random noise than tetris bits.. could just be my retina going overdrive reaching limits.. convincing myself Jesus is staring at me from the sands of time even when the stimulus is changed to a block of cheese.. I'd be too switched off to notice.. lost in thought. Perhaps I've reached a threshold to not process anything at all but to think instead.
Scatters are pretty contentious. Any kid with trained video game retinas would refute such a silly question. I knew straight away there were no faces and then figured out a couple minutes later why. My retina senses are trained from hours of tetris boredom on the train, just another rat trying to make a living.. to not switch off and assume completion patterns dictated to me. Old people see anything. If you pressure them to see a face they'll see a face but it doesn't mean they saw anything at all.. their mouths took precedence over their senses.. probably why people like to ramble as they get older. I'd buy this as social expectation study.. not internal representations because they're lying about a "FACe! There's a FACE!"
Petrossa.. psychology as astrology is a tad harsh? This study smells poorish with contentious intervals and averaging outs. Then wack on lies and mental stuff. Seems more Freudish than cognition. But at least it reassures us people can think when looking at tea leaves.
Ryan: Yeah, I thought that too. The faces will only appear in the average if it's the same face every time (or if they have overlapping features of the same "color").
Using more sophisticated stats it might be possible to pick out more features, although that would also raise the risk of false positives.
The stats involved are actually very similar to the ones used in anaysis of fMRI images. It's the same basic idea of looking for "blobs", I think.
So basically, if your give someone something and tell them to look for something, they are going to try and relate it to something or someone they know?
Your title captured me, and I clicked on it. I like challenges and I thought I would be able to find a face in one of the two images.
I convinced myself I could find it if I studied hard enough, then I scrolled down trying to find an answer. Then I kept reading and figured out that there wasn't an image to be found.
Kind of cool how that happens, and a very interesting blog and study they conducted!
That goes for everything though, if you ask someone to look for something they are going to do it.
Try and look for red cars. You'll be amazed how many you suddenly see where before you only seen a grey/silver mass of tin cans.
If you generate random stimuli and then sort them across some dimension, the result is non-random. This is hardly surprising and is not an example of 'seeing something when it is not there'.
This is like randomly generating numbers between zero and 100 and asking participants to identify primes. They will identify the prime numbers (if they are moderately mathematically competent), but, the way the argument is put here, the numbers are just random so the participant has identified patterns where none are there. Baloney. The averaging across the stimuli shows that at least participants 1 & 2 are identifying stimuli that are, by random chance, more facelike. This is hardly a purely top down process and in fact proves the sensitivity of bottom up processing to statistical regularity and the ability of people to sort along interesting dimensions.
Another way of thinking about it: if you take a signal detection theory model and train or instruct observers on some psychophysical task, then take away the signal, there are still trials where the noise (internal and external) looks 'stimulus-like'. This can reflect noise in bottom up processing, stimuli that are close to the 'signal' condition by random chance, and a host of other factors. NDB.
I think this is a really interesting paradigm with a terrible discussion that massively overreaches.
Oops... NDB was meant to be NBD. No big deal.
CM,
Thanks for the information you gave us.
I just cannot agree to the fact that wasting ressources with badly design unproductive models when to study schizophrenia (for example with fRMI) Top down cognition might be so precious is "no big deal"
it is a big deal to waste ressources and produce flawn models if you take an economical or ethical view at it.
31 January 2012 10:52
neuroskeptic,
Thank to suppress my 31 January 2012 10:52 comment.
it is so sad for me that English -I have a poor command of - is what the latin was to our ancestors wanting to communicate accross the western world. As soon as mandarin Chinese replace English, you will understand what I mean.
Hmmm... How can they claim to have eliminated the bottom-up influences, when the figures clearly show that stimulus structure did play a role in the decision (as shown by the face-like average images)?
You say "the face wasn't there". Well, according to these figures, some "kind-of-facey" signal was actually there. Yes, it was generated by chance - but it was still there.
If I take a bunch of real face images, and hide them under a lot of noise, I would get a result that is similar to these stimuli. If the subject is able to detect which is which above chance, should I call it "pareidolia" as well?
Perhaps it's just me, but something bothers me about the face templates/classification images. They made these images by subtracting the sum of the nonface stimuli from the sum of the "face" stimuli. But they don't normalize by the number of trials, and I think this is a problem.
Say in each stimulus black pixels were 0 and white pixels were 1. Pixel a was white on 3000 of 4000 face trials and white on 4500 of 6000 non-face trials. So, the difference is -1500, and you've a big difference between face and nonface trials. However, that pixel was white on the same proportion of face trials as non-face trials (0.75). In this case, the pixel appeared white on 7500 trials out of 10000, so there was a bias in the stimuli that comes out as a difference between face and nonface stimuli if you don't normalise by number of trials.
Let's say there was no bias, that pixel a was white on half the trials. Again people say there are faces on 4000 trials (i'm using 4000 as in the paper they said there were faces "detected" on something like 40% of trials). so, if these are just randomly picked, we get 2000/4000 times the pixel is white in face trials and 3000/6000 times in nonfacetrials. So the difference is -1000, and again we've got a difference between faces and nonfaces. But, proportionately, it was white 50% of the time for both sets of faces, so there would be no difference if you normalized by number of trials.
Now: if the participants are not responding to face and nonface on 50% of trials respectively, the only way there can be no difference between faces and nonfaces using the subtracted sums method is if there is a bias in the stimuli:
if pixel a is white 2000 times on 4000 trials and 2000 times on 6000 trials, it was white 4000 times on 10000 trials.
Now most of the time obviously the differences won't be as stark as that and will bobble around the middle of the bell curve, but still...
Additionally - where would these numbers - -1500 and -1000 - be on the classification images? no idea, since they don't give a scale, and thus we can't tell which direction the differences are in!
Matt Craddock: Mmm... that's a good point. I had assumed that they averaged the "face" and "nonface" stimuli, and then did the subtraction (because that is equivalent to normalizing).
But according to the paper they didn't. They say that "For each observer, we summed together all of the “face present” noise fields and subtracted from that the sum of all “face absent” noise fields. "
Now because they only classified 46% of the trials as faces, this means that the sum of the "face" stimuli will be "darker" (or lighter, depending on whether black is 0 or 1) than the sum of nonface ones. Because there's just more non-face trials to sum up.
However - I can't see why this would produce a face-like structure (or any structure). It ought to affect all pixels equally.
So I don't think it can explain the "inner faces" but, you're right, it seems a bit odd.
Ah, but hang on. This is only an issue if we assume that the pixel values are 0 and 1. But if they're 1 and -1, it normalizes itself automatically as you sum them (I think?) because (assuming there's no structure) every 1 is cancelled out by a -1.
yeah, true, was just going off the typical convention for images. wish it were better explained in the paper. Same with proper labelling of the figures - it seems implausible that the face-like structure could be because of patterns in the non-face stim, but it'd be nice if it was a bit more clearly labelled etc.
Post a Comment