The basic neurobiology behind the 12-dot illusion | Hacker News Reader

The basic neurobiology behind the 12-dot illusion

theneurosphere.com

171 pointsneurosphere42 comments

The fovea, the area of high-resolution vision in the middle of the field of view, is surprisingly small, just a few degrees across. The resolution falls rapidly outside the fovea [1]. A lot of the detail we perceive in the periphery is actually the brain filling in blanks based on "cached" data.

The resolution drop could in principle be taken advantage of in computer graphics, especially in VR applications with robust enough eye tracking [2].

[1] https://en.wikipedia.org/wiki/Fovea_centralis

[2] https://en.wikipedia.org/wiki/Foveated_imaging

Hah, why not.. We already created perceptually correct audio CODECs. Doing the same for real-time image generation seems like a worthy research area for sure!

True. And it's being done already! Here are just a few papers by my graduate school cohort.

Pdiff ("perceptual diff") uses perceptual metrics including visual acuity to diff two images. This is frequently used for image based regression testing of websites. It's really handy when you need to diff images that don't come out pixel identical every time, like images made with different browsers, or images that involve some randomness.

http://pdiff.sourceforge.net

https://scholar.google.com/citations?view_op=view_citation&h...

http://www.graphics.cornell.edu/pubs/1999/RPG99.html

http://www.graphics.cornell.edu/pubs/2000/PTYG00.html

https://scholar.google.com/citations?view_op=view_citation&h...

Great! Very cool work.

The GI paper double thumbs up.

I always love getting a chance to promote these guys. http://www.getfove.com/

They get it. Eye tracking, foveated rendering, all that.

Any idea when they might launch if ever? Been waiting on that. I built my own gaze tracking system but would like a production-ready model.

What was your setup?

They send out monthly newsletters and things are coming along, just at a slower pace than everyone was hoping for. But no official release date yet.

Pair of Chinese endoscope cameras and a LX1000 webcam + free open source software for gaze tracking. Worked OK.

Interesting idea about the eye-tracking for VR: offloading computation to the brain!

I'll make a correction that while minor, makes a world of difference: filling in the blank is more accurately understood as the brain making bayesian inferences or predictions or best guesses about what's missing or uncertain. The top down interaction occurring there is much more interesting than the term caching implies.

>filling in the blank is more accurately understood as the brain making bayesian inferences or predictions or best guesses about what's missing or uncertain.

Do you have any sources, or further reading, on that?

The way the eye interfaces with the brain never ceases to amaze me. Another fun fact is that it even has it's own error correction mechanisms, one of which you can intentionally miscalibrate in order to see colors that aren't there for days, weeks, and sometimes even months after doing the calibration (which is acheived by looking at a very specific image pattern for a long time). I would not recommend actually doing this, as those who have have reported back that the illusory colors become quite distracting and actually cause a bit of emotional distress after many months, but it's crazy that the brain has this kind of chromatic abboration error correction programmed in in the first place. https://en.wikipedia.org/wiki/McCollough_effect

The color scientists who first told me about the McCollough effect adamantly refused to ever try it, saying it could damage your vision forever.

So I was scared to try it, but eventually I did and for me the effect went away after a couple minutes. Maybe I'm lucky, but the warnings & the legend of the effect lasting months make such a good story, I have to wonder if it's a little overstated.

But -- what if the effect is just as easy to unlearn somehow, as it is to learn? It's existence may have a lot to do with gratings being pretty uncommon in nature, leaving a weak spot in the system that is easily trainable and slow to re-adapt just because we don't stare at gratings very often.

I wonder if image recognition neural networks would be susceptible to this during training

How long did you stare at the test pattern?

For long-lasting effect mis-calibrating any of these low level visual recognizers, you need to look at the test pattern for quite a while. Just a minute or two isn’t going to prompt a long-term effect.

Ah, you're right. That part was left out of the warnings I got. "Jones and Holding (1975) found that 15 minutes of induction can lead to an effect lasting 3.5 months."

I don't remember how long I tried the first time, but I'm sure it was not longer than 2-3 minutes.

The article says an anti-McCoullough effect can override the first effect. Can you also undo or override one ME with another ME using opposing colors?

Your brain can also adjust to your eyes seeing everything upside down.

https://www.theguardian.com/education/2012/nov/12/improbable...

It's terrifying that you can innocently trigger bugs in your visual cortex that will take lots of time to fix.

A terrible idea would be to make a video game (or a short movie maybe) which triggered the effect. Imagine if the background in (e.g.) Super Hexagon were designed for this purpose...

Looking at the first YouTube result of Super Hexagon, since I haven't played it before, holy crap I think this playing game will already do brain damage! ;)

I "induced" the effect several years ago (At least 3 or 4, but probably more), and to this day the grating has a faint, but noticeable color tinge.

At least for me, it seems to be truly permanent.

Crazy! Ever tried to undo it using the opposite stimulus from what you did originally?

For those of you interested in science fiction whose interest was piqued by this comment, here's some required reading:

http://www.infinityplus.co.uk/stories/blit.htm

http://ansible.uk/writing/c-b-faq.html

I'm not sure the central argument fully explains the illusion.

If you can simultaneously see several stars in the sky using your peripheral vision using averted gaze, why not several dots?

I suspect uncertainty plays a part, but image completion from higher-order feedback that complete the lines might drive the illusion more. Put another way, I believe if you remove the gray lines, the illusion ceases to work.

Because stars in the sky is a high contrast visual. That applies here too, I increased the contrast using Photoshop and the illusion is gone, you can see all 12 dots perfectly fine — http://imgur.com/a/G4xR4 I think they are spot on, outside the fovea, contrast detection capability of our eyes dip drastically thereby blending the dots lying outside the fovea with the surrounding grid. Increase the contrast to the point where even the regions outside the fovea can detect it and the illusion goes away.

I don't think that explains it, either: Even in low contrast, the illusion fails w/o the lines, as far as I can tell.

Without the lines... there is much higher contrast.

I'm all for hearing an alternative hypothesis, but right now this article's seems pretty good at covering whta is happening.

I found that once I zoomed in to have only 6 dots on the screen and zoomed out, the illusion disappeared and I was able to see the 12 dots simultaneously.

I tested it on a friend and same thing but he had to stay zoomed in longer than me, for me it was instant, he had to stay focused 5 or 10 seconds.

And now I can't not see the 12 dots even 48 hours later without being exposed to the image.

I'm replying to you rather than making another top level comment to prevent sprawl, but I'm adding a different way the effect changed for me. (I'll probably try yours later, when I'm done playing with it, lol)

For me, holding a "picture" of what the picture is in my head and trying to heat map where I think my view is spotting the dots changed it from seeing about 4 (in a triangle or box) to about 6 (in weird squiggles). I think that actively trying to see it changes how my brain caches information as it moves the eyes' focus.

I was looking into this when the 12-dot illusion came out, and found an interesting refutation of the receptive field theory regarding a related illusion - the original Hermann grid. I'm not sure if it also applies to the 12-dot illusion, where the illusion seems to be more about foveal/peripheral accuracy.

You can read about it here, and play with a demo:

http://www.michaelbach.de/ot/lum-herGridCurved/index.html

http://web.mit.edu/bcs/schillerlab/research/A-Vision/A15-2.h...

I always saw dots between the keys on my Macbook Air. Thanks to you I finally know the name for this phenomenon.

Very interesting, thanks! Especially that second link.

Interestingly, I seem to have a little different effect than they did in the second illusion they present. First of all, I find there are three distinct perceptions, the left-right and up-down that they mention plus spinning (fan motion). I actually didn't see the left-right form at all until getting to the page with the larger star, although when I went back I could see it in earlier pages. On the page with rows of red x and green o with alternating color positions between rows, I see the top row as spinning and the other rows as up-down; after looking at it a bit longer than for the other pages to change it changes to all spinning, then can change back to just the top row spinning. Now when I go back to the similar colored dots page I see the same thing, although I didn't the first time.

Also interesting how you can affect what you see in that illusion - not reliably at first, but with increasing accuracy over a few minutes. Or at least affect what most of them do - that top row on some of them still does its own thing sometimes. I wonder if more complex perception changes would be possible with more practice. Very interesting stuff!

It's pretty surprising to learn how fast visual acuity falls off outside the fovea. Our brains are amazing at making us think we can see a wide field of view when we really can't.

This article was great, fun to read. I think this chart summarizes the whole thing:

https://goo.gl/images/e4JKt4

If you fixate on the dot in the middle, all letters are equally legible to your eyeballs. This lets you see directly the difference in resolution between your fovea and your peripheral vision.

I'm no expert on either field, but this neurobiology explanation sounds a lot like what you read about in recent neural network posts:

You have some base layers taking in the physical information.

Some of those are connected in groups, somewhat like the a convolution stage scanning a sub-area.

Sometimes, some of those areas collectively show something interesting and pass it up. Kinda like max-pooling.

With artifical NNs, there are also images that fool the network in various ways.

One odd thing I noticed. At the size the image displayed on my laptop, I could look at a dot and see the dots to either side, but not the ones above or below, even though they were the same distance apart. I guess at least my fovea must be a horizontal ellipse, rather than a circle.

I noticed the same. I wonder if this is a result of two horizontally separated eyes (assuming too too are not a cyclops), or a training effect of reading horizontal text regularly?

This was a fascinating read about the neurobiology of the visual system. Loved it!

The worst trait I see when domain experts try to write "introductory" material is a depressingly-common amnesia toward what it was like to not already know their subject matter. This article is so badly written, so overly and needlessly wordy, that I'm honestly considering deconstructing it line-by-line as a case study in how not to write articles for the layman.

Here's a brief example, using the most jarring line I've been derailed by so far (I still haven't finished reading this article, getting hung up on runs like this):

> So what is the purpose of lateral inhibition in the retina? Let’s consider what kind of stimuli are optimal for activating this bipolar cell.

"What is the purpose of lateral inhibition in the retina?" Uh, hi - I just got here. We are zero lines from the place where we were first introduced to any of these terms (the last two words of the preceding illustration's detail are literally the first place the phrase "lateral inhibition" even appears).

We've just been shown this concept, at all, for the first time - in a literally microscopic illustration, with some concepts so unexplained they were just left for us to hunt for ("Notice that a single neuron pools information" - notice what? Notice how?) - and now we're supposed to be considering the ultimate purpose for this obliquely-introduced phenomenon? Not only that - we're supposed to be pondering this now using a neurologist's lexicon?

If we were actually supposed to be following what the author is saying here, we'd be given some time to reiterate this concept we were just presented from one angle, to consider it in different approaches, with descriptions we don't immediately understand reinforcing a model we could build with other descriptions we could better understand.

Instead of reinforcing its subject, the article spends time on aside paragraphs mocking what a non-neurological model of human vision might be (talking about Dennett's Cartesian Theatre), despite the way that nobody reading this would have that misconception, and introducing it only adds a concept so unrelated that it actively impedes understanding of the material. Passages like this are just thrown in, like the author wants to say "I know more than you, and I want you to know that I don't just know neurology, I also took a philosophy course."

The line I excerpted doesn't even have to spend more time describing the concept. Indeed - it's actually more effective if you make it less wordy, because a high-level general description tells us which things we don't strictly have to understand to follow the next part. Here's how I'd write that whole paragraph - note how much less I hang meaning on unexplained jargon:

> So, when a wide area of photoreceptors see the same signal, they actually reduce the signal seen by the tight cluster in the center. Why reduce the signal like this? Because this way, when we see a small detail - one that isn't surrounded by a big area of the same kind of light - that wider group of photoreceptors (connected to the horizontal cell) doesn't inhibit the signal. Getting stronger signals for just the smaller details is what makes stuff like dots and edges visually obvious to our human eyes.

It's not perfectly clear - it still needs a round or two of editing to be truly smooth to read - but, even as a rough draft, that's smoother than anything in the linked publication, which reads like didactic sandpaper.

Bit of a harsh critique, though I appreciate reading people's comments on my writing. Ultimately, doing some writing over the past few months has taught me that I shouldn't necessarily underestimate people's ability to understand scientific concepts if only I take the time to explain them properly. I'm sorry that what I wrote wasn't sufficiently comprehensible :-( Also, I never took a philosophy class. None of what I said was meant to be pretentious or superior - I was just putting out there some concepts I found interesting.

If it helps, I had zero prior knowledge of "lateral inhibition in the retina" and I had no problem at all understanding the writing.

Yeah, don't take it personally. Most of what I wrote there was just pent-up generalized frustration with a lot of published material out there. I regret not sufficiently framing my post within that context.

The authors I really want to hit with this kind of critique are the ones who aren't embracing what you describe here:

> Ultimately, doing some writing over the past few months has taught me that I shouldn't necessarily underestimate people's ability to understand scientific concepts if only I take the time to explain them properly.

There's a troublingly large contingent out there that, as XKCD put it, [communicates badly and then acts smug when it's misunderstood][1], [confused about how communication works][2]. These are the ones I really take issue with: the ones who are pleased when the audience who can absorb their material is small, interpreting it as validation of their own abilities (thinking "I'm one of a few people who can understand this"), instead of a refutation of them.

[1]: https://xkcd.com/169/

[2]: https://xkcd.com/1028/ "Anyone who says that they're great at communicating but 'people are bad at listening' is confused about how communication works."

I particularly apologize for the part about it being pretentious to mention the Cartesian Theatre. While I would cut that in later drafts for being irrelevant (which, like all irrelevant material, confuses the subject), it really doesn't come across as anything more than that, and I was especially projecting my frustrations from other stuff I've read when I characterized it like that.

That the writing is subjective and hence too colorful and wordy is evidenced by the remark, "what I find most interesting about them".

Edit: It is however not as long as I derided in another comment. At least there is a link to a relevant paper in the beginning, although further reading didn't appeal to reading that paper. The text is almost half as long as the paper, that's too long for a summary. I'll refrain from more critique as I'm far from an expert on literature.

Your comment is ironically very wordy as well, but I concur.

Edit2: I am not seeing sources for the images.

git to the point. tenthousand lines and one of importance. to paraphrase: as neurons are connected to form fields of perception, we can recognize shapes quicker, but as the fields grow, we cannot resolve smaller details. I'm sure I got some detail wrong, didn't want to copypaste.

Reply on news.ycombinator.com