The Illusion Machine That Teaches Us How We See | Hacker News Reader

The Illusion Machine That Teaches Us How We See

nautil.us

64 pointspmcpinto18 comments

Cool. It seems to make the common argument that "since evolution is random, it makes mistakes, therefore we see illusions." I disagree with this logic. I think it would be truer to say "since visual processing resources are (inherently) limited, evolution makes trade-offs, therefore we see illusions." I see jumping to the most-likely conclusion about what we see as more of a power-minimizing and performance-maximizing trade-off than a design flaw.

It may be a mistake to assume evolution 'designed' anything at all. Its a drunkard's walk over the solution space. As soon as a relative optimum point is reached, evolution tends to stick there and not search for further, more-optimum solutions.

Witness the 'thumb' of a panda, which is uses to strip leaves from the eucalyptus branch. It has a thumb (toe on its paw) but no, evolution chose to extend a wrist bone in a kind of protuberance; the panda pulls the branch between that and its arm to strip leaves. Not ideal; kinda weird really. But evolution doesn't give a flip.

How optimum the solution is is related to how important and conserved the system is. The visual system is very old, and it has unquestionably been subject to intense selective pressure.

Then explain the half-dozen entirely different vision systems extant right now? They are very different; they have all been under selective pressure.

It would be interesting to see if they all succumb to the same kinds of illusions. I guess the important question is whether these illusions are arbitrary and reflect arbitrary choices by evolution, or if they are fundamentally hard and would be hard for any hypothetical vision system. My hunch is that they are fundamentally hard.

Evolution doesn't design, clearly. But its solutions are not determined by chance alone, either. The shapes of the solutions are partly determined by the contours of the available solution space. It's sort of like pressing soft material into a mold, I think. The trajectories of the individual bits of material are governed by chance; but after applying and releasing pressure, the macroscopic object invariably forms itself in the image of the mold.

Sure but the 'mold' is a complex surface with many local maximums. Once evolution reaches a local maximum, then drunkards walk is stalled - it moves incrementally in any directin - things get worse. It cannot leap across the gap to another local maximum; ok its unlikely it will make the leap.

So we get different populations with wildly different solutions to the same problem.

But these aren't so much visual illusions as they are camera illusions. Their whole premise is the fact that they are a 2D projection of 3D objects. There are multiple 3D objects that project to the exact same 2D object, so I don't understand why there has to be a mistake involved.

That would be like saying two times ten is twenty, and four times five is twenty, so division has been designed incorrectly.

I guess the premise is that, given that multiple 3D objects could correspond to what we are seeing, we consistently choose the wrong one. Perhaps to borrow your analogy, it would be sort of like if, upon seeing a multiplication problem equal to twenty, we automatically assumed 2x10, and we perceived an illusion whenever the problem was actually 4x5.

We consistently choose the overwhelmingly more likely one. That's the reasonable thing to do, and it shows in the fact that these illusions take considerable effort to construct.

Our vision works by a projection on to a 2D surface in our eyes. This type of illusion does rely on having a particular vantage point, though...

We rely pretty heavily on 'synthetic aperture' reconstruction. Close one eye, look at a scene with objects at various distances. It looks flat - near and far are hard to distinguish. Move your head right and left (still one eye closed) - the objects jump out at you, nearer ones clearly 'in front' of further ones.

So the 2D retina is just the start - our neural processing jelly is reconstructing a 3D image from it continuously. This affects how we perceive the world drastically. Lots of room to fool the algorithms. Not because the hardware (eye/retina) is faulty; because using fairly simple and efficient hardware, our incredible neural processing elicits tremendous results!

There's a lot of language in this article that is sloppy at best. For example:

> The human brain routinely throws away many possible interpretations of the visual data it receives from the eyes. Given the brain’s limited resources and its need to interpret visual data quickly, it can’t afford to entertain every bizarre interpretation—it simply goes for the explanation that seems most likely, based on its past experiences and built-in visual processing machinery.

That makes it sound like determining the most likely interpretation is cheap, and doesn't constitute "entertaining" every interpretation.

Right. But I think any description would have to be vague given what little we actually know. If it sounded too precise, it would be giving the impression that we know more than we do.

There's a video game titled echochrome that touches on this phenomena. Its puzzle platformer where you have to navigate around obstacles by positioning the camera into various positions until a path forward or the obstacle is obscured by the illusion being projected. Very cool!

On that note, try Antichamber and Perspective, both of which toy very effectively with your perception of 3D spaces.

Also check out monument valley, they just released an expansion too if you've already played it.

There's also Naya's Quest, a free browser game by the creator of VVVVVV. It uses a diagonal isometric view and lets you rotate the camera around to see where the platforms really are. Even with that, it's tricky.

Reply on news.ycombinator.com