Our visual black box is trained on what we see in the real world. We don't process raw sensory data - we keep a more abstracted model in our heads. This is a blessing and a curse - it may either be more useful to have a sparser cognitive inventory, and work on a higher level, or to ignore preconceived notions and try to pick up details about What Actually Is. Marcus Hutter would have us believe this compression is itself identical to intelligence - certainly true for the representation once it hits your neurons. But as a preprocessing step? It's hard to be confident. Is that summarisation by our visual cortex itself a form of intelligence?

Ultimately, we reduce a three-dimensional world into a two-dimensional grid of retinal activations; this in turn gets reduced to an additional layer of summarisation, the idea of the "object". We start to make predictions about the movement of "objects", and these are usually pretty good, to the point where our seeming persistence of vision is actually a convenient illusion to disguise that our vision system is interrupt-based. And so when we look at a scene, rather than even being represented as a computer does an image, as any kind of array of pixels, there is in actuality a further degree of summarisation occurring; we see objects in relation to each other in a highly abstract way. Any impressions you might have about an object having a certain location, or appearance, only properly arise as a consequence of intentional focus, and your visual cortex decides it's worth decompressing further.

This is in part why blind spots (literal, not figurative) can go unidentified for so long: persistence of vision can cover them up. You're not really seeing the neuronal firings, it's a level lower than "sight". It probably isn't possible to do so consciously - we only have access to the output of the neural network, and its high-level features; these more abstracted receptive fields are all we ever really see.

This is what makes optical illusions work. Your visual cortex is pretrained to convert an interrupt stream from your eyeballs into summarised representations. And in doing so it makes certain assumptions - ones which, once correctly characterised, can be subverted with a particular adversarial stimulus. Each optical illusion shows its own failure of the visual cortex to properly represent "true reality" to us. The underlying thing they're pointing at, intentionally or otherwise, is some integrative failure. A summarisation by the visual system that, on further inspection and decompression, turns out to be inaccurate.

We can intentionally exploit these integrative failures, too. Autostereograms have us trick our visual cortex into the interpretation of a flat printed surface as having some illusory depth. Here, that interpretation is intended; if we were forced to see the world as it "truly is", such illusions would be impossible to detect. And so correcting all possible interpretation errors would have other knock-on effects - almost by necessity, there would be a family of phenomena that we would be incapable of detecting, phenomena which rely on the abstraction.

This isn't just "like" an out of distribution error in neural networks, it's the archetypal example of it as wired in biology. And so if you've ever wondered how those errors feel "from the inside" - well, go take a look at some optical illusions, and you'll understand a little better.

Are these errors a bad thing? Not necessarily.

People with grapheme-color synesthesia are arguably people with especially aggressive pattern-matching visual networks. When such a person looks at the above image on the left, it might appear similar to that on the right, in the same vague kind of way that other optical illusions might give the viewer an "impression of colour". And while identifying stealthy digit 2s amongst a pack of digit 5s might not be evolutionarily advantageous, you can imagine how other cases might be far more so.