I believe this illusion is exploiting a quirk of human sensory vision so any machine learning will not be affected. It's the same property that most video encodings and compression algorithms exploit: Our colour and luminance perception does not have equal fidelity, i.e we dont sense as much detail in colour, where as we can sense more detail in luminance... in this case it means we can perceive the stripes, and given stripes alone we can just about discern their colour, but when mixed with a background colour it effectively blends the colour without blending the luminance.

I think this might be the same biological reason that we can't perceive much (or any) colour in low light, rods vs cones.

TL;DR our retinas are not CCDs they are a patchwork of two discrete components.

>I believe this illusion is exploiting a quirk of human sensory vision so any machine learning will not be affected.

That's the point my dude.

Humans see an optical illusion, robots do not. Meaning ML algos that care about seeing things as humans do have a conundrum.

That makes no sense. It's all about the training.

A human can tell ML that the balls in the example are either perceived red (illusion) or reality (brown). The computer will use that training model for future decisions.

Why wouldn’t the robots see it?

Unless your camera is a 3CCD camera, then CCD cameras usually have lower color resolution than luminance resolution, too!

https://en.wikipedia.org/wiki/Three-CCD_camera https://en.wikipedia.org/wiki/Bayer_filter

if you label it as such, ML algorithms will learn it as such, no ?

aha, yes i suppose, learning human fallacies, that would be quite confusing when it's perceiving something quite different to us, it would probably take more effort to learn compared to just manipulating the input image.

This kind of learning bias has already been shown to be eeked out. The 'criminal' classifier that was used in one of the US states (kentucky or tennessee i believe) for sentencing and parole decisions, was shown to carry forward the human biases from its training data. ML is only as unbaised as its training set.

iirc, james-mickens has a usenix presentation about the dangers of using ‘ai’ in this sphere of human activity.

You make a good point, though different people perceive optical illusions differently. Remember the black/blue/gold/white dress? What colour should an AI call the dress when humans can't even decide? So it's actually even more complex than you might think.

This question also applies to audio illusions, such as the (in)famous Laurel/Yanny clip.

"Brown" is already going with "what humans would think", from the sensors that try to capture the image according to the peculiarities of human vision, to whatever representation the system has of the social construct of "brown".