[jl6]

> Models trained on low-pass filtered images maintained high performance even for highly degraded images. More strikingly, models that were trained on high-pass filtered images maintained performance well beyond the point that the degraded images contained no recognisable structures; to the human coauthors and radiologists it was not clear that the image was an x-ray at all.

What voodoo have they unearthed?

[proto-n]

I tend to not believe unbelievable results in machine learning. It's too easy to unintenionally cause some kind of information leakage. I haven't read the paper in detail though, so their experimentation setup could be foolproof, this is not a critique of this paper specifically.

[sidewndr46]

This reminds me of the ML research that could predict sex from an iris. It turns out they were using entire photos of eyes to do this. There are so many obvious cues to pick up on in that case, like eyeliner, eyelashes being uniform (or fake), trimmed eyebrows, general makeup on the skin, etc.

[dragonwriter]

> I tend to not believe unbelievable results

That seems tautologically true.

[JumpCrisscross]

> What voodoo have they unearthed?

Curious for the take not of a neuro-ophthalmologist. If they too are stumped, this may be a path to a deeper understanding our visual system.

Simple transformations obviously discernible to us blind computer vision. (CAPTCHAs.) There may be analogs for human vision which don’t present in the natural world. Evidence of such artefacts would partially validate our current path for artificial intelligence, as it suggests the aforementioned failures of our primitive AIs have analogs in our own.

[civilized]

I think there's a significantly greater than zero chance that they simply botched their ML pipeline horribly and would get their 0.98 AUCs from completely blank images.

[6gvONxR4sf7o]

I think it’s pretty straightforward. Imagine the fourier transforms of some recognizeable audio signals. Maybe a symphony and a traffic jam. They’ll look totally different, even to the naked eye. If you chop off the low frequency components, you can still probably tell which fourier spectrum is which. But now do the same thing in time domain (high-pass filter the audio). It probably won’t be clear that you’re listening to a symphony anymore.

[Der_Einzige]

... Adversarial examples.

It's a whole field of research, and it's pretty trivial to generate them for most classes of ML models. It's actually quite difficult to create robust models that DON'T have this problem...