-
There’s already a lot about this in the media with regard to facial recognition technology, coded by white engineers and developed using white subjects.
-
If you can design an automatic soap dispenser that doesn't recognise black skin, you can be sure nothing bad will will happen if AI isn't trained with diverse inputs.
-
Those dispensers work on infra red. They’re not racist. Skin colour is irrelevant. It’s the position of the hand that’s relevant.
Facial recognition technology struggles with colour. Darker skin is harder to interpret so paler skinned people are more likely to be identified or recognised than darker skinned people. Not related to dark skinned people all looking the same. It’s the shadows / lack of contrast. -
Darker skin is harder to interpret so paler skinned people are more likely to be identified or recognised than darker skinned people.
Dark skin is harder to interpret for the current iteration of the technology. The gadget should never have left QA in that state, and the fact that it did tells you there were no dark skinned people around during testing. Talking about structural racism in these matters isn't unreasonable.
-
Racism always finds a way.
I reckon it'll be fine looking at what it's trying to pick up on, but the previous examples show up stuff that wasn't accounted for because of underlying reasons that people like me obviously hadn't considered, which was all that was being referenced. -
But facial recognition technology is not that good at identifying one of the main groups of people the police like to pick on has a sweet irony to it.
-
Those dispensers work on infra red.
Whilst this is correct.
Skin colour is irrelevant.
This is incorrect.
We know there to be a statistically significant delta in reflectance and absorbance in Photoplethysmography (ie what's used for oxygen saturation and drives the heart-rate data from your garmin/fitbit/Apple Watch) across skin-tones (https://www.nature.com/articles/s41746-021-00408-5 , etc.).
There is also a smaller, but still significant delta in the near-IR and IR bands, up till a wavelength over the 850nm that the IR dispensers use:
(skin color scale: https://en.wikipedia.org/wiki/Fitzpatrick_scale)
(reference: https://opg.optica.org/directpdfaccess/c0a24e54-ec7f-460e-b95000098ac41736_333549/ao-54-35-10559.pdf)The cause for this is- at least partially- Melanin's absorbance spectra, and so whilst no object can be racist, there is a difference in skin absorptivity, which has been- historically- overlooked.
1 Attachment
-
-
This is a problem that is more to do with racist police than racist algorithms.
- The two are closely linked. Where do you think the bad data came from?
- Having a machine learning system do this helps the police disclaim personal responsibility.
- ML systems are effective at finding the implications of the data and validation criteria supplied to them (hence racist algorithms even though nobody explicitly said "We want racist results") and then finding innovative ways to deliver the implicitly desired result.
In other words, the ML systems can significantly amplify institutional racism while enabling humans to evade responsibility. It doesn't mean muich to say "Racist police are the problem". ML adds new problems.
- The two are closely linked. Where do you think the bad data came from?
-
Every day is a school day. I read a bit more on the subject and found this.
https://circuit.bcit.ca/repository/islandora/object/repository%3A2043/datastream/PDF/download/citation.pdf -
So- there is a lot of work done on skin color interaction with wearables (this is at least part of my work currently)- it's just not that easy to solve, yet....
That paper above has a significant flaw, even if the outcome matches, due to additional overlooked (by the authors) complexities that aren't essential to the basic (well-understood) principle of darker= more aborptivity=less reflectance.
-
There's two parts:
The Printshop will use CMYK in combination to macroscopically approximate the same spectra as the skin spectra. At a Microscopic level- well above particle interaction level- this looks like:
Which means that you are probably not going to get the same reflectance properties as skin, where pigmentation is a concentration dependent deposition of a single pigment, with a single spectra.Diffuse reflectance and scattering. (essentially a change in amount reflected due to multiple biological molecules interactions)
Couple them, and you don't have as close an analogue to skin as you might hope.
-
.
2 Attachments
-
-
Murdering cunt invited to UK, so that PM can make money for his family.
Crown Prince Mohammed bin Salman has been invited to visit the UK, the Saudi Arabian embassy has told the BBC.
No 10 said it would confirm the prime minister's engagements in the usual way, with nothing in the diary yet.
But a UK government source confirmed the invite and said there was no reason to think the visit would not happen.
marcomarcos
Ben689908
dubtap
adroit
hippy
starfish&coffee
ric_a5
snottyotter
itsbruce
Dammit
Aldosterone
stevep
WillMelling
NotThamesWater
mespilus
Sumo-
Post a reply
About
In the news
Posted by
@Platini
Yeah, because it's generally trained on systematically racist data.