Silicon Valley doesn't care about poor people: Top AI models kinda suck at ID'ing household stuff in hard-up nations Off-the-shelf object-recognition systems struggle, relatively speaking, to identify common items in hard-up homes in countries across Africa, Asia, and South America. The same software performs better at identifying stuff in richer households in Europe and North America. Though initially shocking, and then not so much when you …
When you go into a radically different society then a lot of 'common household objects' are different to the ones we're used to (and they're often used a bit differently). We wetware types can make intelligent guesses about what we think stuff is because we don't have any nagging probabilities telling us we're wrong. We'll also get corrected by the locals. Software, OTOH, isn't really smart, its just like a bureaucrat with a really big Code of Practice book.
The truth is that rich people in poor countries don't live like the rest of their compatriots, they live like Americans -- or, at least, how they think most Americans live. So obviously made-in-'merka image recognition software will work. But you'll also find that there are a lot of really poor people in the US and their modes of living might surprise both middle class (white) humans and software. It actually wouldn't hurt for people to get out and about a bit more....
It might also be due to the fact that a lot of people have better things to do than rush around taking photos of household items, regardless of where they fit in the socio-economic spectrum (or their skin colour for that matter, I'm going to give you a down vote for mentioning that) - as an example, I'm paid well above average in NZ, I'd even be considered to be paid above average in the US, I own several vehicles, a large house and a beach house with no debt financing on either of them, etc., but I still don't feel the urge to take photos of my soap dish and send them in to some database.
Somehow, there just seems to be more important things to do in life and I suspect I'm not alone in this.
It hadn't occurred to me until you mentioned it, but how representative are their sample images anyway? Even the pump soap dispenser in the sample image is nicely isolated on the sink instead of crammed in between the toothbrush holder and a bottle of lotion.
Do they have images tagged "sink" that include towels/rags hanging on the edge? "Faucet" images that have dishrags hanging off of them? "Towel" images where 3 hand towels of different colours and sizes are crammed together on a small rail?
Even if you somehow provide enough incentive for people to submit "real world" example images for your training set, how much do you want to bet they tidy up a bit beforehand? And if they're scraping their samples from the web, well, nobody is posting anything remotely realistic there...
regardless of where they fit in the socio-economic spectrum (or their skin colour for that matter, I'm going to give you a down vote for mentioning that)
Well, there are the problems then as some of my friends are "of color" shall we say and there are objects in their homes that I'd be hard pressed to identify without knowing the people. Facial recognition also has issues with non-whites due to the data bases and who compiled them.
"Well, there are the problems then as some of my friends are "of color" shall we say and there are objects in their homes that I'd be hard pressed to identify without knowing the people."
Wow! - are you really saying that you go so far out of your way to be "friends" with people who have a different skin colour to you, that there are basic domestic items in their homes that you don't recognize - and this would not apply to some of your "friends" who are of the same skin colour, but a different cultural background eg., Amish?
Apart from coming across as being somewhat condescending, that does suggest a woeful degree of ignorance on your part.
Ok, I'll agree that many people might have something pertaining to their particular cultural background hanging on a wall or scattered around for either functional or decorative purposes, and I might not know exactly what their traditional name for that item is, but I can still recognise what is essentially a knife, a drum, a wall hanging, a harness for a horse and buggy, etc. I might even not know the names of the traditional foods they might eat, or what those foods are made from, but even if seen in isolation I doubt very much I'd ever get confused between food and a bar of soap.
The truth is that rich people in poor countries don't live like the rest of their compatriots, they live like Americans -- or, at least, how they think most Americans live.
Actually, although the richest of Americans have more disposable wealth, as do many Pacific Rim Asians [ yet the upper stratas in Europe also have astounding lifestyles, but less perceived ] it's rich Americans who ape European styles, and are criticised for it by their down-to-earth, more common-man, thick-as-two-planks, plain fellow Americans --- rather than the rich elsewhere copying Yanks.
Which is natural as the USA, until the founding nations become much less represented in the 22nd century, will have just been an off-shoot of European culture.
My soap (in the UK) looks much more like the one attributed to Nepal in that picture. Can't stand those bottled ones: even if they don't stink of perfume, I find them impossible to rinse off (though that wouldn't be such a problem with the harder water much of Blighty - especially the rich bits - has).
More seriously, I expect you're comparing standardised consumer goods owned by rich people with home-made stuff in poorer countries. I wonder how it'd fare with frugal make-do-and-mend folks in rich countries?
The best soap should have a pH below 6, which rules out the standard stuff, which leaves my skin dry and prone to breaking (I have to use rubber gloves for the washing up). Unfortunately, most of the liquid soaps also have all kinds of shit that you don't need along with the perfume that many consumers insist upon.
In many situations water is sufficient to remove whatever is on the skin. Where it is not, I do find the pH 5.5 soaps okay but have also used olive oil to remove oil and grease: will be absorbed quite quickly or can also be washed off reasoably easily.
I was also a bit puzzled by the liquid soap thing. Is it really that common in the UK? I've only ever really noticed it in office/public loos. A bottle of liquid soap just for hand washing appears to not only be significantly more expensive than a bar of soap, but provides for far, far fewer washes. Is that extra cost really worth the tiny little increase in convenience?
In a public or office setting, there's the possibility of cross-contamination with strangers, but in the family home, you are all sharing your germs in so many ways already (which is beneficial in so many ways) that the one benefit of liquid soap is negated.
Not to mention all that waste plastic from the bottles, the mixed plastics due to the pump part and that few people probably bother to separate the pump from the bottle or flush the pump clean before putting it in the recycling bin, hence "contaminated" and unrecyclable.
Liquid soap is very common in Germany, probably because it's less messy than bars of soap. Fortunately, you can get refills in thin polythene containers. But then, seeing as one bottle lasts me more than a year i'm not that bothered about the plastic bit. As you note, recycling isn't that easy and currently not really financially viable. So most of the waste plastic here is also burnt as in combined plants, and the waste is less than one might imagine.
While "pots and pans" can be used for food storage, they are not a refrigerator, so why expect an AI or person to classify them as one when they are being used to store food?
Then there's the difference between a "bed" and a "guest bed"... any guesses? (does Rule 43 apply?)
Personally, I'm now developing a strategy of storing spare batteries in the library and sleeping in a hammock so the robots run out of power and can't find me to murder me in my sleep.
Surely it's just an illustration of the consequences of hand-feeding a learning machine : it will only learn what the developers provide it. Consequently it will only know what they know (or think is worth knowing) and so be biased towards their culture and level of ignorance of other cultures.
It's a problem that will only be resolved when they give the things legs (or wheels) and passports and let them travel and learn what the world is really like outside the lab.
"Surely it's just an illustration of the consequences of hand-feeding a learning machine : it will only learn what the developers provide it. Consequently it will only know what they know (or think is worth knowing) and so be biased towards their culture and level of ignorance of other cultures."
Exactly what I was thinking. Nothing shown here appears to have anything to do with money. Presumably the system would be just as bad at recognising a bar of soap here in the UK, and if anything it's richer middle class types who are more likely to use it instead of pointless disposable bottles. It's simply that the system was developed in one of the richest parts of the world, so it's almost impossible for things it doesn't recognise not to be correlated with lower income. It once again exposes the problem with poorly trained recognition systems that are unable to handle things the developers didn't think of, but it doesn't say anything about Silicon Valley's attitude to poor people.
Of course, it also seems to expose just how terrible these systems are even when supposedly working. In the example shown in the article, only two managed to recognise a bottle of soap, and one of those still thought it was more likely to be a gas tank. They might have been even worse for the Nepali photo, but all except Tencent were completely useless for both photos. Given the point is to recognise the contents of images without human intervention, what this really shows is not that there's bias based on money or culture, but that these systems currently are simply not fit for purpose anywhere.
And there's also the slight drop-off at the very highest income bracket. Maybe the Goop effect? "No no no, we don't use your liquid hand soap, we have a $300 body purification system that harnesses positive ions to gently coax negative particles and lost dirt off of your skin without damaging it."
Actually, if you search the Goop website for "soap", you end up with two liquid hand soaps, eight bar soaps, and three soap-possible/soap-adjacent products.
'A refrigerator in more developed countries have doors, and are made of stainless steel or are painted white, whereas in less developed countries where electricity is scarce, pots and pans are used to store food. Image-recognition models, therefore, won’t know that these simple storage objects are, in fact, refrigerators simply because they haven’t been taught that during the training process beforehand.'
Hang on how is a pot or a pan a fridge?!
Who really cares if an AI model can recognize these household items? It's not like there is going to be an AI self driving car travelling through the house, deciding to avoid the hand soap in a rich household, but plowing through the bathrooms of those who can only afford soap bars.
Reading the post I noticed that only 30,000 photos from 264 homes were used for training. This strikes as too small a sample set of the homes even in the US let alone the world. The results underscore a well known problem with computers 'Garbage In = Garbage Out'. A much wider variety of situations must be carefully documented (probably by hand) before being fed into the idiocy system. Poor quality data, for whatever reason, means the analyses based on the system are at best very limited and more likely basically useless for their intended purpose.
Terrific article! It's ironic though that much of the work of labeling data is conducted in India. Labeling is important because that's the only way so-called AI "learns anything". In other words it's all pattern recognition all the time with nary a semantic in sight. Here's a reference:
https://datadecisioning.com/?brick=ai-works-because-human-trainers-india
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
