[toxik]

Which is inherent to the problem. The domain is the space of _all possible natural images_, it’s so big, it’s ridiculous. Fundamentally, these techniques do not analyze images like humans do, but are rather trained to pick out any salient signal it can latch on to. That seems to be “primates are mostly like humans but darker”, which is superficially true but a pretty weak definition as it includes dark skinned humans.

[pvaldes]

Primates are not darker than humans. Primates is the Order, they came in all colors, including brilliant blue, white with black stripes, orange, yellow or red.

We live in a world when everybody is offended by nothing. The problem is that nobody should be offended by being called a Gorilla. In the same way as nobody is offended by being called an eagle or a wolf. Is a wonderful animal, smart, strong, protective and gentle. What if some idiots used the term pejoratively five generations ago? We know better. Societies can change.

If white people is not being classified as primates, the algorithm should be corrected so they are. Not fixed excluding black people from humanity.

People should be educated also to understand that an AI algorithm is returning probability, not truth

[wizzwizz4]

It's probably because humans are primates – but the AI systems often have to treat “human” as a completely separate category as “primate”, so they have to draw weird, complex boundaries around “primate” (actually “all non-human primates”). When the “primate” classification is stronger than the “human” classification, the system says “primate” rather than “human”, and if it's predominantly been trained on “pictures of white Americans are not pictures of primates”, its “primate” definition might not be skewed to miss everyone else.

I expect you'd get better results if you allowed the system to call humans “primates”, then accept “human primate” as “human” when parsing the output. (That is, leave the “is_primate” output line floating while training on pictures of humans.) I don't know whether that would work, though.

[toxik]

These models typically don’t have hierarchical labels like that, and they apply a softmax to their output - which means /one/ label will be considered correct. (A softmax means taking the exp of your predicted scores, then divide by the sum.)

[wizzwizz4]

I know – but there's no technical reason they shouldn't have more complex relationships between labels (assuming you can train that, which I don't know). If we can't get better training data, at least trying to fix the problem at the algorithms (instead of slapping crude filters on the end of them) would be nice.

[toxik]

I agree in spirit but disagree in practice, I think. Like we said previously, the domain is humongous so even establishing meaningful relationships between labels and sublabels is extremely difficult. Many cases are likely ambiguous too, our human understanding isn’t actually hierarchical - it’s much more elusive. It’s a square peg round hole type problem really, humans don’t really think in terms of labels in the first place, we mostly use them for the purposes of language.