When I think of "biased training data", I think of data that systematically misrepresents reality. Like say taking the prevalence of news reports as a proxy for the prevalence of what's being reported on.
In contrast, the AI ethics people seem to use that term to mean data that doesn't reflect just (as opposed to unjust) causality.
I think there's a deeper issue, which is summed up in the saying "when a measure becomes a target, it ceases to be a good measure": simple correlation-finders - which is what current ML/AI models are - are not fit for the purpose of building control loops. And the "ethics" people are trying to paper this over by manually fixing up all the correlations that people notice aren't (or shouldn't be) casual.
The reason for that is almost entirely the same as your example of biased training data ("taking the prevalence of news reports as a proxy for the prevalence of what's being reported on").
CEOs from large US companies get reported on the most in English language press. Therefore searching in English for "CEO" finds those people.
The reason for that is almost entirely the same as your example of biased training data ("taking the prevalence of news reports as a proxy for the prevalence of what's being reported on"). »
But the CEO essay / example isn't about making search results reflect whatever the real distribution is. It's about making them match arbitrarily chosen statistics explicitly in an attempt to influence how people respond to it.
.
From the essay: "However, these datasets reflect the biases of the society in which they were created and the systems risk re-entrenching those biases.". That's not talking about trying to correct for systematic error a dataset that doesn't reflect what CEOs actually look like, that's talking about CEOs actually not looking like what they should look like; and about manipulating results in support of the social goal of changing what is to be closer to what they feel should be.
Which isn't even really about AI. It's about the proposed feedback loop (CEO demographics) -> (what people think of when they hear "CEO") -> (what kids want to be when they grow up) -> (CEO demographics) and deliberately manipulating that first step in pursuit of a specific social goal. The use of AI is incidental.
Actually I think there are lots of articles about female CEOs, as female CEOs will be more likely to be written about than male CEOs ("the top female CEOs fo 2021" and so on, there are bound to be countless such articles).
No it's not, and if you think that you are misreading it.
The general point of ML ethics is that a machine learning system should not reinforce bias caused by historically biased training data. This is good statistical practice, but unfortunately often gets lost in the "magic AI" bubble.
Avoiding this is like saying "a salary of $110K/year makes you a rich person". It's sort of based on something true (it does put you in the the top quartile of US earners), but is completely useless without correcting at least for location and number of dependants.
That's what ML Ethics is about.
> I've seen smart tech people unable to say that men are on average taller than women because (I believe) they were afraid to say that some "favorable" outcome (tall) was a result of some component that could be genetic.
So you've taken that dumb argument and said all ML ethics arguments are the same?