Underrepresentation is often due to human bias. Occasionally deliberate prejudice, but more often careless assumptions.Jdsk wrote: ↑22 Jan 2022, 9:51amUnderrepresentation in the training set is a different problem from importing bias from human behaviour. (And of course there's a risk of the latter, see next post.)
And we know that underrepresentation can occur. It's been observed in AI with facial recognition and with forensic use of DNA analysis. (And it also commonly occurs in clinical trials for both the majority sex and ethnic minorities.)
Jonathan
Teslas can be programmed to break the law.
Re: Teslas can be programmed to break the law.
Re: Teslas can be programmed to break the law.
Even if that occurs in recruitment of the training set it's a different problem from bias in the AI processes themselves.Bmblbzzz wrote: ↑22 Jan 2022, 3:36pmUnderrepresentation is often due to human bias. Occasionally deliberate prejudice, but more often careless assumptions.Jdsk wrote: ↑22 Jan 2022, 9:51amUnderrepresentation in the training set is a different problem from importing bias from human behaviour. (And of course there's a risk of the latter, see next post.)
And we know that underrepresentation can occur. It's been observed in AI with facial recognition and with forensic use of DNA analysis. (And it also commonly occurs in clinical trials for both the majority sex and ethnic minorities.)
And it can be genuinely difficult to collect enough data from small minorities. (As it often is for the majority sex in clinical trials.)
Jonathan