The sex ratio at birth is actually slightly biased towards men, it takes a few decades to even out. Globally, there are also more men than women, but that's partly due to China and India being artificial outliers, in most countries, women outnumber men.
The article highlights dataset biases that reinforce existing biases in society. It becomes a problem when ML affects the real world. See e.g. predictive policing [1].
White women are the largest demographic in the US, but the US is only a small fraction of the world population. There are more Asian women in the world than there are men & women in the US and Europe combined, but this software fails miserably on Asian women as shown in the article.
It's not a AI bias problem, it's a dataset problem. The issue isn't the algorithm neither are they selling this as a de-facto accurate Business software. You can take it and train the neural nets with whatever dataset you like.
> It's not a AI bias problem, it's a dataset problem
I would argue that it's beside the point. The argument is that the ML systems we generate now and are looking to use for real world systems have biases in them. For example, you do not want a lender auditing system used by banks to incorporate features that end up being proxies for race because the data you used to generate it was biased [1].
9 comments
[ 4.8 ms ] story [ 34.2 ms ] threadWhite men are the largest demographic of the US, so nueral networks will have a bias towards them. It really isn't that complicated.
That is super unlikely, considering that women are born slightly more often then males and women live longer.
[1] https://www.vice.com/en_ca/article/889xyb/over-1000-ai-exper...
Asians are less than 5% of the population in the US.
I would argue that it's beside the point. The argument is that the ML systems we generate now and are looking to use for real world systems have biases in them. For example, you do not want a lender auditing system used by banks to incorporate features that end up being proxies for race because the data you used to generate it was biased [1].
[1] https://news.cornell.edu/stories/2019/01/study-ai-may-mask-r...