4 ms·

What software that turns Obama white can (and can’t) tell us about AI bias

11 points by clouddrover 6y ago

thearchitect1 6y ago
I really don't see the problem. White men are the largest demographic of the US, so nueral networks will have a bias towards them. It really isn't that complicated.
- watwut 6y ago
  > White men are the largest demographic of the US That is super unlikely, considering that women are born slightly more often then males and women live longer.
  - kmm 6y ago
    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.
- djaque 6y ago
  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]. [1] https://www.vice.com/en_ca/article/889xyb/over-1000-ai-experts-condemn-racist-algorithms-that-claim-to-predict-crime https://www.vice.com/en_ca/article/889xyb/over-1000-ai-exper...
- jasongill 6y ago
  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.
  - thearchitect1 6y ago
    Yeah because nobody is using Asian datasets. Asians are less than 5% of the population in the US.
eddhead 6y ago
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.
- djaque 6y ago
  > 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]. [1] https://news.cornell.edu/stories/2019/01/study-ai-may-mask-racial-disparities-credit-lending https://news.cornell.edu/stories/2019/01/study-ai-may-mask-r...
JohnClark1337 6y ago
So go through the dataset and make sure that every race and gender has an equal amount of representation.

11 points by clouddrover 6y ago

I really don't see the problem. White men are the largest demographic of the US, so nueral networks will have a bias towards them. It really isn't that complicated.

watwut 6y ago
> White men are the largest demographic of the US That is super unlikely, considering that women are born slightly more often then males and women live longer.
- kmm 6y ago
  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.
djaque 6y ago
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]. [1] https://www.vice.com/en_ca/article/889xyb/over-1000-ai-experts-condemn-racist-algorithms-that-claim-to-predict-crime https://www.vice.com/en_ca/article/889xyb/over-1000-ai-exper...
jasongill 6y ago
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.
- thearchitect1 6y ago
  Yeah because nobody is using Asian datasets. Asians are less than 5% of the population in the US.

eddhead 6y ago

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.

djaque 6y ago
> 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]. [1] https://news.cornell.edu/stories/2019/01/study-ai-may-mask-racial-disparities-credit-lending https://news.cornell.edu/stories/2019/01/study-ai-may-mask-r...

JohnClark1337 6y ago

So go through the dataset and make sure that every race and gender has an equal amount of representation.