In particular, it is being applied right now to rank Google search results, and probably responsible for lots and lots of Google's profit. You should be skeptical of Google's appraisal of the paper that is material to Google's profit.
Meh, both parrots and language models are inferior to humans in producing language, but one is more useful than the other. And real parrots are also stochastic, like all living things.
After the stochastic parrots have caused the collapse of civilization, we can eat the real parrots.
I don't understand how living things are "stochasic". Can you please elaborate on the matter?
Gebru is the type of person who defines "bias" as anything that isn't sufficiently positive towards people who look like herself, not the usual definition of a deviation from reality as exists. Having encountered AI "fairness" and "bias" papers (words quoted because the words aren't used with their dictionary definitions), it's not even clear to me they should count as research at all, let alone be worth reading. They take as the starting premise that anything a model learns about the world that is politically incorrect is a bug, and go downhill from there.
All politics aside, this is not even true for toy ML problems. If I’m trying to do digit recognition and “all the data I can find” is a billion hand-written 0’s and a million hand-written 1’s through 9’s, naively training on that data will yield a model that’s pretty close to guessing 0 every time.
I think you're trying to say that a large enough dataset will be free of bias. I don't see how that follows. If I train a model on home mortgage decisions, I will replicate the bias on that currently exists on that dataset - https://news.northwestern.edu/stories/2020/01/racial-discrim... - unless there are conscientious choices to reduce that bias.
Researchers in ethics in ML are specifically trying to enable tech companies to do a better job of not replicating bias and justifiably point out where that is occurring.
Third, I would argue that applying an ML model to do something faster if it replicates the bias of a previously human decision is even worse. The bias has taken the human element completely out and systematized the bias and made it possible with even less friction.
> For example, in about 10% audits in which a white and an African-American auditor were sent to apply for the same unit after 2005, the white auditor was recommended more units than the African-American auditor. These trends hold in both the large HUD (Housing and Urban Development)-sponsored housing audits, which others have examined with similar findings to us, and in smaller correspondence studies
They fail to mention how large is the gap. Is the white auditor recommended 102 vs the black auditor 97, or 150 vs 50, or 200 vs 3? Without such critical information it is hard to form an opinion, unless one already has a large bias in accepting discrimination narratives uncritically.
> In the mortgage market the researchers found that racial gaps in loan denial have declined only slightly, and racial gaps in mortgage cost have not declined at all, suggesting persistent racial discrimination. Black and Hispanic borrowers are more likely to be rejected when they apply for a loan and are more likely to receive a high-cost mortgage.
They fail to mention the magic words 'when controlled for income'. America has a huge income disparity problem, which is conveniently forgotten behind the ongoing race (and gender) hucksterism. Assuming we'd wave a magic wand and fix all disparities across visible populations tomorrow, it will still not fix the fact that huge income disparities exist between individuals. Google engineers and researchers get paid 5 times the median national income or more, and (senior) Google management in the 10x to 1000x range. The vast majority of the population is stuck in dead end precarious jobs, with little social mobility, one medical emergency from bankruptcy.
I haven't seen this "usual definition" of bias as a "deviation from reality" that you give anywhere before. Can you say where it is coming from?
BTW, "Stochastic Parrots" is a very descriptive name for the problem
> Moreover, because the training datasets are so large, it’s hard to audit them to check for these embedded biases. “A methodology that relies on datasets too large to document is therefore inherently risky,” the researchers conclude. “While documentation allows for potential accountability, [...] undocumented training data perpetuates harm without recourse.”
Since these models are being applied in a lot of fields that directly affects the life of millions of people, this is a very important and underdiscussed problem.
I really want to read the paper.