Do you think Google, particularly its Fairness in Machine learning team, has adequately addressed the concerns that Umoja Safiya raises? (Keep in mind t hat the searches shown in the introduction to Algorithms of Oppression were conducted several years ago.)
Problem 1
Book: “some of the very people who are developing search algorithms and architecture are willing to promote sexist and racist attitudes openly at work and beyond, while we are supposed to believe that these same employees are developing “neutral” or “objective” decision-making tools.”
Google Talk: Talks about thinking about diverse users in mind, but not about who is choosing this diverse mind. They would then base this data on the group they picked. She said in order to do this fairly they have to grow an inclusive workforce and bring in diverse perspectives
Problem 2
Book: “I was overtaken by the results. My search on the keywords “black girls” yielded HotBlackPussy.com as the first hit”
Google Talk: Talked about Counterfactual Fairness and Equality of Opportunity (Examples from 2 groups should have equal likelihood of being classified correctly/incorrectly)
Google Talk: Google Translate: Translating a non gendered language (Turkish) to English. It translated i am a nurse to she is a nurse and i am a doctor to he is a doctor.
Solution: Instead of making the model how to gender. They gave users both masculine version and feminine versions of the sentence being translated
Problem 3
Book: “This book was born to highlight cases of such algorithmically driven data failures that are specific to people of color and women and to underscore the structural ways that racism and sexism are fundamental to what I have coined algorithmic oppression.”
Google Talk: To me it was pretty broad in saying get a better data set to train models
Google Talk: Gender Shade - a study that was done to look at gender classifiers. Take an image and try to tell if the user is male or female. The error rates for darker females was the highest
Solution: actively went about trying to collect more data. Sought to get more data to represent darker females to train the ML
Problem 4
Book: “The first problem for Google was that its photo application had automatically tagged African Americans as “apes” and “animals.”
Google Talk: Talked about how they give their MLs data to learn from so maybe the data they were given was riddled with people being racist. Their solution was to actively go out and collect data, generate synthetic data, or model could algorithmically look for where errors are occurring and search for that data.
Google Talk: Or if they don’t want their model to make assumptions based are identity terms, then it should not be shown it. Stereoypes in Translation
Problem 5
Book: “The second major issue reported by the Post was that Google Maps searches on the word “N****r” led to a map of the White House during Obama’s presidency, a story that went viral on the Internet after the social media personality Deray McKesson tweeted it.”
Google Talk: I didn’t think she addressed this or something like it
Problem 6
Book: “There is a missing social and human context in some types of algorithmically driven decision making, and this matters for everyone engaging with these types of technologies in everyday life.”
Google Talk: Build with more diverse teams.