Warning! You are probably being recorded (and transcribed).
Approximate overview
When do you anticipate grading stuff we turn in on Friday?
Things due on Friday and turned in on Friday will be returned by Monday.
Everything else will be in a queue of grading which will be one of Sam’s foci for next week.
My goal is that you get stuff back in time to fix it.
Munro, Dan. 2018. Feynman’s Error. https://www.danmunro.ca/blog/2018/11/29/feynmans-error-on-ethical-thinking-and-drifting-nbsp
Feyman was an American physicist who worked on the atomic bomb. The original decision to work on the bomb may have been ethical, but after we killed millions of people with the bomb, he realized that he should have reflected again.
Noble, Safiya Umoia. 2017. Algorithms of Oppression, Introduction. https://www.jstor.org/stable/j.ctt1pwt9w5
Google has always sucked, but in particular ways. For example, searches for certain kinds of people would give you pornographic images, searches for the nword + house would show the white house. Relying on the “knowledge of the masses” is dangerous.
Cahan, E.M., Hernandez-Boussard, T., Thadaney-Israni, S. et al. 2019. Putting the data before the algorithm in big data addressing personalized healthcare. npj Digit. Med. 2, 78. https://doi.org/10.1038/s41746-019-0157-2. https://www.nature.com/articles/s41746-019-0157-2
Machine learning algorithms are used in healthcare.
They can carry biases because of the data sets.
We assume that large data sets are unbiased, but they are not.
There are many categories of people that can be left out of big data sets, such as elderly, poor, rural, etc. Often, big data comes from people who use computers, and these groups of people may be less likely to use computers.
There are for V’s for data: Volume, Velocity, Variety, and Veracity. Too often, we focus on the first two, but the last two are more important.
Barocas, S. and Selbst, Andrew D. 2016. Big Data’s Disparate Impact. California Law Review, Vol. 104, No. 3, pp. 671-732. https://www.jstor.org/stable/24758720
Where we get data from can affect the quality of our algorithms. It might not capture the right information. (repeat for 60 pages)
More data sometimes helps.
Identifying discrimination in data sets can be hard.
The law has not kept up!
Side note: Even though there are known issues, many legislators will not push for changes until a “better” algorithm is available.
Ledford, Heidi. 2019. Millions of black people affected by racial bias in health-care algorithms. Nature. https://www.nature.com/articles/d41586-019-03228-6
Healthcare algorithms tend to devote less per-capita funding to black people.
The algorithm is, itself, using the amount typically spent on people to determine the care.
There can be many hidden factors in determining an ML decision, some of which are inappropriate.
Be careful about what measures you use to determine whether or not your “algorithm” is biased.
Bolukbasi, Tolga et al. 2016. Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings. In Proceedings of the NIPS 2016 Conference on Neural Information Processing Systems. https://proceedings.neurips.cc/paper_files/paper/2016/file/a486cd07e4ac3d270571622f4f316ec5-Paper.pdf
Machine learning algorithms can learn to associate words.
In doing so, they reinforce gender stereotypes.
It is possible to reduce these biases, but you have to pay attention.
We get biases even from seemingly unbiased sources, such as news articles.
Don’t trust your data.
Caliskan, Aylin et al. 2017. Semantics derived automatically from language corpora contain human-like biases. Science 356, 183-186. DOI:10.1126/science.aal4230. https://www.science.org/doi/full/10.1126/science.aal4230
Similar to the previous one.
These algorithms are good at picking up biases, particularly when looking at relationships.
It’s not just gender, its also race.
What responsibilities will (should) you consider as a computing professional?
What should CS@Grinnell teach about these kinds of issues? Where?
You’ve had to deal with a complex transition this semester. Congratulations on getting through.
My ideals of what you take away from CSC-301:
Autry’s goals:
Recent (and not so recent) research suggests that student evaluations of teaching are often influenced by non-conscious and unintentional biases about the gender and/or race of the instructor. According to this research, students systematically rate women of color, white women, and men of color (also older and less attractive people) lower in their teaching evaluations than young, attractive white men, even when there are no actual differences in the instruction or in what students have learned.
From those findings, we, the faculty of Grinnell College, infer that instructors may also be subject to bias based on additional characteristics unrelated to the effectiveness of their teaching. As you fill out the course evaluation and throughout the semester please focus on the effectiveness of the instruction.
Done!