Algorithms are the drivers of decision making in machine learning. They are made of data provided by humans. It is known that humans are biased and error-prone, but this is also the case for algorithms. Both humans and algorithms make decisions based on available data and experience. If the data provided to algorithms or the way it is developed is biased, it could cause algorithm bias.
Interaction bias is one type of bias that is frequently found in datasets. An example for this bias is facial recognition. A study by Buolamwini (2018) found that datasets for facial recognition are mostly composed of lighter-skinned subjects. The study shows that algorithms perform better in identifying lighter-skinned men compared to darker-skinned woman. As a result, this could cause misidentification or no identification at all. This weekend Zoom encountered problems with their algorithm because of the interaction bias. A black student’s head was removed from the Zoom meeting whenever he would use a virtual background. Another issue caused by this bias was in 2019, when a student from Brown University was mistakenly identified as a suspect in Sri Lanka bombings. This resulted in continuing death threats to the student.
As mentioned above, algorithmic bias can have serious consequences. Some companies are already responding to this problem by abandoning their facial recognition products. For example, Amazon put a one-year pause on its facial recognition product, Rekognition, a few months ago. IBM also responded to this problem by terminating all their research on facial recognition. It is clear that the consequences can be severe, but what do you think is the best way to manage algorithm bias?
References
Buolamwini, J. (2018) Gender Shades: Intersectional Accuracy Disparities in Commercial Gender Classification. Retrieved from: http://proceedings.mlr.press/v81/buolamwini18a/buolamwini18a.pdf
Heilweil, R. (2020) Why algorithms can be racist and sexist. Retrieved from: https://www.forbes.com/sites/cognitiveworld/2020/02/07/biased-algorithms/#320881da76fc
Brown, A. (2020) Biased Algorithms Learn From Biased Data: 3 Kinds Biases Found In AI Datasets. Retrieved from: https://www.forbes.com/sites/cognitiveworld/2020/02/07/biased-algorithms/#6ab8acb376fc
Ivanova, I. (2020) Why face-recognition technology has a bias problem. Retrieved from: https://www.cbsnews.com/news/facial-recognition-systems-racism-protests-police-bias/
Dickey, M. (2020) Twitter and Zoom’s algorithmic bias issues. Retrieved from: https://techcrunch.com/2020/09/21/twitter-and-zoom-algorithmic-bias-issues/
