Elijah Weinberg
Pronouns: He/Him
Research Mentor(s): Efren Cruz Cortes
Research Mentor School/College/Department: Michigan Institute for Data Science / Other
Program: UROP
Session: Session 2 (10:00am – 10:50am)
Authors:
Abstract
The study of cognitive science and its connection between mathematical structures and mental representations is new and underdeveloped. Recently a connection between this study and personal bias and algorithmic design in machine learning(ML) has been made. Bias like this have been identified in machine learning used for healthcare, court hearings, language, and more. In this research project, after familiarizing ourselves with ML and cognitive science, we identified one of the contributing factors to ML implementations and software bias to be biased data. ML is created by modeling the software on past data; if the data used is biased, the software will keep implementing that bias. A good example is a software used by judges in the American judicial system; this software is trained with past police records which contain racial and socioeconomic bias. As a consequence, it reproduces results that are more likely to negatively impact people of color and low-income individuals. With all of this knowledge, we set up the following experiment: design code that contains ML bias: survey the participants (CS engineers) to identify cultural backgrounds and demographics: have participants analyze and edit the biased software (Randomized Control Trial): record the results of the study. We will need to find a high level CS engineer to produce the software for us and garner experienced CS engineers to be participants. We hypothesize that the CS engineers won’t be able to sufficiently identify the bias in the data. We will then use statistical tests to confirm our hypothesis. We expect to answer three questions with our results: 1) What sort of bias is easily recognized and which isn’t? 2) In what order is the bias addressed when the participants are altering the code? 3) Is there a connection between cultural backgrounds and which bias was identified in the code?
