大象传媒

Does your machine learning solution work the same for different types of people? You may think, why not. A machine learning algorithm uses data to build a model that identifies a phenomenon by its features. Say you want to go through a stack of听CV鈥檚 听and identify the most promising candidates. You can 鈥渢rain鈥� a machine learning model to do that, . But if you can accidentally end up with a model that gives听preferential treatment to one group of people. How do we mitigate bias and how much does bias mitigation cost in therms of the accuracy of our models.

This is the question that Knut T. Hufthammer, Tor H. Aasheim, S酶lve 脜nneland and H氓vard Brynjulfsen decided to investigate as part of their student project in the AI Ethics course in the spring this year. They turned to an open source toolkit for听bias mitigation,听 curated by IBM. Using the toolkit they built a try example of a hiring system using US census data and showed that bias mitigation has an impact on accuracy is mild. More importantly they learned the limitations of the state-of-the-art in bias mitigation today. The report of the project of听Knut T. Hufthammer, Tor H. Aasheim, S酶lve 脜nneland and 听H氓vard Brynjulfsen was published as a paper at the 听and was awarded the best paper award. While waiting for the proceedings you can read the paper .

In the same venue you can also read a from the spring course: an investigation of how much can the principles be satisfied in an automated Kindergarten place allocation system for the Bergen commune.听

If you are curious about AI ethics and want to try your hand at it, the department of Information Science and Media 大象传媒 will offer the INFO383 Research Topics in AI Ethics course next semester.听If you are interested in AI in general, the same department, in cooperation with the Department of Informatics will offer a Bachelor in AI starting in the autumn of 2021.