无忧视频

HMC CS Researchers Publish Chapter on Algorithmic Biases

无忧视频 computer science professor George Monta帽ez and his students Daniel Bashir 鈥�20 and Julius Lauw 鈥�20 have published the chapter 鈥淭rading Bias for Expressivity in Artificial Learning鈥� in聽ICAART 2020: Agents and Artificial聽Intelligence, part of the Lecture Notes in Computer Science book series (Springer, Cham).

The HMC researchers鈥� chapter about how bias relates to algorithm flexibility (expressivity) was an expanded and completely rewritten version of the lab鈥檚 award-winning 2020 paper for the International Conference on Agents and Artificial Intelligence (ICAART).

Monta帽ez, Bashir and Lauw expanded their original paper by beginning with a definition of the term 鈥渂ias.鈥�

鈥淭he word 鈥榖ias鈥� is a loaded term in machine learning and statistics, with at least four different uses,鈥� says Monta帽ez. 鈥淲e added a section differentiating the meanings of the term and showing how our particular notion of bias, 鈥榓lgorithmic bias,鈥� is not equivalent to the prejudicial biases we rightly try to eliminate in data science. While all prejudicial biases create algorithmic bias, not all algorithmic biases are prejudicial.鈥�

The authors also took advantage of having more time with their research to improve their presentation of the paper鈥檚 core ideas. 鈥淥ften when you present a paper, in having to communicate the ideas simply to an audience, you stumble upon a much better way of presenting your work,鈥� Monta帽ez says.

鈥淎lthough all of the theorems and definitions are equivalent between the original paper and book chapter,鈥� he explains, 鈥渢he extended version in the book introduces all of the key concepts around a geometric idea called聽inductive orientation, which is basically a direction an algorithm 鈥榩oints towards鈥� in high-dimensional space. The degree to which it points somewhere away from the baseline direction is the degree to which it can be algorithmically biased鈥攚e鈥檙e basically measuring how well-aligned an algorithm is with regard to a particular situation we care about. Furthermore, pointing towards one direction means pointing away from other directions, so we see that no algorithm can be well-aligned with all situations. This geometric idea of alignment paints a better intuitive picture of what we mean by algorithmic biases.鈥�

The original paper, 鈥淭he Bias-Expressivity Trade-off,鈥� co-authored by Monta帽ez, Lauw, Dominique Macias 鈥�19, Akshay Trikha 鈥�21 and Julia Vendemiatti 鈥�21, won the Best Paper award at ICAART 2020.

鈥淭his chapter stands essentially as a new paper, which builds on the content of the original conference publication, but improves it in many ways,鈥� says Monta帽ez. 鈥淲e were also fortunate to have Daniel Bashir join us as a co-author; he was responsible for many of the improvements in the new work, including the new section on different biases in artificial learning.鈥�

This publication marks Bashir鈥檚聽third and Lauw鈥檚 fifth with Monta帽ez鈥檚 AMISTAD Lab. In 2020, Lauw received a student researcher award from the Computer Science Department. Bashir was a 2020 CRA Outstanding Undergraduate Researcher honorable mention.

鈥淭he chapter will likely be used by machine learning and AI practitioners who are interested in new ways of looking at and measuring biases in artificial learning systems,鈥� says Monta帽ez. 鈥淗opefully it inspires greater transparency concerning the biases present in all learning algorithms.鈥�