A Truly Unbiased Model

Опубликовано 16 сентября 2021, 23:51

Any dataset has bias, as the data collection will be intervened by human bias and affected by the nature’s biased law. Therefore, any machine learning model based on it will be biased and limited in generalization. Of course, we have a plenty of “debiasing” techniques that deliberately remove the bias of the “majority”, so as to take care of the “minority” --- fairness is believed to be achieved. However, such “taking care of minority” per se is a bias unfair to “majority”. In this talk, I will share a “best of two worlds” model that is truly unbiased. Its design is simple, yet effective and theoretically grounded.

Speaker: Hanwang Zhang, Nanyang Technological University's School of Computer Science and Engineering

Microsoft Research Deep Learning team: microsoft.com/en-us/research/g...