| 2021-12-16 |
December 16, 2021 – This past October, in recognition of Black History Month, I was delighted to be invited to participate in a Black History Month panel, Celebrating Black History Month in AI, organized by the UCL AI Centre. For the event, I was a panelist in a session entitled "How to use AI to benefit the Black community." The aim was to consider existing projects and to elicit discussion on ways to resolve existing structural problems that the Black community continues to face. This prompted meaningful conversations on bias, in particular how to mitigate dataset biases. This is particularly pertinent to me with respect to some of my current research here at Sony AI. In my presentation, I gave a brief introduction to my background and how I first came to be working in the field of AI, followed by an outline of my current research projects and how they relate to biases in visual datasets. Lately, I have spent a considerable amount of time thinking about biases in people‑centric computer vision datasets. Specifically, how datasets can be made to be more inclusive and diverse, as well as current impediments to this. During my talk, I alluded to the fact that datasets are often subject to latent historical and representational biases, reflecting social inequities and representational disparities. Irrespective of their intent to faithfully depict the world, they render very narrow, discrete views of the visual world. What I mean by this is that despite best efforts to make datasets objective representations of the world, they are invariably subjective. Just think about the narratives and reasoning behind the ‘impartial’ taxonomies that we employ to categorise people. As the philosopher Thomas Nagel suggested, it’s impossible to take a ‘view from nowhere’. Therefore, we need to make datasets more inclusive by integrating a diverse set of perspectives from their inception. Datasets are invariably shaped by problems that researchers and practitioners wish to resolve. However, datasets are not only informed by the perspectives of their developers, but equally by those who create the data samples, as well as those tasked with annotating them. For example, when annotating people in image datasets, if we have a homogenous group of annotators, this risks perpetuating harmful social stereotypes. The perception that people have of others is undoubtedly shaped by their own cultural background. This is particularly the case for people’s ability to perceive skin colour. In 2002, researcher Mark Hill found that Black and White annotators perceived greater color variation within their own race. The highlight of the event for me was the keynote talk by Professor Chris Jackson Jackson, who is a British geologist, spoke on the topic of Race, Racism, and Representation. |
