AI needs Diversity to reduce Gender and Racial Bias!

Artificial Intelligence is the new electricity, powering the technological revolution just like electricity enabled, believes Coursera co-founder Andrew. However, AI has a significant gender and racial bias.

MIT discusses how computer vision is great at recognizing light-skinned males but not good at recognizing darker females. The ability of computer vision algorithms to recognize dark-skinned females is 20%- 34% poorer than its ability to recognize light-skinned males.

Research by the University of Colorado Boulder highlights the difficulty in identifying transwomen and transmen. 

In the research paper, Diversity in Faces by IBM Research AI, the authors highlight that most computer vision training datasets are predominantly focused on light-skinned males. Light-skinned people constitute between 80% to 95% of the images in most training databases. The datasets are also predominantly male. Historically as well, camera manufacturers have focused on light-skinned people and paid less emphasis on capturing other skin tones appropriately. 

These results in computer vision algorithms inappropriately classify a throwback image of the former First Lady of the US (FLOTUS) as “a young man wearing a black shirt”! Why? A mere 2.5% of Google employees are Black, as per its 2018 report! Women are also under-represented, comprising around 20% of the workforce in big tech companies as per a report by Bloomberg.

Source: MIT, 6. SI9I. Introduction to Deep Learning

Like many other spheres in life, we need more diversity in AI. We need to actively promote people from diverse and under-represented backgrounds to join and share their views on the development of AI. Otherwise, needless to say, AI will remain biased in terms of culture, race and gender. And like a recent pop song, we’ll be left complaining “Tuada Kutta Tommy Sada Kutta Kutta.”