Artificial Intelligence is sometimes heralded as a panacea, able to improve and automate technology in a way that will solve many of our social, business and wellness issues. But AI also has a well-documented diversity problem, one that can limit the potential of its usefulness or – worse – amplify the implicit biases that are present in our world today.
A recent report from the AI Now Institute found that 80% of AI professors, 85% of AI research staff at Facebook, and 90% of those staffers at Google are male. Further, people of color make up only a small fraction of staff at major tech companies.
This shortfall in diversity can lead directly to shortcomings in the resulting technology. For example, it was recently revealed that Amazon had previously ceased internal trials on an AI-powered tool for evaluating potential hires because it was unwittingly slanted against women. The more women and minorities on these AI research and development teams, the more powerful and robust the technology they will produce.
Pooja Rao is the co-founder and head of R&D at Qure.ai, a healthcare startup using AI to identify abnormalities in X-Rays and CT scans. She says she is very concerned about the lack of diversity in the field of AI, especially when it comes to healthcare. She is trying to raise awareness and promote diversity while growing her business to ensure that new AI-powered tools best serve all the patient populations they are meant to support.
Rao helped me understand how AI can be an important ally for physicians and medical providers when it comes to testing. Today, she said there are more than 280 million imaging tests performed in the U.S. every year – including X-Rays, MRIs, ultrasounds, and more – but only 40,000 radiologists to read them. One result can be inaccurate patient diagnosis. AI can help overcome this shortage by automating that process to improve accuracy for all patients.
AI is the umbrella term for any technology or algorithm that can replicate some aspect of human intelligence. Like the IBM Deep Blue program that beat the chess grandmasters in a popular publicity ploy in the 1990s, these technologies are meant to perform the same computations as the human brain, only faster and more accurately.
Today, AI has become infinitely stronger using sophisticated deep learning algorithms and with increased computing power. It can now visualize and interpret images exactly as a human would and can be trained to recognize specific objects within those images to compensate for America’s lack of radiologists.
Rao was originally trained as a medical doctor, not as a technologist. She first began applying AI during her PhD at the Max Planck Institute, using it to help decode large amounts of genomic data to predict the onset of Alzheimer’s disease. Excited by the prospect of machine learning to aid research and discovery, she taught herself to code in Python so she could process large datasets at scale. After graduation, she began work in data science and bioinformatics roles, gradually morphing from a clinician learning about AI to an AI scientist working on clinical problems.
Rao views AI today as complementary to healthcare professionals, rather than a replacement. Her company is building AI to help take over routine tasks in order to free practitioners to spend more time with patients, fill the gap left by shortages of professionals, and safeguard against errors.
She is ever-mindful of the danger in not giving her AI tools a diverse and balanced deep learning experience. She pointed to the example of Ambien pharmaceutical trials to make her point. She said that scientists tested Ambien on male mice only during its development. After the drug was approved by the FDA and used publicly, it was found to take longer to dissolve in women’s bodies. This forced the FDA to issue new guidelines for use.
Rao wants to prevent these same mistakes within AI by ensuring the diversity of her teams and training. She suggests that other developers can also help combat this by checking the training date on their programs to recast some algorithms with wider data nets. She also said that teams could create a competing AI and compare the results to test its inherent bias.
Ultimately, Rao is calling for more diverse groups of people to be involved in the development, training, and testing of AI. She believes this will bring a more balanced worldview to the field and to the AI it is developing. Her hope is that more women will enter the field in order to help spot the existing biases in systems today and to avoid creating new ones in the AI of tomorrow.
Date: May 28, 2019