Sanmi Koyejo is on a mission to design a fair, robust, and trustworthy machine learning algorithms

Thank you very much for agreeing to do this interview with us. Can you please introduce yourself to our readers? What do you do?

My name is Sanmi Koyejo. I am a husband and father of energetic boys. I am currently at Google, on leave from the University of Illinois at Urbana-Champaign, and moving to Stanford in September 2022.

I grew up in Lagos, Nigeria, and initially moved to the US for undergraduate studies. More recently, I spent time in Accra, Ghana, working with Google. I have lived all around the US including on the east coast, west coast, south, and midwest.

Can you tell us about your education background?

After secondary school in Lagos, Nigeria (equivalent to high school), I completed an undergraduate degree in electrical engineering at the New Jersey Institute of Technology, then moved on to the University of Texas to complete an M.S. and a Ph.D. in Electrical Engineering. I spent three years at Stanford as an engineering research associate in Psychology before starting my academic lab at the University of Illinois.

Did you have to pay for your studies? Did you have any financial burden on your education journey?

I was awarded partial scholarships for my undergraduate degree as part of an honors college program, had some part-time jobs on campus, and was fortunate that my parents covered the rest of my undergraduate education. Most of my graduate education was paid for by research assistantships, with a few gaps that I paid for using savings from internships.

How did you find your passion for Computer Science and Electrical Engineering?

Mine was a fairly typical story in that I enjoyed “fixing” electrical gadgets around the house, though my family tells me that I broke more appliances than I fixed. I was reasonably good at math, so I was in the advanced math classes at my secondary school, and I knew early on that I wanted to be an Engineer, but not much more.

I eventually chose electrical engineering as my undergraduate degree as it seemed the most versatile – however, I switched focus many times before finding machine learning. Following some undergraduate research and an internship, I was convinced I wanted to work on control systems. When I started graduate school, many of the new engineering faculties at the University of Texas worked in wireless communications, and I decided to switch areas to work with them. At this point, two life-changing events happened. First, I took my first courses in machine learning – at the time to help me build better cognitive radios. More importantly, my first Ph.D. advisor and I determined we were not a good match. This was a big deal for me at the time. After leaving their lab, I decided to focus entirely on machine learning, which turned out to be one of the best decisions for my career. It's incredible to look back and realize that I would not have picked machine learning if I had not failed.

There are people who believe in the “geek gene” hypothesis claiming people should have the 'natural talent' to excel in computing sciences. What are your thoughts on this?

I believe most people can learn to be good at most things if they can apply themselves. That said, environmental and social factors, such as early encouragement, sufficient security and support so one can focus on education, and peer social networks can significantly affect who gets the opportunity to do the work we do. Unfortunately, some of these factors are outside of an individual's control.

What is your current research about, and what kind of problems does your work solve?

My current research interests center around foundational questions in trustworthy machine learning. My conceptual work seeks to understand the role of machine learning measures of accuracy, privacy, robustness, and fairness in how we train and deploy models. In other words, I like to think about the role of measurement, the quantities we choose to focus on, and how these choices affect how we conceptualize and solve machine learning problems. Most recently, I have been fascinated with the role of humans-in-the-loop for specifying metrics and auditing the trustworthiness of machine learning.

The applied work I enjoy the most starts with a real-world problem, where machine learning can play a role in the solution. This interest has led me towards applications in healthcare and neuroimaging, where (machine learning) decisions can have a high impact. Success in these areas also requires working closely with domain experts. Further, data are often expensive to collect and high-dimensional, and as a result, most data sets only contain information from a small sample of individuals. Other common themes in these applied problems range from systems issues like learning from siloed distributed data to statistical issues such as adversarial robustness and distribution shift.

You are internationally recognized for your work in machine learning tools that enhance the fairness, robustness, and privacy of algorithms used in healthcare and neuroimaging. How has your work pushed the boundaries of research in fair, robust, and trustworthy machine learning?

I believe that human expertise remains among our best sources of knowledge for advancing trustworthy AI. Hence, my work aims to provide new tools for efficiently learning from humans.

On fairness, my lab developed new approaches for selecting performance and fairness measures from human stakeholders – to help align model performance measures with human stakeholder preferences. In robustness, we showed the vulnerability of distributed machine learning models to relatively simple failures and proposed new robustness metrics and algorithms. Motivated by privacy concerns, we are developing new tools for distributed machine learning in the healthcare and scientific settings, in collaboration with domain experts, that combine many foundational ideas to build and deploy practical systems.

Your recent work has identified racial or ethnic health disparities in COVID-19 patients based on the analysis of chest radiographs. What do you think should be done to end systemic racism in healthcare?

Issues of fairness and systemic racism are especially challenging in healthcare since it's difficult to separate which signals are driven by physiological differences and which reflect human, data, or modeling biases – this is the case with or without machine learning. These issues are exacerbated by the difficulty of evaluating ML interventions in high-stakes healthcare settings– as interventional experiments can be expensive or unethical. Addressing these issues will require the input of advocacy groups and stakeholders to help us recognize our blind spots, thoughtful clinicians that can consider how biases affect their decisions, and new algorithmic tools that can help implement measurement, mitigation, and auditing of decision-making in healthcare.

What is the most surprising discovery you found in your research?

In my applied work, I remember being surprised by the individual differences in patient care that are so closely linked to socio-economic disparities. Thinking a bit further back, some of my collaborative research on the dynamics of brain networks was very exciting when we first found them.

In terms of technical results, I have a line of research that shows how to optimize almost any classification metric using probability thresholding. These results were surprising at the time but may seem obvious in retrospect. These results helped us improve the efficiency of some of our early approaches for metric elicitation.

I know that you are the recipient of numerous awards, including Sloan Fellowship, Kavli Fellowship, IJCAI, Skip Ellis, and an NSF Early Career Award. What motivates you to excel and break barriers?

I am fortunate to have a robust support system, both personally and professionally. I have been very fortunate to work with exceptional students who bring energy and new ideas, excellent collaborators who are willing to work across academic boundaries, and incredible mentors who are my cheerleaders and advocates, often without my knowledge.

You serve as the president of Black in AI, an organization that works to increase the presence and inclusion of Black people in the field of AI. Can you please tell us more about the organization and your future goals?

Black in AI is an organization dedicated to empowering Black people in the field of Artificial Intelligence and advancing the ethical application of Artificial Intelligence technologies. Black in AI has grown significantly in its five years, with a slate of events and workshops and impactful and growing programs in academics, entrepreneurship, advocacy, and research. This success results from tireless efforts by an incredible and passionate group of board members, program leads, and volunteers. Black in AI is also very fortunate to have the support of generous funders who believe in our mission.

Black in AI is currently building the groundwork for a sustainable long-term organization. For example, we have hired full-time leadership and staff for the first time and are working on revamping and accelerating our programs. In the long term, Black in AI has very ambitious goals, and we believe our programs can help close the diversity gap of AI practitioners in academia, industry, and entrepreneurship.

How has your experience been leading Black in AI? What were your expectations? To what extent were your expectations met?

Leading black in AI had been extraordinarily rewarding – far exceeding my expectations. I continuously meet people who tell me how Black in AI changed their life or career by exposing them to new opportunities or mentoring them to successfully gain admission to the top AI graduate programs. Some of the growing pains have been in all the behind-the-scenes logistics and management, and fundraising can sometimes be taxing. However, I am energized by our mission and vision and the passionate people I work with to achieve these goals.

The Black in AI Academic Program has done an astonishing job in supporting black students, from all over the world, to pursue their graduate studies in some of the best schools in the world. What do you think is the main driver of the success of the program?

The academic team believes in the mission of changing educational opportunities for Black people in the field. Many of the students who join the program are also very passionate and put in a lot of work to be successful. Individual passion plays a significant role in the academic program's success.

You are also the General co-chair of NeurIPS 2022, one of the most important conferences in AI research. How are you planning to tackle the diversity crisis in such major AI conferences?

In addition to supporting affinity groups, ensuring accessibility and diversity are essential in all our decision-making, and providing financial support and accessibility for diverse researchers, I believe a long-term view is critical for sustained change. This means we must work toward strengthening the pipeline of diverse students entering the field. For example, this year, NeurIPS is piloting support for early education outreach.

What was the hardest hurdle you encountered in your career? How did you overcome it?

The hardest hurdle in my career was the challenge of switching fields, as I seriously considered leaving the graduate program, and I had to do a lot of soul-searching. I think I got lucky, I found a new adviser relatively quickly, and I had a great support system to keep me going. The situation could have gone differently, as it does for so many others.

What are you most proud of?

I am most proud of my children, who are growing up too fast.

How was your childhood like growing up in Lagos, Nigeria?

I think I was fortunate to have a good childhood. My ensuring memory is how important education was to my parents, and how much they encouraged my siblings and me to take it seriously.

Did you have a role model growing up?

My parents were my role models growing up.

What are your thoughts on Africa? What do you think are the biggest challenges facing the continent?

Africa has significant potential; we are just beginning to see some early fruits in AI. I am encouraged by the success of programs like data science Africa, deep learning Indaba, some of our efforts at Black in AI, and a growing number of AI-focused graduate degree programs. Career opportunities in AI are also improving with local startups, larger companies, and international companies opening offices in the African continent. I am optimistic about the future.

How can AI technologies help developing countries in order to advance their economic growth and competitiveness in the fourth industrial revolution?

Perhaps one lesson to learn from the effect of cellular phones in Africa, for example, in industries like banking, is that new technologies can have a transformative impact when they need not compete with legacy businesses. I believe AI can have such positive transformative effects, however, we should keep in mind that AI is prone to exacerbating societal biases, and ensure guardrails are in place to avoid bad outcomes.

How can excellence in STEM help Africa overcome its challenges?

Broadly, I believe that representation in who gets to build technological tools can significantly affect how the tools get built and applied, and Africans should have a seat at the table. More narrowly, excellence in AI can enable Africans to build tools that address problems others have not considered. The low barrier to entry is part of the magic of working in AI.

In your opinion, why do you think people of African descent are underrepresented in STEM fields? What do you think should be done to motivate more people to get into STEM?

Historically, the leading centers of AI knowledge have been in America and Europe. I believe this is changing, but there is still some way to go. While the educational opportunities are improving, I think there is still scope to improve the career opportunities for individuals, so they are motivated to stay in Africa.

What are important lessons you've learned on your way to where you are now in your career?

Social networks and luck have an outsized effect on success, even in academia. The people you meet will send opportunities your way, recommend you (often without your knowledge) for positions or service, discuss your work with others, and cite it. Nevertheless, “luck favors the prepared,” so there is no substitute for doing the work.

Thinking back, what do you think you would have done differently, if any?

I would have switched to machine learning much earlier in my studies.

What is something that someone wouldn't know by looking at your profile? Any fun facts?

I learned the piano at a young age and played the piano in church for many years. I don’t practice often, so I am not good anymore.

Finally, what are your pieces of advice for the African youth?

I try to avoid giving general advice, but I will try. It is an exciting time to be in the field, with many amazing opportunities to work on high-impact problems. Many organizations also have a mission to empower new participants in the field. I encourage everyone who can to take advantage of these opportunities. Do not feel compelled to work on the same problems as everyone else. If you are interested in a unique problem or application, then great! Use that as your superpower. Finally, your relationships with mentors and peers can go a long way, and you should not ignore them.