Adji Bousso Dieng on AI, breaking barriers, and computing as empowerment

By Molly Sharlach, Engineering Communications

Adji Bousso Dieng wants to create artificial intelligence that can understand the mechanisms underlying complex systems and have a level of control over their behavior — work that could impact fields including science, health care and automated systems.     

Dieng was appointed to the faculty of Princeton’s School of Engineering and Applied Science in October 2020 and starts her position in September 2021, becoming the first Black woman faculty member in the engineering school’s 100-year history and the first Black faculty member in computer science. 

“This shows how far we’ve come and how far we have to go,” she said. “I’m hoping that things will change drastically and that we won’t have to celebrate these things — it should be normalized.”

A native of Senegal, Dieng holds a Diplôme d’Ingénieur from Télécom Paris and a master’s degree in applied statistics from Cornell University. She earned a Ph.D. in statistics from Columbia University in May 2020, and is currently working as a research scientist in AI at Google. She has also worked at the World Bank, and interned at Weill Cornell Medical College, Microsoft Research, Facebook AI Research and DeepMind. After completing her Ph.D. last year, Dieng founded a nonprofit called The Africa I Know, which seeks to change narratives about African history, knowledge and innovation to inspire and empower young Africans — and to provide them with opportunities to pursue careers in STEM (science, technology, engineering and mathematics).

Here, Dieng shares her research vision, her educational background, and her perspectives on breaking barriers as a Black woman in academia and STEM.

Q. What has been your research focus, and how do you hope to expand on this at Princeton?

A. I work on probabilistic modeling, an approach to AI that allows you to incorporate uncertainty and domain knowledge when learning about the structure underlying complex systems for understanding and decision-making.

Key to probabilistic modeling is the ability to come up with an interpretable generative process for data — meaning, thinking about the mechanisms by which data came to be. 

My Ph.D. thesis was about making this process more flexible, and flexibility is what deep learning is about. Deep learning allows you to extract structure from data for prediction purposes using neural networks. My thesis was about taking ideas from deep learning and bringing them into probabilistic modeling, so that you can account for interpretability, uncertainty and domain knowledge while enjoying flexibility when learning about complex systems.

So far, I have applied this to computer vision and language. At Princeton, I am looking to push on the domain knowledge side of probabilistic modeling — giving more power to what humans want out of AI systems in domains such as health care and the sciences. If we want to be able to apply our models and algorithms in those domains, we’ll need to take great care to ensure we understand and trust their behavior. 

This agenda will require collaboration with scientists, with people working on health care and biology. And we can also imagine constraints arising from societal considerations like fairness and privacy, so that will involve collaboration with people from the Center for Information Technology Policy. That’s the vision for what I want my lab to be: a multidisciplinary lab where we will leverage probabilistic modeling to think about how to incorporate desiderata and constraints stemming from domain knowledge into AI systems.

Q. Can you share a bit about your background and how you got into your field?

A. I grew up in Kaolack, Senegal, which is a region in the center, about three hours from Dakar. I was lucky to be sent to school, because my dad didn’t go to school and my mom didn’t finish high school, but she understood that it’s good for kids to have an education.

I was good at subjects like math and physics, but I didn't know what I could do with them beyond school — until I met Cheick Modibo Diarra. He was the first African to work at NASA as an astrophysicist, and he started the Pathfinder Foundation for Education and Development. The foundation organized summer camps that gathered the three girls with the best grades from [about a dozen countries in West and North Africa], so I participated in that and met him. That was the first time I had met an African who was successful in doing impactful things using mathematics and the sciences. So, I wanted to be an astrophysicist — later I figured out that I actually loved math and computer science.

After high school I received a scholarship from the Senegalese government and also won a scholarship that was funded by the foundation, and I decided to go to France. They have an education system that’s very rigorous and also familiar, because in Senegal we have a similar  system. I studied mathematics and engineering — that’s where I learned programming and computer science, including algorithms, data structures and all of that.

I love computer science because it lets you actually build and see things working that have applicability to the real world. Using computing and data to build tools that help you understand the world — I think that’s very empowering.

Q. What led you to continue your education and pursue your career in the United States? 

A. I was seeing all these success stories of people who immigrated to America, and it seemed like there were no glass ceilings. I’ve heard Joe Biden mention that he summarizes America with one word: possibilities. I totally agree with that. There are so many opportunities in this country and the possibilities are limitless. That was very attractive to me, in addition to the quality of education and research.      

Q. What kind of work did you do at Microsoft and DeepMind?

A. During my Ph.D. I had the opportunity to intern at Microsoft Research, Facebook AI Research and [Google’s] DeepMind. At Microsoft I worked on a new class of probabilistic models to help solve one of the main problems when dealing with modeling sequential data, which is that when the sequence becomes too long the dependencies are not well captured anymore. That work has been extended to conversation modeling and to health care.

At DeepMind I worked on solving one problem that was ubiquitous in the deep learning community: the mode collapse problem for generative adversarial networks (GANs). Generative adversarial networks are being used in many applications, especially for generating data, because they’re very good at that, but they run into this problem called mode collapse, where they only focus on outputting certain types of data. For example, a GAN may only output pictures of people with lighter complexions when there are a majority of white faces in the dataset, which is a problem in terms of diversity. It would fail to generate pictures of me, for example, because I have a darker complexion. So, the work I did was to alleviate that problem and make sure that the networks produce a diverse set of outputs.      

Q. Why did you choose to join the Princeton faculty, and what are you looking forward to?

A. I’m excited about starting my own lab and doing research that I care about, and the Princeton context is all the more exciting because of its history with innovation and scientists breaking new ground. 

Also significant is Princeton’s history with racial questions. There was a time when if you were Black you were not allowed to even be a student, if you were a woman you were not even allowed to enroll as a student or be on the faculty. So, I think we’ve come a long way, especially given the recent events where there’s been more reckoning about racial issues happening across the world. 

Joining Princeton is a highlight for me because those issues are real, and they happen in academia, because it took 100 years for a Black woman to be appointed as a faculty member in the engineering school. That shows there is a lot more progress left to be made. Now that one barrier is broken, people who look like me can say, “She did it, so I can do it now.” It’s meaningful to me that I get to start my academic career at Princeton because of that history. I’m hoping that things will change drastically and that we won’t have to celebrate these things — it should be normalized.

Q. Can you tell us a bit about The Africa I Know and your goals for the project?

A. The Africa I Know is a nonprofit organization with two main goals. It provides a publication platform (theafricaiknow.org) for showcasing African capability and contributions to knowledge and innovation to inspire and empower young Africans. The second goal is to provide resources and opportunities to young Africans to pursue education and endeavors in the areas of science, technology, engineering and mathematics.       

The site theafricaiknow.org also has articles on African history. Again, the idea is to showcase historically Africa’s position in the world and its contributions to many of these domains [including women’s rights and higher education] that are often attributed to people not in Africa. We want to rewrite the history from our perspective, and not the perspective of what we read, which is often negatively biased towards Africa