“India is a great place to help humanity using AI”
A lot of researchers and scholars have stepped into the field of artificial intelligence and machine learning. But, while many are focused on improving the algorithms for the machines, Duke University computer scientist and engineer Cynthia Rudin decided to tread on a less covered path — working tirelessly to use the power of AI to help society and serve humanity.
Analytics India Magazine caught up with Cynthia Rudin, Professor of Computer Science at Duke University. Her research work focuses on ML tools that help humans make better decisions, mainly interpretable machine learning and its applications. Recently, the Association for the Advancement of Artificial Intelligence (AAAI) has awarded Rudin the $1 million Squirrel AI Award for Artificial Intelligence for the Benefit of Humanity for her contributions to the area. AAAI is a significant international scientific society that supports AI researchers, practitioners, and educators. It was founded in 1979.
Cynthia: As a PhD student, I was studying the theoretical properties of algorithms. It was definitely fun, but after I graduated, I wanted to see how ML worked in the real world, so I took a job trying to help maintain the NYC power grid with AI. That project was very difficult, and it led me to understand that more “powerful” ML techniques weren’t helping with most of the real-world problems I was working on. On the other hand, if I wanted to troubleshoot, trust, and use ML models in practice, they needed to be interpretable – I needed to explain my models to the power engineers to get their help troubleshooting. So I changed my research agenda to focus on interpretable models.
At the time, there were less than a handful of AI researchers in the world focusing on this, and the ML community as a whole saw no value in interpretability. In fact, they felt it was ruining the “mystique” of ML as a type of technology that could find subtle patterns that humans couldn’t understand. So, it was extremely unpopular. However, working on interpretable models let me have more of an impact on the world (even if it was hard to publish the papers!): I started working on many different healthcare problems, applying interpretable AI to problems in neurology, psychiatry, sleep medicine, urology, and so on.
On winning the award, A ‘New Nobel’
Some of our models are used in high-stakes decisions in hospitals. I also started working on problems in policing and criminal justice. Our code for crime series detection was posted on the internet on my website and picked up by the New York Police Department (NYPD), who turned it into the powerful Patternizr algorithm, which determines whether a new crime (perhaps a break-in) is part of a series of crimes committed by the same people. I worked on criminal recidivism prediction, showing that interpretable models were just as accurate as black-box models, which makes it unclear why we subject people to black-box predictive models in criminal justice.
I also worked on computer vision, showing that even for deep neural networks, we could create interpretable neural networks that were just as accurate. When the machine learning world started trying to “explain” black-box models, I pushed back, telling them how critical it is to use interpretable models rather than “explained” black-box models for high stakes decisions and highlighting the fact that interpretable models are just as accurate if you train them carefully. My unusual work path had also opened a lot of avenues for me to serve on professional and government committees, where I was often the only AI expert, and I’ve also done a lot of volunteer work for professional societies. So when this award came along, it was essentially designed for what I had already been doing but never expected to be rewarded for in a million years.
Cynthia: Let’s think more broadly about the most serious problems affecting our society. To me, two things stand out: safety and inequity.
Cynthia: I admit that I often wish that AI could do more to solve the world’s problems. It would be amazing if AI could help deal with refugee crises, extreme poverty and reverse climate change in a major way. It could happen, I suppose. Who knows, perhaps AI will be used to uncover massive criminal bank accounts with billions of dollars in them, and this money would then be used for building homes for refugees or something. Don’t underestimate the power of AI!
Cynthia: India is a great place to help humanity using AI because there is so much need. The challenge is setting up the problem and figuring out how to get data to support it and translate it into practice. I found that it was less difficult to get models launched in practice if the domain experts could understand them, which is why I went into interpretable machine learning.
Another important tip is that it is quite easy to make mistakes in differentiating between correlation and causation, so be careful when applying ML techniques; they are typically designed for prediction only, so it is important not to interpret the model’s parameters as causal effects. There are special techniques for causal inference instead. (I happen to like almost-exact-matching techniques for this because they are interpretable.)
Cynthia: There are lots of online courses now. For instance, I post all my advanced graduate lectures on YouTube and lecture notes on my website, here. These materials are freely open to anyone. My lectures are somewhat advanced and require knowledge of probability and linear algebra, but there are many popular courses online you could take at a more introductory level.
Cynthia: It’s hard to pin down a favourite! The paper from my lab that is most well-known is this one: Stop Explaining Black-Box Machine Learning Models for High Stakes Decisions and use Interpretable Models Instead. One can find it here.
This paper explains why there is a chasm between explaining a black box and using an inherently interpretable model. It also gives many examples of interpretable machine learning models.
Cynthia: I’m actually not qualified to answer this one because I have never studied it! To be honest, I think it’s getting better now that it’s acceptable to do a much broader type of work in data science than before. My lab has always had lots of women in it. I definitely remember times in the past when I would go to conferences and feel that I didn’t quite belong, either because I was female or because my ideas – and what I thought was important – were different from the mainstream, but I think maybe those days are coming to an end. Or at least I’m hopeful about that.
Kumar Gandharv, PGD in English Journalism (IIMC, Delhi), is setting out on a journey as a tech Journalist at AIM. A keen observer of National and IR-related news. He loves to hit the gym. Contact: [email protected]
Accelerate Data Engineering
27th Oct 2021
How To Start Your Career In Data Science?
28th Oct 2021
Machine Learning Developers Summit 2022
19-20th Jan 2022
Copyright Analytics India Magazine Pvt Ltd
“India is a great place to help humanity using AI”