Machine learning algorithms have become extremely important as they increasingly become the hearts of solutions to high impact problems. Dr. Shivani Agarwal, associate professor at the University of Pennsylvania, asks fundamental questions about the design and analysis of these algorithms and finds ways to apply them to problems across several domains.
Before continuing, it’s useful to have a general idea of what machine learning is. As paraphrased from the all-mighty Wikipedia:
“Machine learning is a field that uses statistical techniques to give computer systems the ability to ‘learn’ from data, without being explicitly programmed.”
Currently, Dr. Agarwal unsurprisingly teaches CIS 520 (Machine Learning) and CIS 620 (Advanced Topics in Machine Learning) at the University of Pennsylvania. Her early love of mathematics is immediately apparent in her theoretical approach to understanding the intuition behind machine learning algorithms. She digs deep into the underlying mathematics that powers them and is definitely not one for hand-wavy proofs.
Outside the classroom, she leads her research group of a few lucky Phds. As Dr. Agarwal recounts how much she loves teaching as well as heading her research group, she says,
“It’s been fantastic. It’s one of the things I enjoy the most about being in academia: being able to mentor new minds who leave as confident scientists.”
Anyway, she and her group work in three key domains:
Exploring fundamental questions of machine learning algorithms (hold your breath for an example)
Exploring intersections of machine learning with economics, psychology, operations research, and other domains
Collaborating with life scientists to create powerful algorithms that can aid drug development and efficacy, as well as general life science research
Let’s talk about number 1. As an example, in machine learning, there’s a class of problems labeled “classification problems.” We might want to predict what color (or class) a bracelet will turn based on observations of a wearer’s mood (we assume that there are finitely many colors). In these types of multi-class problems, the machine learning community used to lack robust measurements for determining the performance of different solutions. Dr. Agarwal’s research helps provide insight into how algorithms can optimize for specific performance measures and when certain performance measures are more suitable than others.
Ok, number 2. The problems that machine learning solves are not necessarily that new. There are problems shared across economics and network theory for which interesting solutions lie in machine learning. In particular, Dr. Agarwal discussed a brief example of being able to learn orderings over noisy datasets containing pairwise relationships.
This brings us to number 3 (my personal favorite). Dr. Agarwals’ research group has made incredible contributions to the life sciences and has impacted real human lives in extremely meaningful ways. Here’s the setting for one of her projects: patients with a certain type of colon cancer would sometimes produce an enzyme called KRAS. If they produced this enzyme, they were immediately given a certain drug. Unfortunately, only 30% of these patients would even respond positively to this drug’s administration. This meant that 70% of the patients given the drug would go through “expensive, time-intensive, and cumbersome treatment with side-effects that don’t benefit them at all.” So, Dr. Agarwal worked with Mitra Biotech to develop a machine learning algorithm that could predict with 95% accuracy whether or not a patient would respond positively to the drug in question. Today, a variant of this algorithm is being used to help mitigate unnecessary suffering by saving the time, health, and money of those who would not respond well to the drug.
Dr. Agarwal’s road to involvement in machine learning started early on. “Mathematics has always felt very natural,” to Agarwal from a young age. This interest materialized in her pursuing an undergraduate degree in mathematics in India, after which she pursued computer science on scholarship at Trinity College in Cambridge. Up until then, she hadn’t the faintest idea what research even was. Then, at Trinity, she describes, with a light in her eyes, how she “met people pursuing PhDs that seemed to keep learning more and more and literally had careers doing this. And that…that just sounded like so much fun.” She then became involved with machine learning research applied to computer vision and then completed her Ph.D. later at the University of Illinois Urbana Champaign.
Dr. Agarwal closed out the conversation with a key piece of advice for students on their own paths in life.
“The earlier that students realize that it’s not the grades that matter but rather the knowledge that matters, the better prepared they will be for life at any stage. It’s very easy to get grades without getting much of knowledge. Knowledge is like this multi-dimensional object that’s really colorful with lots of substance in it. Testing, in any form, is just some low-dimensional projection of knowledge. I’m not saying ignore grades. I’m saying that focusing on knowledge is more important. If you know the full object, then you can handle any projections of it.”