You are here

Co-Designing Tomorrow's Computers

AI expert Hai "Helen" Li discusses how her lab designs new computer technologies, including personalized devices based on personalized algorithms

1. Machine learning is an area that has many, many applications, from image and speech recognition to medical pathology. What does your work focus on? Can you tell us about a couple of projects you’re working on right now?

My broad research interest is hardware and software co-design to enable small, fast and efficient computing systems. Here’s what I mean by co-design: It used to be that engineers designed the hardware for a computer, and computer scientists designed the software. Everything was quite siloed. But the new trend is to design the entire stack—so that as you design the software, you’re considering what hardware resources you have, and how much of those resources you want to use.

For example, image classification can be performed differently on a phone than on a laptop, because they have different capabilities. So when we design the software, we can take into account how exactly the task will be accomplished, and make the operation simpler and more power-efficient, or give it more power and higher performance if the system will allow it. We’re really tailoring the computing power to the particular situation. Eventually, we would like to hand every individual a personalized device that uses personalized algorithms, based on their own use patterns, to exactly fulfill their needs. The growing field of edge computing has some of the same aims: to make better overall models from large datasets, but to tailor devices to the individual user.

Machine learning: the use and development of computer systems that are able to learn and adapt by using algorithms to analyze and draw inferences from patterns in data

illustration of a brain, half composed of circuitsAnother area I work in is designing new computer memory structures. Memory is a storage system. You can imagine information stored like shoes, in rows of shoeboxes, in a warehouse. But most of the shoeboxes have not yet been filled with shoes—out of 100 shoeboxes, let’s say only four contain the shoes you’re looking for, and those shoes are on different rows, at different ends of the warehouse. You have to make four trips to retrieve the shoes you’re looking for. Wouldn’t it be much quicker and more efficient to put all of those pairs of shoes together in one spot, and ignore all the other empty boxes? That’s obviously a simplification, but it hints at the work we do in optimizing memory structure and operations.

"Ultimately, we anticipate systems that, like our brains, can adapt, recognize nuances and teach themselves as they work. It is an elegant, organic and beautiful way to think."


Another one of my major research focuses in the past ten years is neuromorphic computing. Neuromorphic computing is a new way of designing computing systems by mimicking human brain structure and operations. Ultimately, we anticipate systems that, like our brains, can adapt, recognize nuances and teach themselves as they work. It is an elegant, organic and beautiful way to think. Because brains are so different from today’s microprocessors, we must rethink and reconstruct the entire design philosophy and seek new technologies to support it.


Neuromorphic computing: an approach to computing that takes inspiration from the way the human brain recognizes and sorts information

2. How is your work influencing the field?

Our team is among the first to introduce hardware considerations to accelerate deep learning reserch, thereby bridging the gap between the hardware and machine learning communities. For example, our NeurIPS’16 paper on structured sparsity learning regularized the data structure during neural network training to optimize the memory-access pattern, so that we could speed up task execution. (Think again about the shoeboxes. Execution speedup, or retrieving the correct information, can be achieved by putting all the filled boxes together.) The structured sparsity concept has been widely recognized and adopted by industries such as Nvidia, Intel, and Facebook, as well as academia.

3. AI/ML is a wildly popular area of study. How has Duke ECE expanded its offerings so that students have more opportunities for learning in this field?

illustration of an unrolled diploma Duke ECE offers a minor in ML/AI, which provides undergraduates with not only an understanding of the core concepts of ML and AI, but also rigorous and coherent exposure to both fundamental theories and applications. In addition, ECE has recently initialized many new courses from AI/ML fundamentals:

  • ECE 685: Introduction to Deep Learning, by Vahid Tarokh
  • ECE 687: Theory and Algorithms for Machine Learning, by Cynthia Rudin  
  • ECE 660: Computer Engineering, Machine Learning and Deep Neural Nets, by myself and Yiran Chen, and
  • ECE 590: Machine Learning in Adversarial Settings by Neil Gong.

Students can also learn about use cases in various applications, with classes including:

  • ECE 590: Human Element in Cybersecurity, by Missy Cummings, or 
  • ECE 590: Reinforcement Learning at Scale by Henry Pfister and Jay Hineman.

These special topics courses change, so there are frequently new subjects to explore. We have also been working to integrate more AI/ML components into EGR 101: Engineering Fundamentals, as well as senior capstone projects.

Edge computing: Bringing computation and data storage closer to the data’s origin, to increase speed and allow for user customization 

4. Together with Professor Yiran Chen,  you run a large and very productive lab, which is now located in the Wilkinson Building. What is the lab ecosystem like? What kinds of projects do undergraduates work on, and are they mentored by graduate students?

The research in our lab covers a broad range in computer engineering, so we’ve formed several study groups led by senior PhD students. We always look for self-motivated and capable undergraduate students to work closely with graduate students, explore technical research and brainstorm new ideas. With the graduate students’ mentoring, undergraduates develop valuable hands-on skills.

5.What is your favorite thing about the Duke ECE community?

Every faculty member in Duke ECE is singular and exceptional. We don’t have the size of some other ECE departments in other schools, but our productivity and the impact of our research work is outstanding.

I like that our faculty members work closely with other departments and that ECE serves as a meeting place for the applied sciences at Duke. It really is a great place for students who are interested in research studies on electrical and computing engineering in conjunction with other important application domains, such as health data analysis, business and law. I’m very proud to be here and to be a part of it.


Learn more About AI/ML Research at Duke