Lisa Yan

PhD Candidate
Department of Electrical Engineering

Stanford University

Lisa Yan is a fifth-year PhD candidate in Electrical Engineering at Stanford University advised by Professor Nick McKeown and Professor Chris Piech. Her research is in visualizing students in large undergraduate computer science classrooms. Throughout her graduate career, Lisa has been involved in teaching–from computer networks to introductory probability for machine learning. As part of her commitment to teaching, she has researched how to quantify student progress, reported on the benefits of reproducing research in the classroom, and participated in CS Bridge, a non-profit dedicated to introducing high school students around the world to computer science. Lisa is a Stanford Graduate Fellow and a National Science Foundation (NSF) Graduate Research Fellow, and she holds an MS in Electrical Engineering from Stanford University, USA, and a BS in Electrical Engineering and Computer Science from the University of California, Berkeley, USA.

Learning from Massive Collections of In-Depth Student Work

As undergraduate computer science classes grow, instructor workload also increases; at scale, it is hard to know which students need extra help, much less design classroom exercises that inspire, challenge, and accurately assess students. Many introductory computer science courses also use graphics-based assignments, which allow students space to experiment and create, at the expense of being much more complex and difficult to grade.

 

My thesis work is researching how, in these large computer science classrooms, we can process large amounts of complex assignment data in order to better monitor and understand student learning. Such a task involves creating the correct datasets, designing reliable classification models, and devising methods to efficiently visualize and characterize the student learning process to instructors. I use deep neural networks grounded in computer vision research that use the graphics output of programs to understand the milestones of intermediate student work, and I show that a neural network autograder can be designed with relatively high confidence levels. My work goes one step further and moves from single-image graphics assignments to animated, input-based assignments, such as simple Atari-inspired games. I present an autograding agent inspired by Deep Q-Learning that learns to play student-created Atari Breakout games for the purpose of grading. The goal of my research is to use data science as a preprocessor for monitoring student progress so that human teachers can focus their teaching on the students who need it the most.