A Master of Science in Analytics team's deep-learning project uses a one-step neural network classifier to power an app that identifies yoga poses.
At the University of Chicago Master of Science in Analytics (MScA) capstone showcase, seven teams competed for a $5,000 best-in-show prize. After careful deliberation, the jury of five judges conferred the prize on Shahbaz Chaudhary, Emily Coppess, and Jay Ong for their project entitled “Skeletal Insights: Yoga Pose Classification Using Deep Learning,” which was lauded for its end-to-end solution to a very complex problem.
Receiving honorable mention were Steven Blemker, Chirag Dadhaniya, and William Foran for “Great Wolf Resorts Onsite Revenue: Discovering What Drives Spend in Lodges” and Gordon Dri, Leibao Qi, and Monica Vuppalapati for “An Analytical Framework to Study the Diagnostic Pathways of Patients within the NorthShore Health System.”
“It was a great pleasure to see many very successful presentations,” said Sema Barlas, former director of the MScA program and part of the jury for the day’s event. “Many of the presentations were better than what I typically see at professional conferences. It was very difficult for the jury to choose a recipient team for the best-in-show award among many excellent presentations but in the end the team behind Skeletal Insights emerged on top.”
In addition to Barlas, the jury for identifying the best-in-show prize recipient consisted of Medy Agami (MScA Instructor), Ashish Pujari (Alumnus and MScA Instructor), and Daniel Shapiro from Mathematica. The $5,000 best-in-show prize can be used in the upcoming year to attend an analytics-related conference or to take additional courses as an alumni scholar.
Student Project Measures Human Skeletal Movement
What began with a more general idea than yoga, “Skeletal Insights” sought to measure something that is not typically measured. In this case, the project measured human skeletal movement, with the intention of developing a fitness-related application further down the road. Using a camera to estimate skeletal movement from change in body surface shapes, the team was able to extract data and then judge whether it approximated a fitness- or sport-related form.
“You could compare, for instance, how your golf swing relates to Tiger Woods’s,” noted Chaudhary. “Or how the form of your basketball shot compares to Michael Jordan’s.”
“We settled on yoga,” added Coppess, “because it’s a fitness area with concrete labels that reflect a particular skeletal alignment.”
"One of the important learning experiences of the capstone, beyond the analytics techniques we learned, had to do with solving problems as a group."
How Deep Learning is Utilized to Identify Yoga Poses
“Skeletal Insights” compared two applications of deep learning that used image data to classify and determine the yoga pose being represented: a one-step neural network classifier and a two-step model consisting of a pose-extracting neural network feeding into standard classification models like SVM and random forest. The one-step classifier significantly outperformed the two-step classifier.
The team notes that it was Ashish Pujari, their capstone supervisor, who urged them to try the method using the one-step neural network, which is an approach using techniques and algorithms sufficiently new and cutting-edge.
Solving Problems as a Group
“Without Ashish,” Coppess says, “we wouldn’t have finished our project. He was incredibly helpful when it came to understanding these algorithms and giving us recommendations for what to try next. One of the important learning experiences of the capstone, beyond the analytics techniques we learned, had to do with solving problems as a group. You learn to work with people who might think differently from you and in doing so you create something better.”
Chaudhury also noted the value the capstone experience brought beyond the purely analytical by pointing to the leadership skills working on such a project fosters. Ong highlighted how the coursework, though not directly used in their capstone, provided them with the foundation to confidently pursue new and complex areas of analytics.
“Our rigorous MScA training gave us a good intuition on which problems are solvable and worth pursuing,” Ong says. “Having a good sense for general concepts like the bias/variance trade-off was particularly vital as we iterated through different deep learning techniques with limited computational resources.”