ESE PhD Candidate Nandan Tumu defended his thesis on Representation and Interaction for Trustworthy Cybernetic Systems on Monday, August 31st. Advised by Professor Rahul Mangharam, Tumu’s thesis addresses how learning-enabled systems must earn trust to achieve widespread acceptance in society through three dimensions: performance, process, and purpose.
Performance, Tumu explains, “concerns what a system does.” Process concerns “how it produces its behavior.” And purpose addresses “why that behavior is generated and whose objectives it serves.” When studying the three dimensions, Tumu’s thesis studies how “structured representations of uncertainty, physical behavior, and strategic interaction can provide stronger grounds for relying on learning-enabled systems.”
With regard to performance Tumu’s research “develops conformal prediction methods that quantify prediction uncertainty with statistical guarantees.” According to Tumu, “these methods use efficient representations to make predictions from learning-enabled robotic systems more informative and extend conformal prediction to multi-agent robotic systems and online settings in which interaction and distribution shift violate classical assumptions.”
When it comes to process, Tumu’s thesis, “develops physics-constrained learning architectures in which neural networks predict physically meaningful quantities that are propagated through known dynamic models.” As Tumu explains, “incorporating these models into learning provides stronger grounds for understanding and evaluating the processes used to predict and control learning-enable cyber-physical systems.”
As for purpose, Tumu’s work develops models of competitive influence in networked information systems. “Social influence is formulated as a game in which multiple players allocate limited resources to steer opinions evolving over a shared graph,” Tumu says. “These models make strategic objectives, network effects, and long-run outcomes explicit, exploring analysis of the purpose that drives algorithm persuasion.”
Tumu’s research journey included its fair share of obstacles. The most persistent included what to do when his research hit a dead end. To overcome said challenges, “I found that taking a step back can illuminate a new pathway to get to the problem.”
Throughout his time at Penn, Tumu credits his collaborators, lab members, and advisor for their support. In fact, his most memorable moments at Penn include collaborating with lab mates. In addition, Tumu enjoyed working in student government through Graduate Student Engineering Government (GSEG).
When Tumu isn’t working, he enjoys collecting and repairing mechanical typewriters as “a way to disconnect from the digital world.”
Learn more about Tumu’s work here