As part of the Geometric Media Lab at ASU, I have been able to contribute to cutting edge research in the area of computer vision and its applications across healthcare, security systems and national initiatives. My early projects at GML explored time-aware knowledge distillation for gait analysis, where I examined how temporal representations influenced predictive accuracy on human motion data. I was fascinated by how the governing of topology and organization of the data made significant contributions to the accuracy rather than network depth or parameters. This led to a research poster presented at the Fulton Forge Student Research Expo titled : Advancement of Wearable Robotics through Time-series and Video Data Analysis.
Motivated by this, I contributed to my first co-authored paper, the Topological Guidance-based Knowledge Distillation (TGD) framework, which was presented at the ICML 2024 Geometry-Grounded Representation Learning Workshop. In this work, I employed topological data analysis to extract persistence diagrams and images from raw images, encoding geometric structures such as connected loops and components. These rich representations were then used to train a topological teacher network, whose knowledge was distilled in a compact student model. This work was instrumental in exposing me to the art of distillation and efficient models, critical to developing models suitable for real-world deployment.
With the guidance of Dr. Pavan Turaga, I began a funded position at the lab contributing to research initiatives of DARPA and NSF. I led the development of a representation pipeline for space communication data generated by the Interdisciplinary Science and Technology Department at ASU. This dataset models multi-team systems involving humans, robots, and autonomous agents exchanging information under simulated mission conditions. I designed an algorithm that applies temporal filtration topology combined with statistical measures of volume and entropy to transform 1D temporal telemetry into interpretable 2D representations. Further, I employed a ConvKAN (Kolmogorov–Arnold Network) as a nonlinear teacher model to guide lightweight student networks through knowledge distillation by leveraging the representation algorithm. This strategy was able to achieve over 25% improvement in F1 score compared to conventional baselines in space communication networks. The manuscript is currently being prepared for submission to a signal processing journal.
A number of people across the lab have acted as mentors to me for the past three years namely Eunsom Jeon (Assistant Professor at SeoulTech), Rahul Khurana( SWE at Amazon), Keun Park (PhD student at ASU) and more. Through this research experience, I have been able to learn about my research direction and prepare well for a graduate degree.
Summary and Reflection
Overall, the experience that GML has brought to my academic journey has helped define what joy of living means to me. Working across these projects and working with mentors, I was able to define my research and the kind of researcher I want to be. Throughout these projects, I have worked toward methods that can benefit areas such as image classification in defense, robotics applications, and human activity recognition for accessibility purposes, and that is what I see my work at GML ultimately being used for. The ability to contribute to aerospace operators, or to someone who relies on wearable robotics, through methods I helped develop is what joy of living means to me. It is the knowledge that my work can advance AI across fields that matter.