When senior Timothy Wei presented his research at the Vancouver Convention Center during a conference in British Columbia, Canada, in December, he admitted to feeling both confident and nervous: confident of his project’s utility, but nervous about the questions that audience members might ask him.
Wei presented his paper “Dual-Model Distillation for Efficient Action Classification with Hybrid Edge-Cloud Solution” at the Neural Information Processing Systems (NeurIPS) Workshop on Video-Language Models, one of the three primary conferences for machine learning on the international stage. This paper was intended to propose a novel solution to create a less costly yet accurate model for senior fall detection, giving seniors a cheaper alternative to current expensive protection systems.
Wei first began with the USA Computing Olympiad (USACO), advancing from Bronze to Platinum — for which only the top 1-2% of competitors qualify — in the span of a little over two years. To prepare for this competition, he used Codeforces, a widely used programming application to practice solving programming questions in C++. With a rating of around 1900, he ranked in the top 200 in early 2024 out of nearly 9,000 competitors per round in the U.S., a result of hundreds of hours of dedication and hard work.
Despite his accomplishments in USACO, Wei was compelled by the societal benefits of research to change his focus.
“Competitive programming is normally very abstract because problems involve fictional characters and problems,” Wei said. “I wished to see tangible impact from what I spend my time on, not just building on my skills.”
Wei’s first project in January 024 was titled “Enhancing the Binary Classification of Wildfire Smoke Through Vision-Language Models.” Working with a Stanford University computer science undergraduate student, Pranav Kulkarni, he applied Natural Language Processing — a type of text-based machine learning — to camera imagery. The goal was to detect wildfires faster and allow firefighters to act more quickly.
Wei said this first research and presentation opportunity was a strong learning opportunity, particularly bolstering his confidence when giving academic presentations. Initially, Wei felt intimidated by the idea of giving presentations, but after this project, he realized conference attendees valued content over slick presentation skills, reducing his stress and enhancing his focus on the substance of his projects.
Wei’s next two projects were conducted last February and June at UC Santa Cruz. Those projects were called “Optimizing Large Language Models for Dynamic Constraints Through Human-in-the-Loop Discriminators” and “Dual-Model Distillation for Efficient Action Classification with Hybrid Edge-Cloud Solution.”
The first project proposed a new framework for enhancing Large-Language Models (LLMs) — AI systems that process and generate human-like text for different tasks. By testing the method on a travel planning task, Wei demonstrated that the framework delivered better results and required less data to achieve those improvements.
The second project conducted simultaneously focused on creating a system that combines smaller, faster models for efficiency with larger, more powerful ones for accuracy, using Dual-Model Distillation (DMD) to train the model. By teaching the system to decide when to use each type of model, Wei was able to reduce the computational resources needed to run the fall detection tool — one designed for elderly living alone without relatives or other caregivers.
Wei found that at the NeurIPS workshop, attendees often gave helpful advice and proposed applications for the technology that he may not have seen by himself.
“One person who noticed the poster proposed that our group could apply the DMD model to computer chip manufacturing,” Wei said. “This wasn’t what our group originally imagined, but applying this technology to new fields is very intriguing.”
Wei spent over a dozen hours per week conducting his research beginning in early 2024. Wei feels that it was time well spent. He finds the literature review — the process of identifying a project and researching similar work — to be the most challenging step.
However, Wei believes his efforts in computer science have the potential to significantly improve people’s lives. His research has been oriented toward various societal applications — such as wildfire prevention — and he plans to continue this computer engineering focus in college.
“True education isn’t just about enriching our minds; it’s about empowering others,” Wei said. “Through research, we find the tools to transform ideas into actions that improve lives and create a better future for all.”