Engineering Students Show Their Work at R&D Showcase

Authored by:

If you've ever stumbled in a quick-braking bus or made the sweaty trek up a stopped escalator from the depths of a Metro station, you may have thought you were suffering alone.

But all such incidents, from the most minor equipment malfunction upward, are reported to WMATA, Washington, D.C.'s public transit agency. The result is a massive trove of data that could hold the key to preventing similar incidents in the future. Because the data comes from so many sources, however, it arrives in many different formats, which has historically made truly comprehensive analysis difficult.

Enter the data heroes: senior engineering management and systems engineering students from the George Washington University, who worked with WMATA's Safety Informatics and Solutions Team for their senior capstone project this year.

And what they were asked to tackle was no small feat.

"We received this huge Excel file with 32,000 observations," said senior Amna Maqsood, who discussed the project at the School of Engineering and Applied Science's Student Research & Development Showcase last Friday.

Maqsood's was among 119 teams of undergraduates, graduate students, postdoctoral scholars and research scientists displaying posters and interactive stations across two crowded floors of the Science and Engineering Hall. The annual event was founded in 2007 by alumnus Randolph "Randy" Graves, D.Sc. '78.

"This is a fantastic event that brings so much energy to this building," GW Engineering Interim Dean Jason Zara said. He urged attendees to "Enjoy your day and learn, learn, learn!"

"GW Engineering research is growing strongly year by year," said Associate Dean for Research Lijie Grace Zhang, who also praised the GW Engineering community's "remarkable engagement and determination."

Faced with a 32,000-entry Excel spreadsheet, Maqsood and her colleagues had to have both. What's the first step in analyzing something so massive?

"You cry," Maqsood joked. "We were so overwhelmed."

Eventually, the team was able to standardize the sensitive, complex WMATA data set. They used it to create both a comprehensive map of public transit safety incidents across the Washington, D.C., area and a forecasting tool that could help predict and prevent the conditions under which such incidents occur.

Some of the team's findings were astonishing-and extremely useful. For instance, they found that just three locations accounted for 40% of all reported incidents. "We were able to figure out those hotspot locations where incidents are occurring most frequently to try to reduce that to as low as possible," team member Alana Lee said.

The team also proposed a streamlined data architecture that would increase WMATA's responsiveness and efficiency and eventually be able to integrate artificial intelligence (AI) for even more advanced predictive modeling, team member Gladys Fong explained.

"Ultimately, we're hoping to help WMATA inform their safety operations and training," Lee said.

The WMATA collaboration was one of many projects to incorporate AI, either as a central mechanic or a subject in itself. Several capstone teams developed apps with a machine learning component: a sports meetup app helps match users to activity partners at their own skill level, while a nutrition and cooking app personalizes its output based on how a user rates the recipes it suggests.

Master's students Aalekh Bukhariya and Rachit Das explored an ongoing problem in generative AI: its tendency to produce incoherent or factually incorrect hallucinations. (In a recent incident, an elite law firm issued an apology after its lawyers submitted a filing riddled with AI-generated errors including citations of nonexistent cases.)

Bukhariya and Das performed geometric analysis of hallucinatory and correct responses by different AI models, plotting the "hidden states" through which information passed for each. Accurate responses produced scattered plots, suggesting that correct information could be produced via a number of different trajectories. But the graphs representing hallucinations-even across multiple models-tended to look similar, with data tightly clustered in particular areas of the graph.

This consistent geometric signature suggests that AI hallucinations are not as random as they might seem, Das and Bukhariya said.

Further analysis suggested they may be produced by problematic "neurons" within the model, potentially because these assign too much weight to the output format and not enough to the input question. Eliminate these neurons, the team found, and the hallucinations stop.

"When you're looking to mitigate hallucinations, your first instinct might be to go to some external system," Das said. "But what we are pointing towards is that there may be an issue with the architecture itself."

For many presenters, the R&D showcase was an opportunity to take a breath and reflect on their achievements. At a table manned by seniors Abdu Altahir, Connor Cheung, Julian Gross and Joshua Yao, attendees lined up to try something most people would dread in real life: merging onto a busy Los Angeles highway.

Their project worked to improve autonomous vehicle research capabilities at GW's Intelligent Aerospace Systems Laboratory (IASL) by developing a co-simulation linking two open source platforms: the first-person autonomous car simulator Car Learning to Act (CARLA) and the aerial view traffic simulator Simulation of Urban Mobility (SUMO). The team enabled exact synchronization between the two, allowing virtual "drivers" to experience SUMO's realistic large-scale traffic conditions in the high-fidelity, responsive CARLA environment.

"It feels like the end of a very long road, so I'm happy that we get to present it to everybody and show our final product," Gross said. "It's very, very gratifying."