Tech Skills + Appetite for Travel Lead McArthur ’19 to TripAdvisor
In computer science, Generative Adversarial Networks (GANs) learn to imitate and display such visual input as handwriting and human faces. But what happens if GANs are linked and given the opportunity to work together, like trees in a forest? Well, that’s what Ian Nduhiu ’22 and Kenny Talarico ’22 have spent the past couple months trying to figure out.
This summer, under the guidance of Visiting Assistant Professor of Computer Science Dave Perkins, Nduhiu and Talarico are conducting research and running tests on GANs. They are specifically developing a program that, using multiple GANs, generates images of a
Hometown: Utica, NY
High School: Notre Dame High School
Prospective major: Computer science
“language” of symbols they handwrote, hoping that they can make a computer recognize and recreate each symbol.
Nduhiu explained that GANs are made up of computer “generator” and “discriminator” networks that have opposing goals and “learn” from each other to better achieve those goals. He compared a GAN’s generator network – which attempts to mimic an image – and a GAN’s opposing discriminator network – which tries to separate real and generated images – to counterfeiters and policemen competing over money.
“The policemen are getting really good at catching the counterfeit money, and the counterfeiters are getting really good at faking the real money, so the policemen can’t really tell whether the money is fake or not. The whole process is like tug-of-war, and at the end, counterfeiters get really good and you can’t tell the difference between the fake money and the real money,” said Nduhiu.
Prospective major(s): Computer Science, Math
Hometown: Nakuru, Kenya
High School: Moi High School
In the research group’s study, the generator network becomes better at producing images that look similar to the research group’s symbols, like the counterfeiters becoming better at creating fake money.
Through their research, the group aims to expand current knowledge of GANs and their capabilities. Perkins mentioned that their original idea had been to use artificial intelligence to “develop a language from scratch,” but changed the plan along the way and instead designed the language recognition GAN project. Nduhiu said that the use of language in their research is what ultimately helps make their project “unconventional” and a helpful contribution to the field.
For both Nduhiu and Talarico, learning how to do such a project initially proved challenging. “For the first week or two it was definitely difficult because it was stuff we were learning how to do from scratch,” Talarico said, explaining how Perkins had also been unfamiliar with coding GANs. “We were sort of teaching ourselves and he was learning as well.”
Nduhiu and Talarico enjoy their research and believe that they might do similar work in the future. For Nduhiu, working with GANs has piqued his interest in artificial intelligence, encouraging him to learn more about that branch of computer science. Talarico plans to graduate with a minor in linguistics, and this project thus serves as “the intersection of all [his] academic interests.”
Altogether, the group has enjoyed learning about an emerging field of computer science and working through their research together. While they plan to complete the project by the end of the summer, Perkins said that they ultimately have had several ideas about how they might complicate the generator and discriminator network dynamics going forward.