For Selena She ‘23, this summer has been anything but a break. The math and Swarthmore cognitive science double major has spent her summer completing back-to-back, in-person internships, researching artificial intelligence and cognitive science at the Massachusetts Institute of Technology (MIT) and Stanford University. Her work is supported by the Koshland Integrated Natural Sciences Center.

She began her project at MIT immediately after completing the academic year at Haverford. On the Boston campus, she worked with two graduate students on a project studying how effectively natural language processing (NLP) models could be used to make human-like judgments.

“I taught deep learning models to distinguish plausible and implausible sentences and evaluated how well their performances match human judgments,” she said.

She worked at MIT until June 20th, when she began at Stanford University, located in Palo Alto, CA. There, she began her Research Experience for Undergraduates (REU) internship in the Center for the Study of Language and Information (CSLI). At the CSLI, She is one of 16 students conducting research in the various labs, conducting her work with one graduate student and one principal investigator. Her project here also focused on NLP, though a different application of it.

“Here, I try to extract meaningful structures from NLP models with methods like probing classifiers and interchange intervention training (IIT),” She said. The former method involves breaking NLP models down into their component tasks, and analyzing each layer’s contributions to the final product. The latter involves switching out different tasks of the model with tasks from another model, to see how the final output changes as the model is edited. “I also write programs to see how well these models generalize to other things, like how humans categorize objects by their respective categories,” she said.

She explained that she was drawn to these research opportunities by her interest in cognitive science, a field that plays an important role in human life every day.

“We use a lot of highly sophisticated computational models in every aspect of our lives,” she said. “The problem is that many models, especially deep-learning models trained on huge datasets, are ‘information blackboxes.’ In other words, we are not sure what information they capture and how they generate outputs.” Through this line of research, she hopes to make such hidden information transparent and to deconstruct biases embedded in these programs.

She has participated in off-campus research each summer since coming to Haverford. She feels grateful for those opportunities, and as well for being able to major in cognitive science at Swarthmore. However, she also would like to see more cognitive science available on Haverford’s own campus.

"This is a very exciting field, so I would love to see Haverford introduce an artificial intelligence/cognitive science major/concentration soon,” she said. “In the meantime, I am happy to be a resource to anyone who is interested in this line of work."

"Summer Centered" is a series exploring our students' campus-supported summer work.