Research Experience for Undergraduates
Stanford’s School of Engineering is home to amazing researchers and research.
And undergraduates can be a part of it through the Research Experience for Undergraduates (REU) program, which provides opportunities to work on cutting-edge research guided by Stanford faculty. Professors are excited to work with bright, enthusiastic students.
Information about the program will be sent in early January to declared MS&E majors and minors, and students who have begun the declaration process in Axess.
The Department of Management Science and Engineering seeks applications from current Stanford undergraduate students in early February each year and matches applicants with participating faculty members. Students participate in 10 weeks of research with a faculty member, culminating in a presentation to all participants and participating faculty. Weekly group meetings are held with all participants.
Any other interested students may contact Patty De Luca in January. For general advice on how to get involved in research at Stanford, refer to the Stanford Undergraduate Advising and Research (UAR) Guide.
See the kind of work REU participants are doing below:
Summer 2024 REU projects
During the last week of summer quarter, REU participants presented their work on campus. Learn about two of those research projects in more detail below:
The SAT and inequalities in college admissions
Student (major)
Viviana Iglesias (MS&E)
Faculty mentor
Project description
My project investigated the inequalities within the college admissions process, particularly in understanding whether or not the SAT is a major hurdle for students from lower socioeconomic backgrounds. I explored how socioeconomic status (SES) and race correlate with SAT performance and what implications this has for fair access to higher education. By analyzing data and trends post-SFFA v. Harvard College, I aimed to identify whether or not the SAT threatened the existence of a diverse student body.
Tools and methodologies used
For this project, I conducted an extensive literature review, analyzing numerous journal articles and studies that utilized data from college applicant pools. This research involved examining the correlations between SES, race, and SAT performance through statistical data and graphs. By synthesizing findings from these sources, I identified key patterns and relationships that informed my analysis of the broader implications of standardized testing in college admissions.
Findings
The findings suggest that while there is a high correlation between SES and SAT scores, the test remains an effective predictor of first-year college performance by itself and even more so when used in conjunction with high school GPA. However, eliminating the SAT may not fully address disparities, as SES factors continue to influence other academic measures such as high school GPA and extracurricular involvement. Ultimately, while SES plays a significant role in admissions outcomes, attributing most of the SES biases within admissions on SAT scores overlooks broader systemic issues.
Real-world impact
This research highlights the complexities of creating equitable college admissions systems. Policymakers and educational institutions may use these findings to reconsider their reliance on SAT scores and explore more comprehensive approaches to admissions. By understanding how SES influences both test performance and academic success, universities can work towards admissions processes that better account for students' backgrounds, promoting diversity and fairness in higher education.
LLM-powered causal graph discovery: Evaluating on genetic perturbation prediction
Student (major)
Pinlin [Calvin] Xu (Computer Science)
Faculty mentor
Project description
Causal inference aims to uncover intrinsic relations among variables from observed outcomes, but what if we can augment it with a priori knowledge about the variables? In this project we try to help a gene perturbation prediction pipeline with a graph of regulatory interactions distilled from biomedical papers in PubMed using GPT-4. By enhancing the generative model for plausible gene expression profiles of T-cells following CRISPR-Cas9 gene knockouts, we allow better prediction of the cancer-fighting ability of such modified T-Cells.
Tools and methodologies used
The precision and recall of causal graph discovery is benchmarked on several open and closed LLMs; subsequently a knowledge graph is built from part of PubMed and stored in a Neo4j graph database, from which a causal graph used to find and traverse gene regulatory networks is derived using Cypher queries and additional few-shot prompting. The causal inference pipeline is deployed on SOAL and mainly uses causal-learn and scikit-learn.
Findings & real-world impact
Using the LLM-derived causal graph resulted in limited improvement over previously the unperturbed and the Greedy Sparsest Permutations (GSP) algorithm baselines. As the theoretical lower error bound is still an order of magnitude smaller, we hypothesize that we are currently limited by the sparsity of our graph, which may be verified with a denser graph on a smaller set of selected feature genes. Nevertheless, we find the viability of LLM-derived graphs in knowledge retrieval from vast amounts of existing academic literature, and in knowledge distillation from the largest pre-trained models into existing task-specific pipelines. We also plan to test its versatility and create perturbation embeddings for an alternative deep generative approach.