Join us every Friday at Boston University Faculty of Computing and Data Sciences (CDS) for cutting-edge research presentations by CDS PhD students across data science, AI, and beyond.
We are a student-run initiative within the PhD department of Boston University Faculty of Computing & Data Sciences, dedicated to fostering knowledge sharing and academic growth across our community.
Create a space where students can explore, present, and discuss the research topics they're passionate about in a supportive, collaborative environment.
Every Friday from 12:00 to 1:00 PM in CDS 1646, CDS researchers present on work that excites them—whether it's their current research, an inspiring paper they've discovered, or a hands-on workshop in their area of expertise. From artificial intelligence to biological sciences, our seminars cover the full breadth of computer and data science.
The students who run the series week to week.
Freddy is a fifth-year PhD student in the Computing and Data Science department at Boston University, and advised by the fantastic Allison McDonald. His work explores how power dynamics are shifted by technology with a focus on applying human-driven methods to complex issues. Currently, his projects are on 2nd order dynamics in digital spaces within labor unions and the motivations used by cryptographers for their research.
Lingyi Xu is a Ph.D. student in the Faculty of Computing & Data Sciences at Boston University. She seeks solutions to data missingness in multimodal learning across visual, tabular, and text data. Her work investigates how heterogeneous, incomplete data modalities can be represented and aligned to make learning more adaptable and their relationships more interpretable.
Stella is a PhD student at Boston University Computing and Data Sciences, where she works at the intersection of cognitive science and AI.
Her research centers on modeling human decision making, combining neural networks with traditional cognitive models to uncover the psychological principles behind how we choose. She is also building large-scale, high-quality datasets to drive this work forward.
by Matt Shumway & Yuqi Tu
Pretraining representations can improve sample efficiency in reinforcement learning, but evaluating a representation normally requires expensive finetuning across many seeds and tasks. To avoid this cost, prior work scores representations with cheap proxies, such as linear probing against task-derived targets (e.g.\ reward) and direct measures of the representation itself (e.g.\ effective rank), and validates them by their rank correlation () with downstream performance. Such validation is incomplete, as scores the ordering of an entire pool of candidates, whereas the use case that matters in practice, model selection, depends only on the top choices. We therefore propose normalized fraction-of-oracle (), which scores a proxy by the downstream performance of the representation it selects relative to the best available, and argue it should be reported alongside . We also widen the pool of proxies under test: we evaluate a broad range of probing targets, spanning reward presence, multi-step return, behavioral cloning, episode structure, transition dynamics, and temporal structure. We assess seven proxies under both metrics in two model selection scenarios on the Atari benchmark: selecting among pretraining methods and among representations within a single game. We also evaluate their use for early stopping during pretraining. Across more than 20,000 finetuning runs on 15 games, we find an asymmetry: across proxies that capture different signals, proxy rankings shift with the candidate pool and downstream learner, yet their selections remain consistently strong. Representation proxies are therefore more robust as selectors than their rank correlations alone would suggest.
Ready to join our community of learners, researchers, and innovators?