Natalie Abreu

Hi, I'm a first-year PhD student at Harvard University advised by Boaz Barak. Previously, I attended the University of Southern California where I completed a BS in Computer Science and minor in Mathematics. During that time, I interned at Google and at MIT Lincoln Laboratory.
I'm very fortunate to be supported by a Kempner Institute Graduate Fellowship.

Research Interests

I am broadly interested in the foundations of deep learning, particularly with respect to LLMs. This includes questions regarding the training dynamics of large models, how LLMs perform reasoning tasks, and what information these models make use of. I plan to address these questions from both theoretical and empirical approaches in order to build our understanding of deep learning and our mental models of these systems.

Publications

Addressing Discrepancies in Semantic and Visual Alignment in Neural Networks

with Nathan Vaska and Victoria Helus

ICML 2023 Workshop on Data-centric Machine Learning Research|Link

Addressing Mistake Severity in Neural Networks with Semantic Knowledge

with Nathan Vaska and Victoria Helus

NeurIPS 2022 Workshop on Progress and Challenges in Building Trustworthy Embodied AI|Link

Teaching

Theory of Computation
CSCI 475 @ USC, Spring 2023 Course Producer

Introduction to Algorithms
CSCI 270 @ USC, Spring 2022 - Spring 2023 Course Producer

Fundamentals of Computation
CSCI 102 @ USC, Spring 2021 Course Producer