About Me

Hi, I’m Alex Loftus. I am a textbook author, kaggle competition winner, and PhD student with David Bau’s group. I am interested in understanding how we can use techniques from interpretability to understand training dynamics, steer large models at inference time, reduce hallucinations, and create truly customized systems. I have worked as a data scientist, a machine learning engineer, and as a master’s student in biomedical machine learning at Johns Hopkins University.

I’ve been fortunate to work with a number of brilliant people over the years. Here are some fun projects which resulted:

• I worked with Professor Joshua Vogelstein in the Johns Hopkins Biomedical Engineering department on a pipeline to create graphs from MRI data, which led to a paper in-review at Nature Methods.
• I co-authored a textbook on spectral graph theory, which we are publishing with Cambridge University Press.
• We developed an open-source project, Graspologic, which was acquired by Microsoft and used to measure collaboration changes in their workforce during COVID.
• This year, I began giving talks for the San Diego Machine Learning meetup group, where I joined a team competing in the Vesuvius Ink Detection Kaggle Competition. We won 1st place against 1,249 teams for a competition prize pool of $100,000. Our work was featured in Scientific American!
• I made a linear algebra tutoring series for my friend, which builds up the mathematical machinery of neural networks starting from the absolute foundations: dot product geometry and linear algebra.

I have a number of academic side-interests, including spectral theory, information geometry, the history of science and mathematics, the mechanics of the visual system, constitutional law, various causal relationships between geography and history, and ethics (I am a big fan of Kant, Hume, Ross, and some modern ethicists like Susan Wolf). I am an avid traveler and am (slowly) learning Spanish.

Talks & Publications

• NNsight and NDIF: Democratizing Access to Foundation Model Internals: First authorship. Explore large model internals easily.
• A Saliency-based Clustering Framework for Identifying Aberrant Predictions: NeurIPS Workshop Paper, 2023 - won best poster
• 1st Place Solution - Vesuvius Ink Competition: Presentation, 2023, for 60 people. Presenting on our winning solution to a $100,000 Kaggle competition.
• Network Machine Learning with Python: Textbook, 2023, Cambridge University Press. Second author. 550 pages, ~200 figures.
• A low-resource reliable pipeline to democratize multi-modal connectome estimation and analysis: Paper, 2022, Nature Methods, under review. Second author, wrote most of the infrastructure for the codebase.
• ICML Conference Highlights: Talk about new machine learning techniques in drug discovery and medicine presented at ICML
• Working with LLMs: Talk, 2023, for 100 people at the AIML San Diego meetup
• Role of CAMKII in Associative Conditioning and GLR-1 Expression in C. Elegans: Poster, presented at Society for Neuroscience, 2017, Washington, DC. Later author, conducted most of the later experiments.
• Effects of an unc-43 (CaMKII) Gene Deletion on Short-Term Memory for Associative Conditioning in C. elegans: Talk, presented at Psychfest, 2017, Bellingham, WA.

Misc

I grew up in Seattle, WA. I was a competitive Starcraft 2 player in high school (grandmaster league - competed/won in seattle-area tournaments!). I studied behavioral neuroscience during my undergraduate years, with a philosophy minor focused on ethics. I got interested in math and programming and started a computational neuroscience club, where I taught weekly seminars. I also spent a lot of time partner dancing and playing guitar at open mic nights!