Research

When Working Memory is Just Working, not Memory

Incorporated the model of episodic memory (EM) that I built for my senior thesis. The model was a neural network that retrieved “memories” composed of a stimulus accompanied by a drifting context representation that served as a kind of timestamp. The model attempted to complete a task called the N-back, in which it judged whether the current stimulus matched the stimulus from N stimuli ago. We found that, as we added noise to the drifting context representation, the model’s performance degraded in ways that matched patterns traditionally attributed to size constraints on working memory – more evidence that it’s possible the relationship between the two memory “systems” is less clearly delineated than previously thought. Many thanks to my advisor Jon Cohen.

Princeton Corpus of Political Emails

Project based on my junior paper, in which I created a modified version of the OpenWPM web crawler to sign up for email lists on campaign websites. I built a simple analysis / NLP pipeline to study questions like what topics candidates were mentioning in emails and how often they used certain rhetorical techniques (e.g. “us versus them” language). Many thanks to my advisor Arvind Narayanan and the rest of the team at CITP, who did most of the lifting to turn my prototype into a real research asset.

Authorship Attribution through Word-Pair Measures of Custom Embedding Similarity

Another junior paper, in which I built a model that created micro-sized word embeddings from the work of individual authors. I found ways to geometrically compare these word embeddings, and found that geometric similarity between two word embeddings was very predictive of shared authorship. Many thanks to my advisor Brian Kernighan.

Other Writing

From a summer journalism program in Greece: On the hunt for urban renewal? Come back Monday

From the TechCongress blog: On Sci-Fi & Tech Policy

Working on adding more here :)