Ari Benjamin

In 2015, I quit a PhD in molecular simulation after witnessing neural networks translate between human languages. Somewhere inside they represented a sentences ‘meaning’, and I realized I needed to understand deep learning. I am now a computational neuroscientist specializing in neural networks and their relationship to the brain.

About half my work is theory aimed at understanding what neural networks learn, and why. I have shown that that networks naturally develop ‘efficient’ codes for the world when they learn with gradient descent (Nature Communications, 2022); and showed in spotlighted work at NeurIPS that any single network can be understood as a Bayesian ensemble, (NeurIPS, 2024). I am currently interested in the nature of beliefs in large language models.

I am also quite interested in bridging neuroscience and machine learning. I am especially drawn to what today’s neural networks don’t yet capture: neuromodulators like serotonin, the diversity of cell types in the brain, and how brains keep learning for a lifetime without overwriting what they know.

All theorists should also work with data, at least sometimes. Wanting to understand cellular diversity, I built TissueFormer, a transformer that extends single-cell foundation models to predict tissue-level phenotypes from populations of single cells (BMC Bioinformatics, 2026. In the past I have worked extensively with electrophysiological recordings to predict and decode neural activity.

I am currently a postdoctoral fellow in the laboratory of Tony Zador at Cold Spring Harbor Laboratory. I completed my PhD with Konrad Kording at the University of Pennsylvania.