hiya.
i’m henry, a post-doctoral research fellow in the AI Lab at Princeton.
in the AI Lab, i work with Jon Cohen, Tom Griffiths, and Sarah-Jane Leslie.
before that i did my phd in informatics (computer science) at the University of Edinburgh, where i was advised by Kenny Smith and Ivan Titov
i’m a computational linguist working on learning, and structure. with an emphasis on models of language. if that’s of interest, get in touch
how do deep learning models learn to do so much, so well?
My research tries to understand what learning looks like on a representational level. I focus on Information Theory, building efficient, scalable, approaches to interpretability, that let us better understand how large-scale neural networks, work. This helps to provide insight into how learning may work in humans and other species – and helps us build better models by understanding the representational effects of different design decisions.
selected publications
Representations as Language: An Information Theoretic Framework for Interpretability
Representations as LanguageAn Information Theoretic Framework for InterpretabilityThis paper appeared as a talk at the International Meeting of the Cognitive Science Society in 2024.AbstractLarge scale neural models show impressive performance across a wide array of...
Anaphoric Structures Emerge Between Neural Networks
Anaphoric Structures Emerge Between Neural NetworksWithout explicit efficiency pressures.This paper appeared at the International Meeting of the Cognitive Science Society in 2023.AbstractAnaphors are ubiquitous in human language; structures like pronouns and ellipsis...
Compositionality With Variation Reliably Emerges in Neural Networks
Compositionality With Variation Reliably Emerges in Neural NetworksEmergent representations in a multi-agent model are rife with the kinds of variation ubiquitous across natural languages. This paper appeared at ICLR 2024.AbstractWe re-evaluated how to look for...
Meta Learning to Compositionally Generalise
Meta Learning to Compositionally GeneraliseIntroducing domain-general biases via optimisationThis paper appeared as a talk at the Meeting of the Association of Compositional Linguistics (ACL) in 2024.AbstractNatural language is compositional; the meaning of a sentence...