Research

Research Technician · Hayden Lab, Department of Neurosurgery, Baylor College of Medicine · May 2024 to September 2026

The lab records single neurons in the brains of epilepsy patients while they listen and talk. This page is what I did there: the recordings I ran, the pipelines I built, the papers that work went into, and one analysis project of my own.

What I did

Linguistics

The language side of the lab’s analyses.

Did all of the linguistics for the lab’s Nature Human Behaviour paper (356 hippocampal neurons, 10 patients): every word of the podcasts tagged for part of speech, dependency relations, syntactic depth, clause boundaries, position within the clause, and word frequency.
Extracted the word embeddings those neurons were compared against, from five models: GPT-2, Llama-3 and DeBERTa for words in context, GloVe and Word2Vec for words on their own.
Built a 57-feature linguistic annotation of every word for my own grammar-and-meaning analysis.

Recording

Getting the data, at the hospital bedside.

Ran research recording sessions with epilepsy patients at two hospitals, Baylor St. Luke’s and Texas Children’s, working alongside neurosurgeons, epileptologists and nurses.
Recruited and enrolled patients into studies, under IRB protocols and HIPAA.
Traced noisy electrode bundles to an electrical fault rather than their location in the brain, and showed which part of the noise re-referencing could and could not remove.
Wrote the lab’s standard operating procedures for data handling and electrode reconstruction, and trained lab staff on them.

Pipelines

Turning raw recordings into data people can analyse.

Spike sorting and quality control, built solo: 98% agreement with expert curation, with manual review cut from 28% of the data to 5%. Adopted lab-wide and run daily by other staff.
Electrode localization: 203 contacts on 19 leads localized within 0.16 mm of the hand-validated result with no manual steps. The lab’s ten-step reconstruction procedure now runs end to end, with a person approving the result before anything is uploaded.
Speech transcription and word-level alignment for patient recordings, run locally so the audio never left the site.
Automatic redaction of patient identifiers from clinical log files.

Modelling

Asking what each neuron was tracking.

Poisson and logistic regression models of single-neuron firing, checked with cross-validation, permutation tests, bootstrap confidence intervals and confound-matched controls.
Datasets of 435 neurons in the main analysis and 1,008 pooled across 14 patients.
Embeddings from 26 language models, extracted on a GPU cluster as a job array that finishes in about two minutes.

Papers it went into

I am a co-author on nine papers from the lab. These four are in journals; the other 5 are preprints. The first is the one I contributed most to.

Attention is all you need (in the brain): Semantic contextualization in human hippocampus Nature Human Behaviour, 2026 · In press
Equal-contribution second author. I did the linguistics and the language-model embeddings, and helped design the analysis, run it and write the paper. It shows that hippocampal neurons track where a word sits in its clause, and that their response to a word carries a weighted mix of the words before it, weighted much as a language model’s attention would.
Plasticity and language in the anaesthetized human hippocampus Nature, 2026 · Published · 654, 714–723
A population code for semantics in human hippocampus Nature Neuroscience, 2026 · Published online 30 Sep 2026
Shared neural geometries for bilingual semantic representations in human hippocampal neurons Cell, 2026 · Published · 189(16), 5065–5080

All nine, with citations →

Multi-panel figure: a brain with the hippocampus highlighted, spike rasters under spoken words, a sentence drawn as a tree of nested clauses, and heat maps of neurons ordered by the word position they respond to. — Hippocampal neurons track where a word sits in its sentence. The sentence tree in panel D is the kind of structure I computed for every sentence in the recordings. Figure 1 of the preprint of Katlowitz, Belanger et al., bioRxiv (2025). © the authors.

Multi-panel figure: a recording probe being placed in a human hippocampus, brain scans locating it, the waveforms of single neurons, and their spikes aligned to the words of a spoken sentence. — A recording probe in a living human hippocampus, the single neurons it picked up, and their spikes under the words of a story. Figure 1 of Katlowitz et al., Nature 654, 714–723 (2026). CC BY-NC-ND 4.0.

Click a figure to open it full size.

My own analysis

Grammar and meaning in single neurons

Unpublished. Manuscript in preparation, so the numbers may change.

When someone hears a sentence, do neurons in the hippocampus track its grammar, its meaning, or both? And do language models split the two the same way the brain does?

An explainer I animated for the method: how a sentence’s grammar becomes numbers. “She played the piano” and “He broke the window” share no words but have the same structure, so they come out as the same vector. No recordings in it: the method only. 79 seconds, silent.

What held up

Grammar was the stronger signal in the hippocampus and the anterior cingulate, but not in orbitofrontal cortex. The pattern was the same for people listening to a podcast and for live conversation (14 patients, 81,370 words).
Grammar and meaning are carried by overlapping groups of neurons, along directions that are partly separate, and both load mostly on a single shared axis of population activity.
Across 26 language models, from 0.1 to 32.6 billion parameters, the grammatical part of a model predicted the neurons better than its semantic part in 23 of the 26. Bigger models did no better than small ones.

What I threw out

A large grammar-over-meaning effect at the level of whole sentences. It was sentence length, not grammar, and it vanished once length was fully controlled.
A trend for bigger language models to match the brain better. It held across 10 models and disappeared with 26.
A finding that models keep grammar in shallower layers than meaning. It reversed on the larger set of models.

Analysis I built for other studies

Music: Poisson regression with nested likelihood-ratio tests for a piano-listening study, including pitch encoding and correction for clock drift between recording systems.
Bilingual listening: Decoders for grammatical gender and verb conjugation in recordings from people listening to Spanish.

Methods and tools

Poisson and logistic GLMs Cross-validation Permutation tests Bootstrap confidence intervals Confound-matched controls PCA / CCA UMAP Python MATLAB R PyTorch scikit-learn Hugging Face SLURM GPU clusters