Research

Most of my work is self-supervised pretraining on fMRI and structural MRI: how to represent the data (volumes, cortical flat maps, parcels), how to scale it, and how to evaluate what the models learn. I care most about representations that transfer across people, scanners, and tasks, including clinical ones like the early detection of neurodegeneration.

Selected publications

Google Scholar
Comparison of volume, cortical flat map, and parcellation representations of fMRI data.

ICML 2026 Spotlight

Scaling Vision Transformers for Functional MRI with Flat Maps

Connor Lane, Mihir Tripathy, et al.

We adapted Vision Transformers to fMRI using cortical flat maps and trained CortexMAE on 2,100 hours of open brain activity data. We also released Brainmarks to evaluate trait prediction and cognitive state decoding.

arXiv CortexMAE Brainmarks
Architecture of the Algonauts 2025 model for predicting brain responses to natural movies.

Algonauts 2025 4th place

Predicting Brain Responses To Natural Movies With Multimodal LLMs

Cesar Kadir Torrico Villanueva*, Jiaxin Cindy Tu*, Mihir Tripathy*, et al.

* Equal contribution

Our multimodal pipeline combines video, audio, and language features to predict brain responses to movies. It placed fourth in Algonauts 2025, with a mean Pearson correlation of 0.2085 on withheld, out-of-distribution movies.

arXiv GitHub
Functional neuroimaging model architecture: masked brain-parcel time series receive spatial and temporal embeddings, then pass through a transformer encoder and decoder to reconstruct the signal.

Biological Psychiatry 2025

Transformer-Based Foundation Model for Functional Neuroimaging

Teddy Akiki, Mihir Tripathy, et al.

A transformer-based foundation model approach for functional neuroimaging, published in Biological Psychiatry, Volume 97, Issue 9, pages S160–S161.

Biological Psychiatry
MindEye2 architecture, from shared-subject fMRI representations to image retrieval and reconstruction.

ICML 2024 Spotlight

MindEye2: Shared-Subject Models Enable fMRI-To-Image With 1 Hour of Data

Paul S. Scotti, Mihir Tripathy, et al.

We trained a shared model across seven participants, then adapted it to a new participant using as little as one hour of fMRI data for image retrieval and reconstruction.

arXiv Project Page GitHub