Papers

Dissecting EEG-Language Models: Token Granularity, Model Size, and Cross-Site Generalization

We investigate how token granularity and model size affect EEG-language model performance in both in-distribution and cross-site scenarios, and find that token granularity is a critical, task-dependent scaling dimension for clinical EEG models, sometimes more important than model size.

Lightweight Transformer for EEG Classification via Balanced Signed Graph Algorithm Unrolling

We build lightweight and interpretable transformer-like neural nets by unrolling a spectral denoising algorithm for signals on a balanced signed graph—graph with no cycles of odd number of negative edges. Experiments show that our method achieves classification performance comparable to representative deep learning schemes, while employing dramatically fewer parameters.

VoxelFormer: Parameter-Efficient Multi-Subject Visual Decoding from fMRI

VoxelFormer is a lightweight transformer architecture that enables multi-subject training for visual decoding from fMRI.

Neural and Computational Mechanisms Underlying One-shot Perceptual Learning in Humans

In this paper, we investigate the neural and computational mechanisms underlying one-shot perceptual learning in humans. By introducing a novel top-down feedback mechanism into a vision transformer and comparing its representations with fMRI data, we find high level visual cortex as the most likely neural substrate wherein neural plasticity supports one-shot perceptual learning.

Longitudinal deep neural networks for assessing metastatic brain cancer on a large open benchmark

We present NYUMets-Brain, the world’s largest, longitudinal, real-world dataset of cancer consisting of the imaging, clinical follow-up, and medical management of 1,429 patients. Using this dataset we developed Segmentation-Through-Time, a deep neural network which explicitly utilizes the longitudinal structure of the data and obtained state-of-the-art results at small (<10 mm3) metastases detection and segmentation.

Health system-scale language models are all-purpose prediction engines

We trained a large language model for medical language (NYUTron) and subsequently fine-tuned it across a wide range of clinical and operational predictive tasks, and found that it outperforms traditional models while being much easier to deploy.

Automated, Scalable and Generalizable Deep Learning for Tracking Cortical Spreading Depression Using EEG

We present a graph neural network that is able to track cortical spreading depressions in scalp EEG signals. We show that our model is scalable to different densities of EEG and generalizable to different head models.