Feb 2024 – May 2024 · Georgia Institute of Technology

Neural Language Encoding Models

Research Project (Coursework)

Implementation and analysis of semantic encoding models based on Huth et al. (2016) to predict brain responses to natural language stimuli. Built regression models using word embeddings and Finite Impulse Response (FIR) to map semantic features to BOLD fMRI responses across cortical voxels.

Key Contributions

Implemented encoding models using English1000 word embeddings (985 dimensions, 10,470 features) to represent semantic content of naturalistic story stimuli
Built FIR models to predict hemodynamic responses across 2-8 second windows, achieving significant prediction accuracy for semantic-selective voxels
Analyzed 20+ best-predicted voxels to identify semantic selectivity patterns, revealing distinct cortical representations for court-related, color, and attribute categories
Compared semantic vs. phonemic encoding models via correlation analysis, demonstrating superior performance of semantic features in predicting language-responsive brain regions
Processed and downsampled stimulus representations to match fMRI temporal resolution, handling high-dimensional feature spaces across multiple story datasets

ResearchfMRINLP