[Paper Review] Neural Topographic Factor Analysis for fMRI Data

Eli Sennesh, Zulqarnain Khan+5 more|arXiv (Cornell University)|Jun 21, 2019

Functional Brain Connectivity Studies32 references4 citations

TL;DR

This paper proposes Neural Topographic Factor Analysis (NTFA), a probabilistic generative model that infers structured embeddings for participants and stimuli in fMRI data, enabling disentangled modeling of individual and stimulus-specific neural responses. NTFA improves predictive generalization on unseen data and provides uncertainty estimates, outperforming PCA and SRM baselines in capturing neural response differences across individuals and stimuli.

ABSTRACT

Neuroimaging studies produce gigabytes of spatio-temporal data for a small number of participants and stimuli. Rarely do researchers attempt to model and examine how individual participants vary from each other -- a question that should be addressable even in small samples given the right statistical tools. We propose Neural Topographic Factor Analysis (NTFA), a probabilistic factor analysis model that infers embeddings for participants and stimuli. These embeddings allow us to reason about differences between participants and stimuli as signal rather than noise. We evaluate NTFA on data from an in-house pilot experiment, as well as two publicly available datasets. We demonstrate that inferring representations for participants and stimuli improves predictive generalization to unseen data when compared to previous topographic methods. We also demonstrate that the inferred latent factor representations are useful for downstream tasks such as multivoxel pattern analysis and functional connectivity.

Motivation & Objective

• To address the challenge of modeling individual differences in fMRI data despite small sample sizes.
• To develop a probabilistic factor analysis model that explicitly represents variation across participants and stimuli.
• To enable better predictive generalization to unseen participant-stimulus combinations compared to existing topographic methods.
• To provide uncertainty estimates for embeddings, allowing confidence in detecting meaningful neural response differences.
• To demonstrate utility in downstream tasks such as multivoxel pattern analysis and functional connectivity.

Proposed method

• NTFA extends Topographic Factor Analysis (TFA) by learning a neural network prior that maps participant and stimulus embeddings to conditional distributions over spatial factors and weights.
• It uses a multilayer perceptron to predict factor locations, sizes, and weights based on learned embeddings for participants and stimuli.
• The model employs a structured probabilistic generative framework where each fMRI scan is modeled as a linear combination of spatially and temporally coherent factors.
• Embeddings for participants and stimuli are trained end-to-end via variational inference to maximize the marginal likelihood of observed fMRI data.
• Uncertainty in embeddings is captured through the posterior variance, enabling confidence assessment of differences between participant or stimulus representations.
• The model is evaluated using both synthetic data and real fMRI datasets, including a pilot study and two public datasets (Lepping et al., 2016; Haxby et al., 2001).

Experimental results

Research questions

• RQ1 Can NTFA recover the true underlying cluster structure of participants and stimuli in synthetic fMRI data with distinguishable neural response patterns?
• RQ2 Does NTFA infer meaningful and interpretable embeddings that reflect known neural response differences across stimulus categories in a pilot fMRI study?
• RQ3 How does NTFA’s predictive performance on held-out participant-stimulus pairs compare to PCA and SRM baselines?
• RQ4 To what extent do the inferred embeddings align with previously reported findings in public fMRI datasets such as those involving emotional stimuli and object recognition?
• RQ5 Can the uncertainty estimates from NTFA’s embeddings be used to reliably detect significant differences in neural responses between stimuli or participants?

Key findings

• In synthetic data, NTFA successfully recovered the ground-truth cluster structure of participants and stimuli, demonstrating its ability to model underlying neural response patterns.
• In the in-house pilot study, NTFA inferred stimulus embeddings that revealed distinct neural response patterns across threat-relevant stimulus categories, consistent with the experimental hypothesis.
• On the Lepping et al. (2016) dataset, NTFA inferred embeddings that aligned with known differences in neural responses to emotionally valenced sounds and music between major depressive disorder (MDD) and control participants.
• On the Haxby et al. (2001) dataset, NTFA captured known functional specializations, with embeddings showing clear separation between face, object, and scrambled image categories.
• NTFA outperformed PCA and SRM baselines in predictive generalization, particularly in zero-shot settings where the model predicted responses for novel participant-stimulus combinations.
• The uncertainty estimates from NTFA enabled reliable detection of significant differences: embeddings with means separated by several standard deviations were confidently distinct, supporting robust inference of neural response differences.