Estimating the functional dimensionality of neural representations.

Area of Science:

• Neuroimaging
• Cognitive Neuroscience
• Data Analysis

Background:

• Multivariate functional magnetic resonance imaging (fMRI) analysis highlights the significance of information within voxel patterns.
• Estimating the underlying dimensionality of neural representations is crucial for interpreting these patterns.
• fMRI data noise inflates dimensionality estimates, complicating the assessment of true underlying dimensionality.

Purpose of the Study:

• To develop a novel approach for identifying brain regions with reliable task-modulated signals.
• To derive an estimate of the functional dimensionality of neural signals in fMRI data.
• To address the challenge of noise-induced inflation of dimensionality estimates.

Main Methods:

• Combined singular value decomposition (SVD) with cross-validation for low-dimensional projection of voxel-response patterns at the single-subject level.
• Measured reconstruction goodness using Pearson correlation with a test set to assess significance across participants.
• Employed hierarchical Bayesian modeling to estimate underlying dimensionality and its uncertainty across participants.

Main Results:

• Validated the method on simulated data, demonstrating accurate recovery of true dimensionalities.
• Applied the method to three visual stimulus processing fMRI datasets.
• Demonstrated that the method can identify reliable, task-modulated signals potentially missed by other models.

Conclusions:

• The developed method accurately estimates functional dimensionality in fMRI data, overcoming noise limitations.
• Functional dimensionality estimation can aid in evaluating model-based analyses and identifying task-modulated signals.
• This approach reveals functional differences across brain regions and assesses task-related complexity in neural patterns.