Noise correlations in the human brain and their impact on pattern classification.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Machine Learning

Background:

• Multivariate decoding methods like multivoxel pattern analysis (MVPA) are vital for brain imaging data analysis.
• The exact information source for MVPA's success, particularly the role of individual voxel selectivity versus inter-voxel correlations, is debated.
• Existing theories often focus on aggregated voxel selectivity, overlooking the potential contribution of neural noise correlations.

Purpose of the Study:

• To investigate the role of noise correlations between voxels in functional magnetic resonance imaging (fMRI) multivariate decoding.
• To extend computational theories of noise correlations from neuronal populations to fMRI data.
• To determine if noise correlations contribute to the success of multivoxel pattern analysis (MVPA) in heterogeneous neural populations.

Main Methods:

• Applied multivariate decoding techniques, including MVPA, to fMRI data.
• Measured noise correlations between voxels during rest and task periods.
• Utilized computational simulations to model the influence of selectivity and noise correlations on decoding.
• Examined the relationship between voxel selectivity, classification weights, and noise correlations.

Main Results:

• Decoding performance in MVPA is enhanced when voxels with higher noise correlations are selected for classifier training.
• Voxels with strong selectivity or high MVPA weights tend to exhibit high noise correlations with voxels representing alternative classes.
• Simulations confirmed that noise correlations generally enhance decoding in fMRI data and can be distinguished from selectivity effects.
• Above-chance classification accuracy in fMRI data is modulated by the magnitude of noise correlations.

Conclusions:

• Noise correlations between heterogeneous voxels significantly contribute to the success of multivariate decoding in fMRI.
• The findings challenge theories emphasizing only voxel selectivity, highlighting the importance of neural noise structure.
• Understanding noise correlations is critical for optimizing decoding algorithms and interpreting brain imaging data.