Limits of decoding mental states with fMRI.

Area of Science:

• Neuroscience
• Cognitive Science
• Machine Learning in Neuroscience

Background:

• Multi-voxel pattern analysis (MVPA) is widely used to decode mental states from brain activity.
• The necessity of fixed, fine-grained patterns for accurate decoding has been proposed.
• Previous studies often assume decoder precision is crucial for mental state discrimination.

Purpose of the Study:

• To investigate the spatial precision and redundancy of multi-voxel decoders for brain activity.
• To evaluate the actual contribution of decoder patterns to decoding mental states.
• To compare the performance of multi-voxel decoders with simpler brain activity maps.

Main Methods:

• Analysis of decoder performance across various tasks, including spatial smoothing and voxel subsampling.
• Distinguishing between discrimination (telling states apart) and identification (naming a specific state) performance.
• Utilizing simple similarity metrics to explain decoding performance.

Main Results:

• Decoder patterns were spatially imprecise, unaffected by spatial smoothing.
• Decoders showed high redundancy; 10% of voxels retained full performance.
• Performance was comparable to using raw brain activity maps as decoders.
• Identification performance was poor even with adequate discrimination.
• Similarity metrics explained significant variance in discrimination performance (91% within-subject, 62% across-subject).

Conclusions:

• The efficacy of current multi-voxel decoders for mental state decoding may be overstated.
• Spatially imprecise and redundant patterns suggest decoders are not as informative as assumed.
• Across-subject decoders are potentially superfluous and inappropriate for reliable decision-making in neuroscience research.