Feature-reweighted representational similarity analysis: A method for improving the fit between computational models,...

Area of Science:

• Cognitive Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• Representational Similarity Analysis (RSA) compares representational spaces using (dis-)similarity matrices but assumes equal feature importance.
• Classical RSA may underestimate model-brain correspondence and lead to suboptimal model selection.
• Multivariate decoding offers flexibility by reweighting features, a capability not present in traditional RSA.

Purpose of the Study:

• To evaluate feature-reweighted RSA (FR-RSA) for computational models, assessing improvements in representational similarity matrix (RSM) correspondence and model selection.
• To introduce voxel-reweighted RSA, applying FR-RSA principles to fMRI voxels for enhanced analysis of brain activity patterns.
• To generalize findings across diverse datasets and data modalities using popular deep neural networks.

Main Methods:

• Applied FR-RSA to computational models (deep neural networks) and brain/behavioral data across multiple datasets.
• Developed and tested voxel-reweighted RSA by optimally weighting fMRI voxels.
• Compared FR-RSA results with classical RSA and proposed updated noise ceiling computations.

Main Results:

• Reweighting model units significantly improved the fit between model and target RSMs from fMRI and behavioral data.
• FR-RSA affected model selection outcomes, highlighting its impact on comparative analyses.
• Voxel-reweighted RSA demonstrated even more pronounced improvements in RSM correspondence.
• Classical noise ceilings were found to be exceeded by FR-RSA, necessitating updated computation methods.

Conclusions:

• FR-RSA broadly validates its utility for enhancing the fit between computational models, brain, and behavioral data.
• The method provides a better framework for adjudicating between competing computational models.
• FR-RSA applied to brain measurement channels represents a promising new approach for assessing representational space correspondence.