Abstract
Multivariate analysis of fMRI data has benefited substantially from advances in machine learning. Most recently, a range of probabilistic latent variable models applied to fMRI data have been successful in a variety of tasks, including identifying similarity patterns in neural data (Representational Similarity Analysis and its empirical Bayes variant, RSA and BRSA; Intersubject Functional Connectivity, ISFC), combining multi-subject datasets (Shared Response Mapping; SRM), and mapping between brain and behavior (Joint Modeling). Although these methods share some underpinnings, they have been developed as distinct methods, with distinct algorithms and software tools. We show how the matrix-variate normal (MN) formalism can unify some of these methods into a single framework. In doing so, we gain the ability to reuse noise modeling assumptions, algorithms, and code across models. Our primary theoretical contribution shows how some of these methods can be written as instantiations of the same model, allowing us to generalize them to flexibly modeling structured noise covariances. Our formalism permits novel model variants and improved estimation strategies: in contrast to SRM, the number of parameters for MN-SRM does not scale with the number of voxels or subjects; in contrast to BRSA, the number of parameters for MN-RSA scales additively rather than multiplicatively in the number of voxels. We empirically demonstrate advantages of two new methods derived in the formalism: for MN-RSA, we show up to 10x improvement in runtime, up to 6x improvement in RMSE, and more conservative behavior under the null. For MN-SRM, our method grants a modest improvement to out-of-sample reconstruction while relaxing an orthonormality constraint of SRM. We also provide a software prototyping tool for MN models that can flexibly reuse noise covariance assumptions and algorithms across models.
Introduction
Functional MRI (fMRI) data is inherently structured as a matrix, where one dimension represents space (voxels) and the other represents time (TRs). Despite this inherent structure, traditional analysis methods often treat these dimensions asymmetrically. Some models account for correlations between voxels while assuming temporal independence, while others do the inverse. This methodological fragmentation has led to "siloed" approaches where popular techniques like Representational Similarity Analysis (RSA) and Shared Response Mapping (SRM) are treated as entirely different animals, each requiring bespoke software and optimization algorithms. Furthermore, ignoring the "Kronecker-separable" nature of fMRI noise—where noise at one voxel and time point is correlated with others—can lead to biased estimates and inflated false-positive rates.
The Matrix-Normal Formalism
To address these issues, we propose using the Matrix-Normal (MN) formalism to unify these disparate methods. An MN distribution describes a random matrix where rows and columns have their own distinct covariance structures. This approach allows us to treat regression-based methods (like RSA) and factor-analysis-based methods (like SRM) as two sides of the same coin. In the MN framework, they share the same fundamental structure; the primary difference is whether the design matrix is observed or latent.
Key methodological contributions include:
- MN-RSA: A reformulated version of Bayesian RSA that models spatial and temporal noise simultaneously. Crucially, its parameter count scales additively with the number of voxels, whereas previous state-of-the-art methods scaled multiplicatively, enabling much more efficient whole-brain analysis.
- MN-SRM: A version of Shared Response Mapping that replaces the traditional, rigid orthonormality constraint with a more flexible Gaussian prior, improving the model's ability to reconstruct actual brain activity.
- ECM Algorithm: We derived an Expectation-Conditional-Maximization algorithm to fit these models efficiently, ensuring the number of parameters remains manageable as more subjects are added to a study.
Empirical Evaluation and Results
The framework was validated using synthetic data to establish ground truth, alongside real-world datasets including the Human Connectome Project (HCP), the Raider dataset (subjects watching Raiders of the Lost Ark), and the Sherlock dataset (subjects watching Sherlock).
Our results demonstrate significant advantages for the MN-based methods:
- Speed and Accuracy: MN-RSA achieved up to a 10x improvement in runtime and was up to 6x more accurate (lower RMSE) than the previous state-of-the-art (BRSA) on large-scale synthetic problems.
- Fewer False Alarms: Under null conditions, MN-RSA proved to be the most conservative method, correctly identifying a lack of structure where other methods produced spurious patterns.
- Improved Reconstruction: MN-SRM consistently outperformed traditional SRM in out-of-sample reconstruction, proving that the flexible Gaussian prior is a more realistic assumption than strict orthonormality for brain mapping.
Conclusion and Future Work
The MN framework provides a powerful, unified toolkit that improves both the efficiency and accuracy of fMRI analysis. While current iterations assume uniform temporal noise structure across voxels, future work will aim to incorporate voxel-specific noise structures and explore more complex, non-linear relationships. By providing a software prototyping tool that allows researchers to "plug and play" different noise models, we hope to foster further integration across the fMRI analysis landscape.
Summary updated