TY - JOUR T1 - Understanding multivariate brain activity: evaluating the effect of voxelwise noise correlations on population codes in functional magnetic resonance imaging JF - bioRxiv DO - 10.1101/592618 SP - 592618 AU - Ru-Yuan Zhang AU - Xue-Xin Wei AU - Kendrick Kay Y1 - 2019/01/01 UR - http://biorxiv.org/content/early/2019/03/31/592618.abstract N2 - Previous studies have shown that neurons exhibit trial-by-trial correlated activity and that such noise correlations (NCs) greatly impact the accuracy of population codes. Meanwhile, multivariate pattern analysis (MVPA) has become a mainstream approach in functional magnetic resonance imaging (fMRI), but it remains unclear how NCs between voxels affect MVPA accuracy. Here, we tackle this issue by combining voxel-encoding modeling and MVPA. We focus on a well-established form of NC, tuning compatible noise correlation (TCNC), whose sign and magnitude are systematically related to the tuning similarity between two units. We first replicate the classical finding that TCNCs impair population codes in a standard neuronal population. We then extend our analysis to fMRI data and show that voxelwise TCNCs do not impair and can even improve MVPA accuracy when TCNCs are strong or the number of voxels (i.e., pool size) is large. Further computational analyses demonstrate that the discrepancy between the effect of TCNCs in both neuronal and voxel populations can be explained by units’ tuning heterogeneity and pool sizes. Our results provide a theoretical foundation to understand the effect of correlated activity on population codes under different neural measurements. Our results also suggest that the conventional MVPA approach conflates several key aspects of neural coding, and future fMRI research may need to characterize different properties (e.g., NC) of multivariate responses to assess the accuracy of population codes.SIGNIFICANCE STATEMENT Noise correlations (NC) is the key component of the multivariate response distribution and thus characterizing their effect on population codes is the cornerstone for understanding probabilistic computation in the brain. Despite extensive studies on NCs in unit recording experiments, little is known with respect to their role in functional magnetic resonance imaging (fMRI). We characterize the effect of voxelwise NC by building voxel-encoding models and directly quantifying the amount of information in multivariate fMRI data. In contrast to the detrimental effects of NC implied in unit recording data, we find that the voxelwise NCs can enhance information codes in several situations. Our work provides a new approach to characterize population responses in fMRI research. ER -