What do across-subject analyses really tell us about neural coding?

A key challenge in human neuroscience is to gain information about patterns of neural activity using indirect measures. Multivariate pattern analysis methods testing for generalization of information across subjects have been used to support inferences regarding neural coding. One critical assumption of an important class of such methods is that anatomical normalization is suited to align spatially-structured neural patterns across individual brains. We asked whether anatomical normalization is suited for this purpose. If not, what sources of information are such across-subject cross-validated analyses likely to reveal? To investigate these questions, we implemented two-layered feedforward randomly-connected networks. A key feature of these simulations was a gain-field with a spatial structure shared across networks. To investigate whether total-signal imbalances across conditions-e.g. differences in overall activity-affect the observed pattern of results, we manipulated the energy-profile of images conforming to a pre-specified correlation structure. To investigate whether the level of granularity of the data also influences results, we manipulated the density of connections between network layers. Simulations showed that anatomical normalization is unsuited to align neural representations. Pattern similarity-relationships were explained by the observed total-signal imbalances across conditions. Further, we observed that deceptively complex representational structures emerge from arbitrary analysis choices, such as whether the data are mean-subtracted during preprocessing. These simulations also led to testable predictions regarding the distribution of low-level features in images used in recent fMRI studies that relied on leave-one-subject-out pattern analyses. Image analyses broadly confirmed these predictions. Finally, hyperalignment emerged as a principled alternative to test across-subject generalization of spatially-structured information. We illustrate cases in which hyperalignment proved successful, as well as cases in which it only partially recovered the latent correlation structure in the pattern of responses. Our results highlight the need for robust, high-resolution measurements from individual subjects. We also offer a way forward for across-subject analyses. We suggest ways to inform hyperalignment results with estimates of the strength of the signal associated with each condition. Such information can usefully constrain ensuing inferences regarding latent representational structures as well as population tuning dimensions.