Feature-based attention (FBA) selectively enhances processing of an attended feature at the expense of unattended or task-irrelevant features. Recent studies showed that FBA modulates the perceptual space with both a monotonic profile (i.e., feature-similarity gain) and a non-monotonic profile (i.e., surround suppression). A significant question arises regarding the neural mechanism of the non-monotonic surround suppression effect. Previous studies have suggested that two candidate neuronal mechanisms could underlie these attentional modulations: a shift of neuronal tuning preference toward the attended feature, or a multiplicative gain modulation that scales the overall responses without changing their tuning property. Yet the empirical evidence for these mechanisms is still lacking. In the current work, we explored how these neuronal mechanism manifest at the level of fMRI BOLD measurement using a simulation approach. Specifically, we employed an encoding/decoding approach by first simulating voxel responses from neuronal population assuming either mechanism and then applying a regression-based inverted encoding model (IEM) and a Bayesian method to decode population representations. We found that both methods captured the signature patterns associated with these different neuronal mechanisms. In our second aim, we systematically varied the correlation structure of voxel noise to further compare these different multivariate methods in a stimulus classification task. Our results showed a clear advantage of the Bayesian method over IEM, suggesting that the Bayesian method was superior for deciphering neural representation given the prevalent noise correlation and variable tuning width in the brain. In sum, our current simulation work may provide a proof of concept for future empirical studies investigating cortical mechanism of FBA using non-invasive methods, as well as guidance for choosing suitable methods in these investigations.
Read
