Abstract
Noise in the nervous system makes it impossible to infer with absolute precision the presented stimulus from a cortical response. Rather, the information contained in neural activity is uncertain. How best to characterize this uncertainty in cortical stimulus representations? We recently demonstrated that the trial-by-trial imprecision in perceptual representations can be reliably extracted from the human visual cortex, using fMRI in combination with a novel probabilistic decoding approach (van Bergen, Ma, Pratte & Jehee, 2015, Nature Neuroscience). Here, we present two new developments in this probabilistic analysis technique, which yield impressive improvements in decoding performance. First, we improved the estimation of an important component of the decoding model, namely the spatial noise correlation structure between fMRI voxels. Second, we implemented sampling techniques to appropriately account for the full range of decoding models that are plausible given a set of (noisy) training data. We applied this augmented decoding algorithm to cortical activity measured in an fMRI experiment, in which human participants viewed grating stimuli and reported their orientations. We found that the ability to decode stimulus information from visual cortex improved markedly on several fronts. First, the decoder’s estimates of the presented orientations were much more precise than observed previously, with mean errors decreasing by about 30%. Second, the augmented decoder was better able than its predecessor to gauge the uncertainty with which stimulus information is represented in cortical activity. Specifically, we observed a significantly stronger link between the degree of uncertainty decoded from visual cortex and variability in the observers’ behavioral responses. Together, these findings indicate that the augmented probabilistic decoder provides an exceptionally clear window onto the precision with which stimulus information is encoded in cortical activity.
Acknowledgement: This work was supported by ERC Starting Grant 677601 to J.J.