Jeffrey Nestor¹ (), Karen Tian¹, Angus Chapman¹, Jenny Motzer¹, Rachel Denison¹; ¹Boston University

Motivation: Humans are constantly faced with uncertainty about the state of the world due to neural and environmental noise. One can compensate for sensory uncertainty by integrating prior information about a stimulus feature, but it is unclear how uncertainty representations unfold over time and dynamically incorporate prior information. Here we first sought to establish whether trial-by-trial uncertainty representations are decodable from electroencephalography (EEG) data. Then we leveraged uncertainty decoding to ask whether and how prior information sharpens sensory representations over time. Methods: We recorded EEG data while, on each trial, participants estimated the location of a low-contrast grating target and then reported their positional uncertainty by adjusting an arc to bet on the precision of their estimate. Targets were either preceded by a spatial cue which provided a Gaussian prior over target location or an uninformative neutral cue. We trained probabilistic decoders on EEG data which estimated a posterior probability distribution over target locations for each trial and each time point, allowing us to extract time-resolved location predictions as well as estimates of sensory uncertainty. Results: Behaviorally, reported uncertainty correlated across trials with position error. In neutral trials, EEG decoders predicted target location above chance starting ~150 ms after stimulus onset and continuing through the response period. Decoded uncertainty predicted decoding error both across time points and across trials. Importantly, we also observed trial-by-trial correlations between reported uncertainty and decoded uncertainty, demonstrating that uncertainty decoded from EEG meaningfully reflected sensory uncertainty. In spatially cued trials, decoded uncertainty was significantly reduced relative to neutral trials even before target onset, showing that the brain represents predictive spatial information in preparation for the target. Conclusion: Dynamic, behaviorally-relevant representations of sensory uncertainty can be decoded from EEG data and are shaped by prior information in an anticipatory fashion.

Acknowledgements: Startup funding from Boston University to R.D., National Defense Science and Engineering Graduate Fellowship to K.T.