Separating predicted content from predicted uncertainty
Most work on prediction in the brain has asked where a predicted stimulus is represented. This study asks a different question: does neural population activity also carry how uncertain that prediction is? The paper treats a prediction as having two parts, the objective content (here, a predicted location on a ring) and the subjective uncertainty attached to it.
The distinction matters because the two can be measured separately. A decoder can estimate the most likely predicted location from brain activity, and the same generative model can estimate how spread out that estimate is. The spread is the uncertainty. The authors' claim is that this spread is not noise to be averaged away but a signal the brain carries.
A dual task that makes uncertainty reportable
Twenty-two healthy participants (16 female, 18–24 years old) performed a prediction task inside a MEG scanner; data from two were excluded for head motion greater than 20 mm, leaving 20. On each trial, a dot appeared on a ring, then a second dot appeared 30°–60° away, clockwise or counterclockwise. After a blank interval, a beep cued participants to saccade to where they predicted a third dot would fall, keeping the same angular distance and direction.
Participants then adjusted a marker to report that location, and separately adjusted an arc centered on it to report their uncertainty: a longer arc meant higher uncertainty. Points were awarded only if the true target fell inside the arc, and the award decreased with arc length (10 d⁻¹, where d is arc length in polar angle degrees). This scoring rule gives participants a reason to report uncertainty honestly rather than always drawing a wide arc.
Each run had 40 trials, and participants completed 8 to 12 runs within 1.5 hours. MEG was recorded continuously at 1000 Hz with a 306-channel whole-head system.
Decoding a probability distribution, not a point
The analysis used a generative model, TAFKAP, that describes how a noisy neural population response is produced from a stimulus value. Inverting that model on single-trial MEG data yields a posterior distribution over the predicted location rather than a single best guess. The width of that posterior is the decoded uncertainty.
This is the key methodological move. Because uncertainty is read out from the shape of the reconstructed distribution, it can be quantified trial by trial, using the variability in neural activity as informative structure. The authors ground this in probabilistic population codes, the idea that neural populations encode both the most likely value and its associated uncertainty through the shape of the population response.
Decoded locations track what participants actually did
Before trusting the uncertainty estimate, the study checked the content estimate. Decoded target locations correlated with the true target locations, confirming that the decoded locations carried information about the actual stimuli. Signed decoding errors also correlated with signed behavioral errors across participants, so trial-by-trial biases in the decoded location lined up with biases in participants' reports.
The uncertainty measure was then validated against error variability. Decoded uncertainty correlated with the variability of decoding errors, quantified with a within-participant quantile-binning procedure, and with the variability of behavioral errors. In other words, when the model said the prediction was less certain, both the decoder and the participant were more variable.
Decoded uncertainty matches reported uncertainty
The central result is a positive correlation between decoded uncertainty and reported arc length across participants (r = 0.028, t(19) = 2.45, p = 0.024, Cohen's d = 0.547). The correlation coefficient is small in absolute terms because it is computed across participants on trial-wise relationships, but the effect was statistically reliable.
Three control analyses tested alternative explanations. Decoded uncertainty from the 100 ms before the second dot appeared, a window reflecting sensory processing of the first dot before a prediction could be formed, did not predict reported uncertainty (r = −0.003, p = 0.830). Controlling for trial order left the correlation significant (r = 0.031, p = 0.005), and decoded uncertainty showed only a non-significant decreasing trend over trials (p = 0.072). Controlling for absolute report error also left the correlation significant (r = 0.030, p = 0.033), while decoded uncertainty was not significantly related to absolute report error after controlling for reported arc length (p = 0.227).
A further analysis compared reliability indices. Reported uncertainty tracked report error magnitude, decoded uncertainty tracked decoding error magnitude (r = 0.314, p < 0.001), and the trial-wise reliability of the two was correlated (r = 0.032, p = 0.028). The authors note these reliability indices are composite measures, so this exploratory analysis does not identify a unique mechanism.
The uncertainty signal builds just before the response
A time-resolved analysis used a 60 ms sliding window with 8 ms steps, spanning 250 ms after the second dot's offset to the onset of the saccade cue. The correlation between decoded and reported uncertainty rose over time, reaching significance between 438 ms and 470 ms after the second dot's offset (cluster-based permutation test, all p < 0.05). The authors interpret this as uncertainty representation becoming more prominent as the prediction is formed.
Source localization with an LCMV beamformer on the 100 ms before the saccade cue pointed to a fronto-striatal cluster including ventromedial prefrontal cortex, ventral striatum, and putamen (peak MNI coordinates X = 12, Y = 15, Z = 0). Neither canonical motor-planning regions nor early visual regions showed significant activation, which the authors read as evidence against purely sensory or motor explanations. Critically, activity in that cluster did not itself correlate significantly with decoded uncertainty (p = 0.139) or reported arc length (p = 0.977), so the localization is exploratory anatomical context, not proof that the cluster encodes uncertainty.
Prediction uncertainty is not precision-weighted error
The paper draws a careful boundary. A body of work has identified neural correlates of precision in auditory, perceptual, and sensorimotor predictive processing, but those studies mainly examined precision-weighted prediction errors. This study does not directly test predictive coding models, where precision refers to the expected reliability of prediction errors and is implemented as gain modulation of prediction error responses.
The decoded uncertainty here is the uncertainty of the prediction itself, and the authors suggest it may arise from at least two sources: variability in how the first two dot locations are internally represented and maintained, and variability in extrapolating the upcoming third dot. These sources can fluctuate trial to trial without any change in external input.
A further limitation concerns the response. Because participants reported uncertainty with a motor response, the authors note that future work could dissociate prediction uncertainty from motor preparation by orthogonalizing uncertainty magnitude and motor-response magnitude, or by counterbalancing the response mapping.
Uncertainty as part of the prediction, not a byproduct
The study's distinctive contribution is that prediction uncertainty was decoded directly from neural activity, rather than inferred indirectly. Earlier Bayesian observer work proposed that perceptual estimates combine sensory input and predictions weighted by their respective uncertainties, but did not decode prediction uncertainty itself. Here the uncertainty is an empirical per-trial measure.
The authors also argue this extends probabilistic population coding from bottom-up sensory uncertainty to top-down predictive processes, supporting the hypothesis that encoding uncertainty through population activity may be a general coding scheme. That is the authors' interpretation, not something the decoding results alone establish.
For readers weighing the evidence, the priority is the comparison between the two windows: uncertainty decoded before the prediction could form did not predict reports, while uncertainty decoded later did, and the correlation peaked in the final tens of milliseconds before the response. That timing pattern, together with the control for report error magnitude, is what separates a genuine prediction-uncertainty signal from residual sensory activity or a generic sense of difficulty.
Sources used in this answer
Neural representation of the subjective uncertainty of prediction in the human brain
Prediction uncertainty was decoded trial by trial from MEG signals using a generative model, and decoded uncertainty correlated with participants' reported uncertainty, with the correlation emerging 438–470 ms after the second dot's offset, shortly before the prediction response.
