Extended Data Fig. 2: Orientation and uncertainty decoding performance.

The orientation of the presented stimulus, and associated uncertainty, decoded from activity patterns in areas V1-V3. (a) Orientation decoding performance was quantified by means of the circular equivalent of the Pearson correlation coefficient between presented and decoded orientations. Correlation coefficients were computed for each subject individually and then averaged across subjects (N = 32). Presented and decoded orientations were significantly correlated (z = 83.58, p < 0.001, r = 0.60, 95% CI = [0.58, 0.61]). (b-d) To assess the degree to which the decoder captured uncertainty contained in neural population activity, we compared decoded uncertainty to behavioral variability, the rationale being that a more precise representation in cortex should also result in more precise behavioral estimates (see also¹⁰). (b) Corroborating our approach, we found that decoded uncertainty was greater for oblique compared to cardinal orientation stimuli (correlation distance-to-cardinal and decoded uncertainty: z = 2.95, p = 0.002, ρ = 0.025, 95% CI = [0.0083 0.041]). This finding was paralleled by the imprecision in observer behaviour (correlation distance-to-cardinal and behavioral variability: t(287) = 13.60, p < 0.001, r = 0.63, 95% CI = [0.55, 0.69]). (c-d) In addition, behavioral orientation responses were more precise when the decoded probability distributions indicated greater certainty in cortex, (c) both across orientation stimuli (correlation decoded uncertainty and behavioral variability: t(287) = 2.30, p = 0.011, r = 0.13, 95% CI = [0.019, 0.25]), and (d) when controlling for orientation (t(286) = 1.68, p = 0.047, r = 0.099, 95% CI=[−0.017, 0.21]). Altogether, this further underscores the validity of the decoding approach and shows that decoded uncertainty reliably characterizes the degree of imprecision in cortical representations of the stimulus (see^10,18 for further proof of this approach). Note that these are partial residual plots, which is why the data is centered around 0. Error bars (a-b) represent ± 1 s.e.m. (c-d) Shades of red indicate ten equal-size bins of increasing decoded uncertainty, dots represent individual observers (N = 32).