Fig. 4: Comparing the temporal dynamics of decoding luminance polarity versus hue.

a Classification accuracy for the identity problems (see Fig. 3a, c). The data used to decode luminance identity was subsampled so that the maximum classification accuracy for luminance identity was the same as that of hue identity (61% of the trials were used). With the subsampled data, luminance polarity decoding peaked at 103.4 ms [100, 110], which is not different from the time of peak decoding of the complete data set (Fig. 3a). Inset shows the differences between the peaks across the 1000 bootstrapped samples across the 18 participants. Accuracy for luminance polarity peaked 18 ms before hue (p = 0.003, one-sided test comparing the bootstrapped distribution of the times to peak; the p value indicates that three of the 1000 bootstrapped times to peak of the hue problems were earlier or equal to the bootstrapped times to peak of the luminance problems.). The magnitude of classifier accuracy corresponding to the cessation of the stimulus (open arrowhead) was greater for luminance polarity than hue (p < 0.001; exact p values not possible, see “Methods” section; not adjusted for multiple comparisons; p value indicates that none of the 1000 bootstrapped decoding problems showed higher classification accuracy at stimulus cessation for hue compared to luminance polarity). Shading shows the SE of the bootstrap samples. b Classification accuracy for the generalization problems. The data used to decode the generalizing-luminance-polarity problem was subsampled to match the initial peak of generalizing hue (16% of the trials were used). The subsampled generalizing-luminance-polarity problem peaked at 99.8 ms [95, 110], which is not different from the time of peak decoding for the complete data set (Fig. 3a). Accuracy for luminance polarity decoding peaked 19 ms before hue (p = 0.017). The magnitude of classifier accuracy corresponding to the cessation of the stimulus was not greater for luminance polarity than hue (p = 0.3). p-values in b computed in the same way as for panel a.