Attention amplifies neural representations of changes in sensory input at the expense of perceptual accuracy

Attention enhances the neural representations of behaviorally relevant stimuli, typically by a push–pull increase of the neuronal response gain to attended vs. unattended stimuli. This selectively improves perception and consequently behavioral performance. However, to enhance the detectability of stimulus changes, attention might also distort neural representations, compromising accurate stimulus representation. We test this hypothesis by recording neural responses in the visual cortex of rhesus monkeys during a motion direction change detection task. We find that attention indeed amplifies the neural representation of direction changes, beyond a similar effect of adaptation. We further show that humans overestimate such direction changes, providing a perceptual correlate of our neurophysiological observations. Our results demonstrate that attention distorts the neural representations of abrupt sensory changes and consequently perceptual accuracy. This likely represents an evolutionary adaptive mechanism that allows sensory systems to flexibly forgo accurate representation of stimulus features to improve the encoding of stimulus change.

The median value of direction tuning curve shifts, induced by attended and unattended direction changes of +25 o for each monkey individually. p is the p-value of the two-sided Wilcoxon signed rank test for distributions with zero median. p* is the p-value of the paired two-sided Wilcoxon signed rank test for the null hypothesis that the difference between paired samples comes from a distribution with zero median.

Supplementary Note 5
Control analysis 1: We convolved spike trains with a Gaussian kernel to compute the spike density functions. As summarized in Supplementary Table 2, the shifts of direction tuning curves in both attended and unattended conditions are independent of the standard deviation used to smooth the spike trains.  Table 2. The median value of direction tuning curve shifts induced by attended and unattended direction changes of +25 o (n = 52 cells). p is the p-value of the two-sided Wilcoxon signed rank test for distributions with zero median. p* is the p-value of the paired two-sided Wilcoxon signed rank test for the null hypothesis that the difference between paired samples comes from a distribution with zero median.

Control analysis 2:
To make sure that direction tuning curve shifts do not depend on the tuning characteristics of MT cells, we repeated the analysis by applying the following inclusion criteria: (1) neurons were highly direction selective (response to the preferred direction was at least 5 times larger than the response to the anti-preferred direction), (2) the neurons were well-tuned (goodness of fit > 0.7). 25 out of 52 cells fulfilled these criteria. The results of this analysis (Supplementary Table 3: Control analysis 2) were consistent with those presented before (Fig. 2b).

Control analysis 3:
We performed a control analysis to show that the direction tuning curve shifts are not affected by the selection of the post-change time window. This analysis used an analysis time window from +150 to +250 ms following the direction change to assess the post-change direction tuning curves. The results (Supplementary Table 3: Control analysis 3) confirmed those reported in the article (Fig. 2b).  Table 3. The median value of direction tuning curve shifts induced by attended and unattended direction changes of +25 o . p is the p-value of the two-sided Wilcoxon signed rank test for distributions with zero median. p* is the p-value of the paired two-sided Wilcoxon signed rank test for the null hypothesis that the difference between paired samples comes from a distribution with zero median.

Control analysis 4:
We demonstrated that the direction tuning curve shift induced by attended and unattended direction changes is independent of the symmetry of the function employed to fit the neural responses to the different directions. We fitted the pre-and post-change data separately with skewed von Mises functions (see Methods). In two separate analyses, we included tuned (goodness of fit greater than 0.7), direction-selective (response to preferred direction at least 5-fold larger than response to anti-preferred direction) cells based on the fit of the data with each of symmetric (Supplementary Fig. 5a) and skewed ( Supplementary  Fig. 5b) von Mises functions each time. The results show that regardless of the symmetry of the function used to model neural responses to different directions, post-change direction tuning curves in attended condition are shifted and the shift is significantly larger than be explained by sensory adaptation in unattended condition. Figure 5. Distribution of direction tuning curve shifts caused by attended (blue) and unattended (red) direction changes of +25 o across MT cells. The direction tuning curves of each cell were modeled by skewed von Mises functions (see Methods). The median values of the shifts and the corresponding p values are labeled (vertical dashed lines; * p < 0.05; ** p < 0.01; *** p < 0.001; **** p < 0.0001 two-sided Wilcoxon signed rank test for distribution with a median of 0; # p < 0.05; ## p < 0.01 paired two-sided Wilcoxon signed rank test). Unfilled bars represent cells for which the magnitude of tuning shift was greater than 40 o . (a) Inclusion was based on the results of fitting the data with the skewed von Mises functions: 34 out of 52 neurons fulfilled the inclusion criteria. Tuning shifts were significant in both attended (p = 0.00004, two-sided Wilcoxon signed rank test for distribution with zero median) and unattended (p = 0.009, two-sided Wilcoxon signed rank test for distribution with zero median) conditions. Moreover, the attention-related increase in the tuning shift was statistically significant (p = 0.02, paired two-sided Wilcoxon signed rank test for the null hypothesis that the difference between paired samples comes from a distribution with zero median). (b) Inclusion was based on the results of fitting the data with the (symmetric) von Mises functions: 25 out of 52 neurons fulfilled the inclusion criteria. Tuning shift was significant in both attended (p = 0.0005, two-sided Wilcoxon signed rank test for distribution with zero median) and unattended (p = 0.03, two-sided Wilcoxon signed rank test for distribution with zero median) conditions. Moreover, the attention-related increase in tuning shifts was statistically significant (p = 0.009, paired two-sided Wilcoxon signed rank test for the null hypothesis that the difference between paired samples comes from a distribution with zero median). Source data are provided as a Source Data file. Figure 6 illustrates how the direction tuning shift of single cells induced by an abrupt direction change might result in the overestimation of a perceived direction change in both attended (left column) and unattended (right column) conditions. Supplementary Figure 6a plots the tuning curves of four representative neurons (different colors) prior and subsequent to the direction change of +25 o with solid and dashed curves, respectively. In each condition, we used the median of the direction tuning parameters across all cells in the corresponding condition to create the tuning curves of the representative cells. This is justified because there was a weak correlation between the pre-change preferred directions of MT cells and the direction tuning shifts induced by the direction change (Pearson r = 0.27, p = 0.053 for attended; and Pearson r = 0.08, p = 0.6 for unattended. Spearman r = 0.19, p = 0.2 for attended; and Spearman r = 0.12, p = 0.4 for unattended). Supplementary Figure  6b depicts the population response profiles to a stimulus moving in 25 o (solid black curve) and the direction change of 25 o (dashed gray curve). Vertical solid black lines show the motion direction, the downward gray triangle and the vertical gray dashed lines indicate the pre-and post-change motion directions, respectively. The population post-change response profiles show peaks shifted away from the post-change direction of +25 o . This shift has the same magnitude but opposite sign as the neural direction tuning curve shifts. As single cells exhibit larger tuning shifts in the attended condition than the unattended condition, the postchange response profile in the attended condition is shifted more than that in the unattended condition.