Fig. 4: Model comparison of participants’ choices reveals divisive normalization with early and late noise.

a–d Predictions from different alternative models under the specific experimental design, with fixed pairs of targets and varying distracter mean values (x-axis) and distracter variance (color-coded lines). The early noise of V₁ and V₂ were fixed to match the overall choice accuracy observed from the data. Two levels of late noise were applied, annotated by text. e Context effects on conditional choice accuracy between the targets, aggregated by experimental conditions. All option values were scaled to the minimum target within each subject. Changing trends along the x-axis were disclosed by implementing sliding windows over V₃ (window span = 0.3, step size = 0.015). The solid lines and the shaded areas indicate the mean choice accuracy and the standard deviation at each sliding window. Linear trends were indicated by the straight lines fitted in the constrained range of scaled V₃ (0.2 to 0.8). f Sliding windows of choice accuracy across different levels of standard deviation of the scaled V₃, coded in color lines. Each line represents a subset of data within a narrow range of distracter variance (standard deviation ±0.07) around the value indicated in the legend; thus a wider window span (0.5) and coarser step size (0.03) were applied for visualization.