Fig. 2.
Time measurement and production task, gain = 1 and 1.5. a Trial structure. Each trial began with the presentation of a red fixation spot and a gray “Go target”. Subjects had to measure a sample time interval ts demarcated by two flashes (“Ready” and “Set”). After Set, subjects had to press a key (“Go”) to produce an interval as close as possible to the correct interval tc = gain × ts. In the “identity” context, the correct interval was the same as ts (gain = 1), whereas in the “remapped” context, the gain was 1.5. The ratio between the distances of Go and Ready to the fixation point was equal to the gain factor. After the response, subjects received feedback via a colored circle, whose position relative to the Go target indicated the magnitude and sign of the error. The color of the circle indicated whether the error was smaller than a fixed threshold (green for a “hit,” white for a “miss”; the example illustrates a hit trial). b Behavior. Produced interval of a representative subject in the identity (gray) and remapped (red) contexts (filled circles: mean, shaded regions: mean ± one standard deviation; dashed line: correct intervals). Solid lines represent the mean responses of a Bayesian observer-actor model (see Methods) fit to the subject’s data separately for the two contexts; the dash-dot line in the gain = 1.5 condition corresponds to the prediction for the remapped context under the null hypothesis, using parameters of the model fit to the identity context (H0: no additional MTN). The subject’s behavior shows excess bias beyond what was predicted assuming no additional MTN. c √VAR vs. BIAS for the two contexts (gray: identity, dark red: remapped), as well as the prediction for the remapped context assuming no additional MTN (empty circle). This prediction underestimates RMSE indicating larger MTN for the remapped context. A substantial portion of increased RMSE was due to an increase in BIAS (“excess bias”). Dashed quarter circles illustrate combinations of BIAS vs. √VAR giving rise to equal RMSE; error bars represent 95% confidence intervals estimated using a bootstrap procedure (n = 1000)