Figure 3.

Multisensory bounded evidence accumulation model for unifying choice, reaction time (RT), and confidence. (a) Noisy momentary evidence is constructed as a reliability-weighted sum of heading evidence from visual and vestibular cues (red and blue distributions in leftmost panel). In this example, the two cues are conflicting and of different reliability (variance), and the resulting combined distribution (green curve) is biased towards the more reliable cue (here, shifted rightwards). Evidence favouring each of the two alternatives is accumulated separately and the choice is determined by which accumulator reaches its bound first; RT is given by the time of bound crossing (plus a small non-decision time). Two simulated trials are depicted. In both, the rightward accumulator wins the race, but in example trial 2, the decision maker is less confident because the losing accumulator was closer to the bound as compared to trial 1. (b) The model asserts that the brain has implicit knowledge of the relationship between the amount of evidence gathered by the losing accumulator and the probability (shown as the log odds) of making a correct choice, and uses this knowledge to compute confidence. (c) Confidence rating (mean ± s.e.) as a function of RT quantile, plotted separately for the three different heading angle magnitudes (|hdg|, leftward and rightward pooled) and three modality conditions: vestibular (platform motion only), visual (optic flow only), and combined. Data are from the same dataset as figure 2 (n = 5 subjects). (d) Simulations from the model also exhibit a negative relationship between confidence and RT, conditioned on |hdg|.