Abstract
Both humans and animals have been shown to have temporal biases in perceptual decision-making tasks, ranging from an overweighting of evidence early in the trial (primacy bias) to an overweighting of evidence late in the trial (recency bias), even when the optimal strategy requires an equal weighting. Recent work showed how the specific bias observed in an experiment can be predicted from the statistics of the sensory stimulus, and that the observed primacy effects are due to a type of confirmation bias that arises from approximate hierarchical inference that is ubiquitous during decision-making (Lange et al. 2021). Here, we mapped out the parameters and properties of this confirmation bias by a series of five new experiments, each involving 10 human observers performing a coarse orientation discrimination task. In the first sequence of experiments, we kept the number of independent pieces of evidence (stimulus frames) constant at 10, while we varied the duration of each frame from 42ms to 83ms to 166ms. In the second sequence of experiments, we kept the duration of the stimulus fixed at 420ms while we varied the number of stimulus frames from 5 to 10 to 25 . First, we found that the primacy bias is remarkably robust, dominating the temporal dynamics in all of the above experiments involving frame durations from 16 ms to 166ms, and total stimulus durations from 420ms to 1.66s. Importantly, we found it to have similar strength across all durations when measured in terms of stimulus frames rather than physical time, both supporting a computational rather than biophysical basis, and demonstrating that the hierarchical inference computations adapt to the temporal statistics of the inputs. These results also argue against an internal integration-to-bound process as the cause of the empirically observed primacy bias.