Neurally-Informed Models of Protracted Sequential Sampling of Long Noise-free Stimuli

Sequential sampling models have provided a rich theoretical framework in the study of perceptual decision-making and quantitatively capture diverse behavioural data. However, research has increasingly highlighted that for the long-duration, statistically stationary stimuli commonly studied, it is difficult to definitively establish the operation of protracted sampling or temporal integration as opposed to non-integration strategies like extremum-detection, especially when based on behaviour alone. Here, we attempted to resolve such mechanistic details in the interesting case of judging subtle feature-differences (contrast) in a long (1.6-sec), noiseless stimulus, by jointly analysing the accuracy of delayed reports and the dynamics of a centroparietal electroencephalographic signal (‘CPP’) reflecting decision formation. Accuracy steadily increased across four covertly-manipulated evidence durations and the CPP remained elevated throughout the stimulus period, together indicating protracted sampling. Models fit to CPP dynamics as well as accuracy resolved more details than those fit to accuracy alone, such as the setting of a bound. However, a convincing joint neural-behaviour fit could be achieved not only by a temporal integration model, but also by an extremum detection model that evoked a stereotyped signal ‘flagging’ the first single-sample bound crossing. This illuminates interesting, testable alternatives to integration, and the limited extent to which combining neural decision dynamics with behavioural information can uniquely identify underlying mechanisms in some behavioural scenarios.