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 the operation of protracted sampling, and of temporal integration as opposed to non-integration strategies like extremum-detection, may not be as ubiquitous or as experimentally identifiable as once thought. Here, we attempted to resolve such mechanistic details in the context of a task for which the role and extent of temporal integration is uncertain - discriminating feature-differences in a long-duration, 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. Jointly fitting CPP dynamics as well as accuracy uncovered features such as the setting of a bound, which were unresolvable in fits to accuracy alone. However, a convincing joint neural-behaviour fit could be achieved not only by a temporal integration model, but also by an extrema detection model that evoked a stereotyped signal ‘flagging’ the first single-sample bound crossing. We explore features of neural motor preparation signals that align with the predictions of either model, and reflect on challenges for neurally-constrained modelling when addressing scenarios that produce limited behavioural data.