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

Perceptual decision behaviour is known to be well-captured by models based on sequential sampling and temporal integration, but doubts have been raised about the generality and identifiability of these operations. Here we used neurally-constrained modelling to probe their role and temporal extent in the uncertain case of perceptual judgments about long-duration stimuli with no physical noise but weak evidence. We found that accuracy steadily improved across four covertly-manipulated evidence durations, indicating protracted sampling, but these delayed behavioural reports alone were insufficient to establish the operation of integration or of a decision-terminating bound. We then elaborated the models to prescribe how they would generate decision variable signals as well as choices, and fit them additionally to the average dynamics of a centroparietal electroencephalographic signal ( CPP) that traces decision formation. This established the setting of a bound and ruled out some non-integration mechanisms. However, one extrema detection model, which evokes a stereotyped signal flagging the first bound-exceeding sample, rivalled the integration model in reproducing the evidence-dependent buildup dynamics of the CPP, alongside behavioural data. Moreover, the two models captured different features of neural motor preparation signals but neither captured all of them. We discuss the implications for the generality of integration and the technical challenges of neurally-informed modelling given limited behavioural data.