Predictions through evidence accumulation over time

Abstract

It has been proposed that the brain specializes in predicting future states of the environment. These predictions are probabilistic, and must be continuously updated on the basis of their mismatch with actual evidence. Although electrophysiological data disclose neural activity patterns in relation to predictive processes, little is known about how this activity supports prediction build-up through evidence accumulation. Here we addressed this gap. Participants were required to make moment-by-moment predictions about stimuli presented in sequences in which gathering evidence from previous items as they were presented was either possible or not. Two event-related potentials (ERP), a frontocentral P2 and a central P3, were sensitive to information accumulation throughout the sequence. Time-frequency (TF) analyses revealed that prediction build-up process also modulated centrally distributed theta activity, and that alpha power was suppressed in anticipation to fully predictable stimuli. Results are in agreement with the notion of predictions as probability distributions and highlight the ability of observers to extract those probabilities in a changing environment and to adjust their predictions consequently.

“Hit the missing stimulus”. A simultaneous EEG-fMRI study to localize the generators of endogenous ERPs in an omitted target paradigm

“…The pP2 component in omitted target paradigms has never been reported before. Apparently, it is part of the stream of cognitive processing for the detection of salient events relevant for the current task: its role has been associated 64,65 to accumulation of sensory evidence leading to target recognition and decision making (a process referred to as “perceptual decision making” 66 ). For these reasons, the pP2 was also described as the correlate of the stimulus-response mapping process 67 .…”

Weak proactive cognitive/motor brain control accounts for poor children’s behavioral performance in speeded discrimination tasks

“…In visuo-motor tasks after stimulus onset, in addition to the well documented sensorial (P1 and N1) and endogenous (N2 and the P3) components, three more components have been identified: the prefrontal N1, P1 and P2 (pN1, pP1 and pP2 respectively) peaking between 110 and 300 ms after the stimulus onset and localized in the rostral part of anterior insula Sulpizio et al, 2017). While the pN1 and pP1 are related to top-down perceptual processing associated with stimulus physical salience and awareness and sensory-motor integration respectively Sanchez-Lopez et al, 2017), the pP2 has been associated with evidence accumulation process, that is the efficacy of the stimulus-response mapping Darriba & Waszak, 2018;Di Rollo et al, 2016;Perri, Berchicci, Lucci, Spinelli, & Di Russo, 2015b, 2015a. These anterior insular ERP components were recently characterized by the present research group; however, in the past, several research groups found that anterior ERP components were generically localized in frontal areas (Foxe & Simpson, 2002;Potts, Liotti, Tucker, & Posner, 1996).…”

Neural Encoding of Auditory Statistics

“…The CPP, on the other hand, is later, having similar latency and topography to the P3b response, which has been linked to context updating in working memory because of expectation violations ( Donchin and Coles, 1988 ; Polich, 2007 ; Romero-Rivas et al, 2018 ; Darriba and Waszak, 2018 ). Additionally, in contrast to the MMN response, the P3b is associated with changes in global regularities encompassing higher-order statistics ( Bekinschtein et al, 2009 ; Wacongne et al, 2011 ; Chennu et al, 2013 ) and more complex stimuli ( Chernyshev et al, 2016 ).…”

“…To illuminate how this joint representation is constructed, we used a computational model grounded in Bayesian accounts of statistical predictive coding in the brain ( Knill and Pouget, 2004 ; Tenenbaum et al, 2006 ; Daunizeau et al, 2010 ; Pieszek et al, 2013 ; Wilson et al, 2013 ). This model embodies several theoretical principles of predictive processing: that the brain maps sensory inputs onto compact summary statistics ( Brady et al, 2009 ; McDermott et al, 2013 ); that the brain entertains multiple hypotheses or interpretations of sensory information ( Mumford, 1991 ); and that the brain incrementally updates its predictions over time based on evidence from new inputs ( Darriba and Waszak, 2018 ). The D-REX model and its multifeature extension presented above represent a computational instantiation of these theoretical principles that can be used to interpret experimental results.…”