Neural surprise in somatosensory Bayesian learning

Abstract

Tracking statistical regularities of the environment is important for shaping human behavior and perception. Evidence suggests that the brain learns environmental dependencies using Bayesian principles. However, much remains unknown about the employed algorithms, for somesthesis in particular. Here, we describe the cortical dynamics of the somatosensory learning system to investigate both the form of the generative model as well as its neural surprise signatures. Specifically, we recorded EEG data from 40 participants subjected to a somatosensory roving-stimulus paradigm and performed single-trial modeling across peri-stimulus time in both sensor and source space. Our Bayesian model selection procedure indicates that evoked potentials are best described by a non-hierarchical learning model that tracks transitions between observations using leaky integration. From around 70ms post-stimulus onset, secondary somatosensory cortices are found to represent confidence-corrected surprise as a measure of model inadequacy. Indications of Bayesian surprise encoding, reflecting model updating, are found in primary somatosensory cortex from around 140ms. This dissociation is compatible with the idea that early surprise signals may control subsequent model update rates. In sum, our findings support the hypothesis that early somatosensory processing reflects Bayesian perceptual learning and contribute to an understanding of its precise mechanisms.

Surprise: a unified theory and experimental predictions

“…To do so, we consider a generative model of a volatile environment (Fig. 1A) that captures a few key features of daily life and unifies many existing model environments in neuroscience and psychology (Behrens et al, 2007;Daw et al, 2011;Findling et al, 2021;Gijsen et al, 2021;Gläscher et al, 2010;Glaze et al, 2015;Heilbron & Meyniel, 2019;Horvath et al, 2021;Huys et al, 2015;Liakoni et al, 2021;Mars et al, 2008;Meyniel et al, 2016;Nassar et al, 2012;Nassar et al, 2010;Ostwald et al, 2012;Wilson et al, 2013;Xu et al, 2021). The generative model describes the subjective interpretation of the environment from the point of view of an agent (e.g., a human participant or an animal).…”

“…Specifically: B. Standard generative model for studying passive learning in volatile environments (Adams & MacKay, 2007;Fearnhead & Liu, 2007;Liakoni et al, 2021;Nassar et al, 2012;Nassar et al, 2010;Wilson et al, 2013), C. Generative model corresponding to variants of bandit and reversal bandit tasks (Behrens et al, 2007;Findling et al, 2021;Horvath et al, 2021), where the cue variable X t = A t is a participant's action, D. Generative model for modeling human inferences about binary sequences (Gijsen et al, 2021;Maheu et al, 2019;Meyniel et al, 2016;Modirshanechi et al, 2019;Mousavi et al, 2020), and E. classic Markov Decision Processes (MDPs) (Daw et al, 2011;Gläscher et al, 2010;Huys et al, 2015;Lehmann et al, 2019;Schultz et al, 1997;Sutton & Barto, 2018), where the cue variable X t = (A t−1 , Y t−1 ) consists of previous action and observation. See Appendix A: Special cases and links to related works for details.…”

“…4 variable Z (e.g., the amplitude of the EEG P300 component (Kolossa et al, 2015;Meyniel et al, 2016)) is sensitive to or representative of surprise. Given two measures of surprise S and S , a typical experimental question is which one of them (if any) more accurately explains the variations of the variable Z (Gijsen et al, 2021;Kolossa et al, 2015;Mousavi et al, 2020;Ostwald et al, 2012;Visalli et al, 2021); see Fig. 2A1.…”

Human Belief State-Based Exploration and Exploitation in an Information-Selective Symmetric Reversal Bandit Task

“…Expected Bayesian surprise is one of many quantities that have been proposed to formally capture immediate information gain and thus myopic exploration (Schwartenbeck et al, 2019). As alluded to in the Introduction, we here opted for Bayesian surprise due to its putative representation in the human neurocognitive system (Itti & Baldi, 2009;Ostwald et al, 2012;Gijsen et al, 2020).…”

Gated recurrence enables simple and accurate sequence prediction in stochastic, changing, and structured environments

“…We also compared them to two types of heuristics which perform very well in this environment: the classic 'delta-rule' heuristic ( Rescorla & Wagner, 1972 ;R. S. Sutton & Barto, 1998 ) and the more accurate 'leaky' heuristic ( Gijsen et al, 2021 ;Heilbron & Meyniel, 2019 ;Meyniel et al, 2016 ;Yu & Cohen, 2008 ) (see Methods for details). To test the statistical reliability of our conclusions, we trained separately 20 agents of each type (each type of network and each type of heuristic).…”