Neurocomputational Underpinnings of Expected Surprise

Abstract

Predictive coding accounts of brain functions profoundly influence current approaches to perceptual synthesis. However, a fundamental paradox has emerged, that may be very relevant for understanding hallucinations, psychosis, or cognitive inflexibility: in some situations, surprise or prediction error-related responses can decrease when predicted, and yet, they can increase when we know they are predictable. This paradox is resolved by recognizing that brain responses reflect precision-weighted prediction error. This presses us to disambiguate the contributions of precision and prediction error in electrophysiology. To meet this challenge for the first time, we appeal to a methodology that couples an original experimental paradigm with fine dynamic modeling. We examined brain responses in healthy human participants (N = 20; 10 female) to unexpected and expected surprising sounds, assuming that the latter yield a smaller prediction error but much more amplified by a larger precision weight. Importantly, addressing this modulation requires the modeling of trial-by-trial variations of brain responses, that we reconstructed within a fronto-temporal network by combining EEG and MEG. Our results reveal an adaptive learning of surprise with larger integration of past (relevant) information in the context of expected surprises. Within the auditory hierarchy, this adaptation was found tied down to specific connections and reveals in particular precision encoding through neuronal excitability. Strikingly, these fine processes are automated as sound sequences were unattended. These findings directly speak to applications in psychiatry, where specifically impaired precision weighting has been suggested to be at the heart of several conditions such as schizophrenia and autism.SIGNIFICANCE STATEMENT In perception as Bayesian inference and learning, context sensitivity expresses as the precision weighting of prediction errors. A subtle mechanism that is thought to lie at the heart of several psychiatric conditions. It is thus critical to identify its neurophysiological and computational underpinnings. We revisit the passive auditory oddball paradigm by manipulating sound predictability and use a twofold modeling approach to simultaneous EEG-MEG recordings: (1) trial-by-trial modeling of cortical responses reveals a context-sensitive perceptual learning process; (2) the dynamic causal modeling (DCM) of evoked responses uncovers the associated changes in synaptic efficacy. Predictability discloses a link between precision weighting and self-inhibition of superficial pyramidal (SP) cells, a result that paves the way to a fine description of healthy and pathologic perception.

Keywords: Bayesian learning; EEG-MEG fusion; dynamic causal modeling; mismatch negativity; predictive coding; trial-by-trial modeling.

Figure 1.

A, Experimental design. Schematic view of the predictability manipulation applying to typical oddball sound sequences. PC (left, green) involves cycles of ordered transitions between segments of repeating standards (chunks), which become shuffled in the UC (right, red). Deviant probability remains the same in both context (p = 1/6). Gray rectangles delineate an exemplary cycle for both sequences. S: Standard, D: Deviant. B, Group-average difference responses. For each modality (EEG, left; MEG, right), scalp maps of grand-average difference (deviant – standard) responses at latency showing a significant predictability effect for both contexts (PC: green; UC: red). Middle plots: traces at sensors showing a significant MMN reduction under predictability (EEG: FCz; MEG: MLT16; location on related scalp map is represented by a black circle); gray areas indicate the significant time intervals for these sensors (permutation tests with multiple comparison correction, p < 0.05).

Figure 2.

Neurocomputational framework. Representation of the current approach deployed at both the cognitive and physiological levels to address the automatic adaptive learning at play during auditory processing, and to disambiguate the mapping of precision weights and prediction errors onto physiological responses. First-level analysis (upper panel): the expected perceptual learning of the oddball rule is first tested at the computational level (left) as well as its physiological implementation within a fronto-temporal hierarchy (right). Second-level analysis (lower panel): adaptation of this learning to the manipulation of predictability is then tested through the examination of model parameters for each condition (UC, PC), both at the computational (left) and physiological (right) levels. Gray boxes highlight the specific differences that were tested. Different learning time constants τ (left) would support hierarchical learning with opposite effects on precision weighting and prediction errors, that are testable (hence separable) using DCM. First-level and second-level rules are described in Figure 1A. Dynamic models: pl (perceptual learning), dcm (dynamic causal model); cortical sources: HG (Heschl's gyrus), PP (planum polare), IFG (inferior frontal gyrus), SF (superior frontal); experimental contexts: PC/pc (predictable context), UC/uc (unpredictable context). D: deviant. Forward/self: DCM forward/self-inhibition connection strength parameters.

Figure 3.

Perceptual learning models. A, Each cluster of interest is represented (orange) over the inflated cortical surface of the SPM template brain (Mattout et al., 2007). These six clusters are left, right HG (lHG, rHG), left, right PP (lPP, rPP) and left, right IFG (lIFG, rIFG). Total number of nodes in each cluster is indicated in parenthesis. B, Bayesian surprise as a function of $\mathit{\tau }$ (arbitrary units, a.u.). Illustration of different BS trajectories obtained with varying $\tau$, for the first 100 stimuli of a typical UC oddball sequence. Two comments should be made: (1) BS decreases as $\tau$ increases and (2) whatever $\tau$, BS is larger for deviants (D, black squares) than for standards (S, gray squares). C, Learning model predictions of the MMN amplitude as a function of $\mathit{\tau }$ (group average) for UC (red) and PC (green) sound sequences (see Materials and Methods). Note that in both contexts, MMN amplitude decreases similarly as $\tau$ increases.

Figure 4.

DCM. A, Cortical sources for DCM analysis. Each source is indicated schematically with orange dots on the inflated cortex, with corresponding MNI coordinates (mm) in parenthesis. B, Model families. Upper row, Schematic view of the five model families designed to test DCM architecture in deviance processing. Bottom row, The two model families of DCM input, HG and HG-IFG. Color codes of extrinsic connections (conn.) and DCM source (or node) are provided in the legend. C, Model spaces. Network structure analysis (left): DCM specifications for each of the 36 models (in columns). Frontal, backward and intrinsic trial-specific gains, as well as modulatory connections correspond to binary options (enabled = 1, disabled= 0) applying to the entire network. Standard-to-deviant modulation analysis (right), following the same logic of display. Backward trial-specific gains were disabled or applied onto either extrinsic or intrinsic connections depending on modulatory connections (as detailed in the main text).

Figure 5.

Computational modeling of deviance processing. Family-wise Bayesian model comparison. For each cluster and at each time sample, family inference provides the estimated posterior family exceedance probability of each model family (fam_null: black, fam_noL: orange, fam_L: pink).

Figure 6.

DCM of deviance processing, using p-MEEG. A, Family inference for DCM architecture: family exceedance probabilities (upper left) and corresponding network for the winning family A5 (lower left). B, Family inference for DCM inputs. C, Family inference for standard-to-deviant trial-specific modulation for the forward connectivity. Family exceedance probabilities of the model families with disabled (0) and enabled modulation (1) is provided (left) and group-level BMA estimates of gain values (in condition UC) averaged over the DCM network is represented (right). Prior value of gain (light purple) was set to 1 assuming no standard-to-deviant modulation. Labels for the cortical sources and model families are provided in the main text.

Figure 7.

Effect of predictability on auditory processing. A, Cognitive modeling (perceptual learning). Left plot, Effect of predictability on learning parameter $\tau$. Posterior estimates of $\tau$ averaged at the group level and over the six clusters exhibited a significant difference between conditions. Middle plot, Downweighting of past observations obtained with the posterior estimates of $\tau$ for conditions UC (red) and PC (green), respectively. Right plot, Group-level observed (Obs.) and predicted (Pred.) MMN amplitude within each cluster, and for condition UC (red) and PC (green). Each value gathers time samples that exhibited a significant learning effect in the first-level analysis and was computed following the scheme described for the pseudo-MMN computation (detailed in the main text). B, Physiologic counterpart (DCM). Effect of predictability onto effective connectivity obtained with the fusion of EEG and MEG DCMs. Left and middle plots, Posterior estimates of forward and self-inhibition strengths measured in both conditions (UC: red, PC: green), averaged over the group and over the connections within the DCM. SP: superficial pyramidal; a.u.: arbitrary units. Right plot, Source-based self-inhibition strengths showing a significant predictability effect. Values for each condition (UC: red, PC: green) are displayed over a typical DCM graph representation. Labels of cortical sources as in Figure 4.