Predictive Coding and Hierarchical Sensory Prediction in the Human Brain

Abstract

The predictive processing framework posits that the brain continuously generates predictions about incoming sensory information. We used 7T fMRI with high-resolution multiband sequences to examine hierarchical prediction error signals across visual cortex in 156 participants viewing dynamic visual scenes. Bayesian multilevel modelling revealed significant prediction error activity in early visual areas (V1/V2) and stronger feedback-related activity in higher areas, consistent with predictive coding theory. These findings advance our understanding of hierarchical neural computations underlying perception.

Introduction

The brain is not a passive receiver of sensory information but an active predictor continuously generating expectations about the world. The predictive processing framework, formalized in computational neuroscience, suggests that perception arises from the minimization of prediction errors across hierarchical neural systems (Friston, 2010; Clark, 2013). This theory has profound implications for understanding both normal perception and pathological conditions including schizophrenia and autism spectrum disorders.

Recent advances in neuroimaging methodology—particularly high-field 7T fMRI and multiband acquisition—enable unprecedented spatial and temporal resolution of neural activity in sensory hierarchies. However, direct evidence for hierarchical prediction error signals in human visual cortex remains limited. The present study bridges this gap by examining the laminar and regional distribution of prediction error signals during naturalistic visual perception.

Method

Participants

We recruited 156 healthy adults (aged 18–40 years) from the Ottawa community via Prolific Academic. Inclusion criteria required normal or corrected-to-normal vision and no history of neurological or psychiatric disorder. Participants were pre-registered on the Open Science Framework (protocol available at osf.io/7xkqm). All procedures were approved by the institutional review board and informed consent was obtained in writing.

Procedure

Participants underwent a single 60-minute 7T fMRI scanning session on a Siemens MAGNETOM scanner equipped with a 32-channel receive coil. During scanning, participants viewed three 6-minute video clips of natural scenes (forests, urban environments, face interactions) while fixating on a central crosshair. Video frames were presented at 30 Hz with occasional jitter (±80 ms) to induce prediction errors. High-resolution multiband echo-planar images were acquired (1.2 mm isotropic voxels, TR = 1.5 s, multiband factor = 3). Anatomical T1-weighted images were also acquired for co-registration.

Analysis employed custom SPM12 code for motion correction, normalization to MNI space, and laminar-segmented region-of-interest analysis. We modelled prediction errors as the squared Euclidean distance between successive frames in a 64-dimensional feature space extracted from a pretrained convolutional neural network (ResNet-50). Bayesian multilevel regression with weakly informative priors estimated the relationship between prediction error magnitude and fMRI signal across visual areas.

Results

Across all participants, we identified robust prediction error-related activity in early visual cortex (V1/V2), with peak activation in superficial layers consistent with feedforward input (t = 4.23, p < 0.001, posterior probability = 0.94). Notably, higher visual areas (V3/V4/LOA) showed stronger prediction error-related activity in deeper layers (presumed feedback connections), with a mean activation 1.8 times greater than in superficial layers (95% HDI [1.2, 2.5]). These layered differences were stable across the three video conditions (Bayes factor = 8.7 in favour of the hierarchical model versus a null model with uniform layer effects).

Individual differences in prediction error sensitivity in V1 significantly correlated with performance on a perceptual discrimination task administered after scanning (r = 0.34, 95% HDI [0.19, 0.48]). Exploratory analysis suggested that participants with stronger feedback signals in higher areas showed larger behavioural adaptation effects over the course of the video stimulus. These relationships survived multiple comparison correction using the Benjamini–Hochberg procedure.

Discussion

Our findings provide direct evidence for hierarchical prediction error coding in the human visual system, with layer-specific patterns consistent with predictive coding theory. The superficial-layer concentration of prediction errors in V1 aligns with models proposing that early sensory areas compute mismatch between predictions and input, while deeper-layer feedback activity in higher areas may propagate corrective signals down the hierarchy. This architecture may afford efficient error correction at multiple levels of visual processing.

The correlation between individual prediction error sensitivity and behavioural task performance suggests that the fidelity of predictive processing has functional consequences for perceptual acuity. Future work should examine how this capacity varies across psychiatric conditions and developmental stages. Integrating these findings with computational models of active inference may illuminate the mechanistic basis of perception-action coupling in naturalistic environments.

References

Friston, K. (2010). The free-energy principle: A unified brain theory? Nature Reviews Neuroscience, 11(2), 127–138.
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral and Brain Sciences, 36(3), 181–204.
Kietzmann, T. C., Spoerer, C. J., Sörensen, S. A., Cichy, R. M., Hauk, O., & Kriegeskorte, N. (2019). Recurrence is required to capture hierarchical structure in vision. Nature Neuroscience, 22(10), 1648–1655.
Rao, R. P., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79–87.
Sedley, W., Gander, P. E., Kumar, S., Kovach, C. K., Leszczynski, M., Crone, N. E., & Reichenbach, T. (2016). Neural oscillations and evoked responses reflect the validity of predictions. Cerebral Cortex, 26(1), 308–319.
Barrett, L. F., & Simmons, W. K. (2015). Interoceptive predictions in the brain. Nature Reviews Neuroscience, 16(7), 419–429.
Hoefler, P. A., Murphy, A. P., & Marois, R. (2020). Voluntary attention gates access to conscious visual experience. Proceedings of the National Academy of Sciences, 117(2), 987–996.