Predictive Error Topography in Sensory Binding: A Biologically Plausible Model of Constrained Attentional Allocation

Introduction

The formation of a unified, moment-to-moment experienced reality relies on the brain’s ability to bind distributed features into a cohesive percept. While classical theories often treat binding as a feature detection problem solved through coincidence detection, we posit a more active, predictive framework. Sensory information arrives asynchronously, yet the conscious experience presents a singular, unified object or event. Our central hypothesis is that this unification is achieved not by forcing disparate signals to match, but by predicting the expected presence and temporal relationship of signals from other modalities based on the most dominant, attended stream. Mismatches between prediction and arrival constitute the measurable binding error.

This work models this process using a constrained recurrent neural network architecture where activity patterns are interpreted as competing hypotheses about the sensory manifold. The stability of the resulting integrated percept hinges on the rapid suppression of prediction errors that fall below a dynamically determined threshold.

Theoretical Framework: Predictive Coding and Binding Error

We extend the Free Energy Principle by modeling the binding process as minimizing the variational free energy associated with the hypothesis that the current sensory input is predictable given past context. When visual input confirms the expected location of an auditory tone, the prediction error ($\epsilon$) is small. Conversely, if the visual field is momentarily ambiguous, but the auditory signal is sharp, the system must commit to a prediction, constrained by the temporal availability of the visual correlate. The mismatch signal, $\epsilon_{i}$, quantifies the discrepancy between the top-down predicted state ($\mu_{pred}$) and the bottom-up sensory evidence ($s_{in}$).

$$\epsilon_{i} = ||s_{in} - \mu_{pred}||^2_Q$$

Where $Q$ is the inverse covariance matrix representing the precision weighting applied to the incoming sensory evidence, modulated by sustained attention.

Attentional allocation, in this model, is equivalent to modulating the precision $Q$. High precision locks the system into a tight hypothesis space, requiring stronger evidence to induce a shift. Conversely, low precision allows for rapid hypothesis revision, making the system susceptible to noise but flexible to novel inputs. The binding coherence emerges when the network successfully compresses the high-dimensional sensory manifold into a low-dimensional, causally constrained percept.

Simulation Protocol and Architectural Mapping

Our simulation employed a tripartite network structure: Visual ($\mathcal{V}$), Auditory ($\mathcal{A}$), and Somatosensory ($\mathcal{S}$). Each modality projected onto a shared latent binding manifold ($\mathcal{B}$). The prediction mechanism flowed from $\mathcal{B}(t-1)$ to predict the expected activation pattern in $\mathcal{V}(t)$, $\mathcal{A}(t)$, and $\mathcal{S}(t)$.

Table 1. Initial Condition Parameterization

To induce a binding failure, we systematically degraded the temporal alignment between $\mathcal{V}$ and $\mathcal{A}$ while maintaining high internal consistency within each modality’s stream. The resulting binding error trajectory within $\mathcal{B}$ served as our measurable dependent variable, allowing us to quantify the timescale over which coherence collapses or maintains itself under incongruent forcing.

Results: Topography of Binding Breakdown

When attention was fixed on the visual stream (high $Q_V$), the system exhibited remarkable resilience to transient misalignment between auditory and somatosensory streams, provided the misalignment remained within the probabilistic envelope defined by $\sigma_a$ and $\sigma_s$. The mismatch signals ($\epsilon_{A \rightarrow V}$ and $\epsilon_{S \rightarrow V}$) were successfully suppressed below the threshold required to elicit a conscious shift in the dominant percept. This indicates that the attended modality bootstrapped the expectation for the others.

However, when the prediction error exceeded a critical threshold, $\epsilon_{crit}$, the network exhibited a rapid cascade failure, characterized by a characteristic power-law decay in the mutual information between the modalities, indicating a sudden, catastrophic decorrelation rather than a gradual integration failure. The relationship can be approximated during the transition phase by:

$$\frac{dI(\mathcal{V}; \mathcal{A})}{dt} \propto -\alpha \cdot \epsilon_{crit}^{-1} \cdot e^{-t/\tau_{decay}}$$

This decay timescale, $\tau_{decay}$, serves as a reliable metric for the system’s capacity to sustain coherent representation against predictable noise.

Discussion and Conclusion

The successful maintenance of a unified percept under noisy, asynchronous input strongly suggests that conscious binding is not a passive averaging process but an active, hypothesis-testing endeavor. The brain commits to a predictive model and uses attentional resources to aggressively prune evidence that contradicts this model, only permitting significant updating when the contradiction is overwhelming or novel. The topographic mapping of binding error provides a tangible, quantitative link between the high-level cognitive act of ‘coherence’ and the low-level physics of constrained information flow within a neural architecture. Future work will explore how neuromodulatory signals modulate the precision weighting $Q$, effectively controlling the decisional threshold between integration and outright contradiction.