Are cortical microcircuits optimized for information flux? -- A simulation-based reverse engineering study

Abstract: A sufficiently large information flux in recurrent neural networks, quantified by the mutual information between successive network states, is considered a prerequisite for rich information processing capabilities. This raises the question of whether biological neural networks, such as cortical microcolumns, may be structurally organized to enhance information flux. To investigate this possibility, we study a simplified model of the cortical layer 5 architecture, in which a densely and strongly interconnected core population is embedded within a larger supporting network. Surprisingly, we find that the embedding network exerts a pronounced flux-enhancing effect on the core dynamics. Systematic reverse-engineering analyses reveal that the embedding network provides two key contributions: first, it generates effective biases that shift core neurons into a higher-entropy operating regime; second, it supplies stochastic fluctuations that prevent the network from becoming trapped in simple fixed-point or oscillatory attractors through the mechanism of Recurrence Resonance. We further show that the information flux can be increased even beyond the biologically embedded case by applying individually optimized biases to the core neurons, and that these biases can emerge from a simple self-organization principle. Our findings are relevant both for the functional interpretation of biological neural circuits and for the design of artificial recurrent systems such as reservoir computers.

• The paper demonstrates that embedding cortical cores within a layer-5-like circuit boosts information flux by approximately 6× through effective bias, especially via inhibitory inputs.
• It employs a simulation-based reverse engineering approach using an Embedded Core Model with binary Boltzmann neurons to replicate key statistical features of cortical wiring.
• The study shows that optimized, neuron-specific bias adjustments enabled by homeostatic plasticity can elevate information flux beyond that achieved by network embedding alone.

Information Flux Optimization in Cortical Microcircuits: A Simulation-Based Reverse Engineering Analysis

Introduction

The question of whether biological cortical microcircuits are structurally optimized for maximal information flux has deep implications for both neuroscience and the design of artificial recurrent neural systems. This work investigates this hypothesis by means of simulation-based reverse engineering applied to a biophysically motivated abstract model of cortical layer 5 microcircuits. The authors develop an Embedded Core Model (ECM) incorporating a densely connected excitatory core, a sparsely wired excitatory periphery, and an inhibitory interneuron population, capturing key experimentally reported statistics of cortical wiring.

Information flux—quantified as the mutual information between consecutive network states—is employed as the central dynamical metric. Systematic interrogation of the ECM reveals nontrivial mechanisms underlying the regulation of information processing capacity, implicating both static and dynamic modulatory influences from the embedding network. The findings have broad consequences for interpreting the functional architecture of cortical columns and for guiding bias and noise modulation strategies in RNN engineering.

Model Architecture and Methodology

The ECM is structured with $N_t = 125$ binary Boltzmann neurons partitioned into three populations: 10 core excitatory units with strong and dense interconnections, 90 peripheral excitatory units wired sparsely, and 25 inhibitory interneurons. Connectivity reflects a biologically informed log-normal weight distribution observed in cortical circuits ([23]). Boltzmann unit dynamics facilitate tractable analysis and enable exact control of firing statistics through temperature and bias parameters.

Information flux is evaluated via intra- and inter-triplet mutual information, leveraging subnetwork statistics to approximate the high-dimensional core’s global flux. This method balances computational feasibility with the necessity of capturing essential dynamical dependencies.

Major Findings

Embedding Network’s Role in Enhancing Core Information Flux

A salient result is that embedding the core within the full layer-5-like circuit increases the core’s information flux by a factor of $\sim6\times$ (flux indicator $f=139.3$ in the embedded model vs. $f=23$ in isolation). This enhancement does not require closed-loop feedback from core to periphery; rather, the effect is strictly feedforward. Lesion experiments show that the main contribution arises from inhibitory interneuron input to the core. The peripheral (excitatory) population also contributes, but to a lesser extent.

Biases and Fluctuations: Distinct and Synergistic Effects

Statistical analysis of embedding network inputs shows that mean shifts (biases) from peripheral and interneuron populations move core neurons into higher-entropy regimes, while temporal fluctuations serve a noise-like role. Setting embedding network inputs to their temporal means retains the majority of the flux enhancement ($f=109.5$), confirming the dominant influence of effective bias. Temporal fluctuations supply an additional, but smaller, benefit. Importantly, substituting real embedding fluctuations with independent Gaussian noise of identical variance fails to recover full flux enhancement, indicating the criticality of the specific (non-Gaussian, structured) statistics present in the real connectivity.

Analytical Single-Neuron Theory

The work uses a single noisy, self-coupled Boltzmann neuron to dissect mechanisms. Analytical calculations demonstrate a resonance in flux as a function of noise amplitude (Recurrence Resonance), confirming prior findings ([19], [39]). However, optimal, properly tuned constant biases can enhance information flux far more effectively than noise, especially in systems with strong positive feedback. This insight generalizes to multineuron systems and is supported by numerical experiments.

Bias Optimization and Local Adaptation

By sweeping and evolutionarily optimizing neuron-specific bias vectors in the isolated core, information flux can exceed that achieved with embedding network input ($f=203.9$ vs. $f=139.3$). Importantly, local homeostatic plasticity—implemented as adaptive feedback driving the time-averaged neuron output to $1/2$—leads to identical optimal bias configurations, providing a plausible biological mechanism for information flux maximization.

Theoretical and Practical Implications

Computational Neuroscience

The ECM shows that cortical core-periphery architecture and structured inhibitory/excitatory input can robustly place core circuits into regimes of high entropy, optimizing mutual information flow without the necessity for closed feedback loops. This reconceptualizes the role of peripheral and interneuron populations as modulatory agents establishing computationally privileged dynamical states, as opposed to mere conveyors of content signals.

The work draws explicit connections to the "edge of chaos" regime in RNNs ([40], [41]), demonstrating that maximum core neuron entropy—obtainable via locally tuned biases—serves as a principled route to this high-capacity operating point. The bias-targeting mechanism aligns with homeostatic plasticity and BCM-like regulatory theories ([47], [48]), but with precise information-theoretic justification for target firing rates.

Reservoir Computing and Artificial RNN Design

The results are highly relevant to the engineering of reservoir architectures and other RNN implementations. Traditional gain/noise tuning (e.g., adjusting spectral radius or adding Gaussian noise) is less effective than precise, neuron-specific bias control for maximizing information flux. The demonstrated efficacy of homeostatic, local bias adaptation provides a blueprint for robust, self-organizing RNN modules. Moreover, reliance on non-Gaussian, structured perturbations instead of white noise suggests new avenues for noise engineering in artificial systems.

Core-Periphery Motifs in Complex Networks

The identified function for the core-periphery motif—in maximizing core information flux—offers a functional interpretation of architectural features found across complex networks, including the "rich club" organization of large-scale brain connectomes ([50], [51]).

Limitations and Open Questions

• The hypothesis that large information flux always correlates with optimal computational performance requires broader scrutiny. Tasks with divergent statistical or temporal complexity may exhibit distinct optimal operating points.
• The weak contribution of temporal fluctuations from the embedding network, except when non-Gaussian, motivates a systematic analysis of how the structure and statistics of such fluctuations interact with circuit topology to shape information dynamics.
• The potential for active, content-dependent computation in the peripheral network, as opposed to mere modulation of the core, remains unexplored.
• The generalization of these findings from cortical microcircuit scale to mesoscale or macroscale brain networks is an open and promising direction.

Conclusion

This study provides rigorous simulation and analytical evidence that the structural organization of cortical-like microcircuits can substantially enhance information flux within strongly interconnected cores, primarily through static effective biases supplied by surrounding neural populations and secondarily via selectively structured fluctuations. The results suggest that layer 5 microcircuits are, at least in part, organized to optimize information transfer, not by maximizing randomness but by tuning cores to high-entropy, deterministic regimes accessible with homeostatic, local regulation. These findings bear directly on both the interpretation of biological recurrent networks and the principled engineering of artificial systems, motivating further inquiry into the interplay between structure, intrinsic dynamics, and optimal information processing.

Reference:

"Are cortical microcircuits optimized for information flux? -- A simulation-based reverse engineering study" (2605.14680)