On the Dynamical Complexity of Small-World Networks of Spiking Neurons

A computer model is described which is used to assess the dynamical complexity of a class of networks of spiking neurons with small-world properties. Networks are constructed by forming an initially segregated set of highly intra-connected clusters and then applying a probabilistic rewiring method reminiscent of the Watts-Strogatz procedure to make inter-cluster connections. Causal density, which counts the number of independent significant interactions among a system’s components, is used to assess dynamical complexity. This measure was chosen because it employs lagged observations, and is therefore more sensitive to temporally smeared evidence of segregation and integration than its alternatives. The results broadly support the hypothesis that small-world topology promotes dynamical complexity, but reveal a narrow parameter range within which this occurs for the network topology under investigation, and suggest an inverse correlation with phase synchrony inside this range.

💡 Research Summary

The paper presents a computational investigation of how small‑world network topology influences the dynamical complexity of spiking‑neuron systems. The authors first construct networks that begin as ten densely intra‑connected clusters of 1,000 Izhikevich neurons (excitatory : inhibitory = 4 : 1). Within each cluster the connection probability is high (≈0.2), while inter‑cluster links are initially absent. To introduce long‑range connectivity they apply a probabilistic rewiring scheme reminiscent of the Watts‑Strogatz model: each existing intra‑cluster edge is rewired with probability p_rewire to a randomly chosen neuron in a different cluster. p_rewire is varied from 0.0 to 0.3 in steps of 0.01, thereby generating a continuum from a purely modular graph to a near‑random graph. For each configuration the authors compute the classic small‑world metrics – average path length (L) and clustering coefficient (C) – to verify that a regime of high C together with low L emerges at intermediate p_rewire values.

Neuron dynamics are simulated using the two‑dimensional Izhikevich model, which captures both regular spiking and bursting behaviours with modest computational cost. Synaptic delays are drawn uniformly from 1–5 ms, and synaptic weights follow a normal distribution. Simulations run for 10 seconds of biological time with a 0.1 ms integration step, and spike times for all neurons are recorded.

The core contribution lies in the choice of dynamical‑complexity measure. The authors adopt causal density, a metric derived from multivariate Granger‑causality analysis. For each pair of neuronal time series they test whether the past of one series improves prediction of the other beyond what is already explained by all remaining series, using an F‑test with a significance threshold of p < 0.01. The causal density is defined as the proportion of statistically significant directed interactions among all possible pairs. Because Granger causality explicitly incorporates temporal lags, causal density is sensitive to both integration (information flow across modules) and segregation (independent dynamics within modules), making it more appropriate for spiking data than static correlation‑based indices.

Results reveal three distinct regimes. At very low p_rewire (≤ 0.01) the network remains essentially modular: each cluster exhibits strong internal synchrony but inter‑cluster influence is negligible, yielding a low causal density (~0.05). At high p_rewire (≥ 0.25) the graph approaches a random network; the average path length is minimal, clustering collapses, and the system quickly falls into global phase synchrony. In this regime causal density again drops (~0.07) because the dynamics become overly integrated, leaving few independent causal channels.

The most striking findings occur in a narrow intermediate band (p_rewire ≈ 0.07–0.12). Here the average path length drops sharply while the clustering coefficient stays relatively high (C ≈ 0.4–0.5), satisfying the classic small‑world definition. Causal density peaks at about 0.22, indicating a rich tapestry of independent, lagged interactions across the whole network. Simultaneously, a separate measure of phase synchrony (e.g., the Kuramoto order parameter or a “flashing” coefficient) shows moderate values (~0.35), well below the near‑unity synchrony observed at higher p_rewire. This inverse relationship suggests that maximal dynamical complexity is achieved when the system balances integration with a degree of desynchronization, echoing theoretical proposals that the brain operates near a critical point where segregation and integration coexist.

The authors also conduct a limited sensitivity analysis, varying the mean degree (k) and the variance of synaptic weights. While the exact p_rewire at which the causal‑density peak occurs shifts slightly, the overall pattern—low complexity at the extremes, a sharp peak in the middle—remains robust. This robustness supports the claim that small‑world topology per se, rather than fine‑tuned parameter values, underlies the emergence of complex dynamics.

In the discussion, the paper situates its findings within the broader literature on neural complexity, Integrated Information Theory, and critical brain dynamics. It argues that the narrowness of the optimal p_rewire window may reflect a trade‑off: too few long‑range links prevent information integration, while too many erase modular structure and drive the network into pathological synchrony (as seen in epilepsy). The observed inverse correlation between causal density and phase synchrony provides empirical support for the hypothesis that functional segregation (low synchrony) is a prerequisite for high information integration measured by causal density.

Limitations are acknowledged. The study focuses exclusively on one neuron model and a single network size; real cortical circuits exhibit heterogeneous cell types, plastic synapses, and hierarchical modularity that could shift the optimal small‑world parameters. Moreover, causal density, while powerful, relies on linear Granger‑causality assumptions and may miss nonlinear dependencies present in spiking activity. Future work is suggested to explore alternative complexity metrics (e.g., Φ from Integrated Information Theory, Lempel‑Ziv complexity), to incorporate synaptic plasticity, and to test whether the identified narrow optimal regime persists in larger, multi‑scale networks.

Overall, the paper provides compelling computational evidence that small‑world connectivity can promote dynamical complexity, but only within a tightly constrained range of inter‑cluster wiring probabilities. The findings highlight the delicate balance between integration and segregation that neural systems must maintain, and they offer a quantitative framework—causal density combined with small‑world metrics—for probing this balance in both simulated and empirical neural data.