Technical Summary: Functional Equivalence Without Structural Convergence

Problem Statement

The study investigates a fundamental question in the optimization of neural systems: when independent optimization processes with different representations and operators are tasked with maximizing the same functional of network structure, do they converge on a single architecture, or do they discover a diverse set of structurally distinct solutions? This inquiry addresses the concept of degeneracy—the capacity of structurally different elements to yield the same function—a property observed in biological neural circuits (Prinz et al. 2004; Edelman and Gally 2001). The paper specifically asks whether high-performing networks found via different evolutionary routes share a common structural signature or if the "functional equivalence set" is populated by architecturally diverse networks.

Methodology

Experimental Setup

The study employs a controlled Monte Carlo simulation on symmetric weighted networks with node counts $N \in \{8, 12, 16\}$. Edge weights are constrained to $[-1, 1]$ and rescaled to a unit spectral radius. Three distinct dynamical regimes are tested:

1. Linear Equilibrium: Instantaneous response to Gaussian noise.
2. Linear AR(1): Recurrent process with stable spectral radius.
3. Nonlinear Rate Network: A recurrent network with tanh activation and additive noise.

Search Routes

Four search strategies were compared under equal evaluation budgets (2,000 evaluations for linear models; 1,000 for nonlinear):

• DNA-encoded GA: A genetic algorithm using a discrete 4-letter genome (A, C, G, T) mapped to edge weights via a positional code.
• Continuous GA: A standard genetic algorithm with continuous genes and Gaussian mutation.
• Evolution Strategy (ES): A $(1+10)$ ES with multiplicative step-size adaptation.
• Random Search (RND): The best result from a budgeted random draw, serving as a baseline.

Functionals (Objectives)

Three information-theoretic functionals, serving as proxies for neural dynamics richness, were optimized:

• Neural Complexity ($C_N$): A measure balancing integration and segregation based on subset mutual information.
• Phi-like Measure (PHI): A normalized minimum-information-partition metric (Gaussian mutual information over bipartitions).
• Synergy Dominance (SYN): The negative of Gaussian O-information, measuring synergy versus redundancy.

Structural Descriptors and Metrics

To assess structural convergence, five permutation-invariant descriptors were computed for optimized networks:

1. Sorted spectrum of the weight matrix.
2. Signed triad census.
3. Integration–segregation profile.
4. Modularity ($Q$).
5. Distribution of linear impulse-response magnitudes.

Convergence Ratio ($D$): Structural similarity was quantified by the ratio of the mean cross-route distance to the mean distance between random networks ($D < 1$ implies convergence closer than random; $D \approx 1$ implies no convergence). Additionally, nonlinear perturbation responses were simulated to test functional equivalence beyond linear proxies.

Key Results

1. Functional Performance

All optimization routes significantly outperformed random search across all functionals and dynamics.

• ES Superiority: The Evolution Strategy consistently achieved the highest scores (e.g., normalized $C_N$ of 0.86 vs. 0.46 for random search in the baseline).
• Encoding vs. Algorithm: The DNA-encoded GA and Continuous GA were statistically indistinguishable for $C_N$ and PHI, suggesting the encoding itself did not drive the performance gap. However, ES significantly outperformed the Continuous GA, indicating that optimization operators (selection, step-size adaptation) contributed to the performance difference.

2. Structural Convergence

No Universal Convergence: Independent routes did not converge on a single architecture.

• Spectral Descriptors: The spectral convergence ratio $D$ was generally $>1$ (e.g., 1.19 for $C_N$), indicating that optimized networks from different routes were often more dissimilar than random pairs.
• Descriptor Dependence: Apparent convergence was highly dependent on the descriptor. While the integration–segregation profile showed convergence ($D \approx 0.52–0.68$), this was partly circular as it is derived from the objective function itself.
• Dynamics Dependence: Convergence was not robust to dynamics. In the nonlinear model, spectral convergence for $C_N$ showed modest convergence ($D=0.85$), but this did not hold for other functionals or dynamics (e.g., AR(1) showed $D=1.34$).

3. Functional Equivalence and Diversity

• Diverse High-Scoring Sets: The set of networks achieving high scores (normalized $\ge 0.8$) was architecturally diverse. Their pairwise spectral distances were only modestly closer to each other than random networks (e.g., $0.79$ times the random distance at $N=8$).
• Scaling: As network size increased ($N=8 \to 16$), the diversity of high-scoring sets increased slightly, though the data did not definitively establish that degeneracy scales with size.
• Perturbation Response: Nonlinear perturbation responses showed at most modest convergence ($D \approx 0.90$ for $C_N$), confirming that structurally distant networks can exhibit similar functional responses, but the evidence was mixed and not a "clean collapse" onto one profile.

4. Robustness and Repair

• Fragility: Optimized networks were brittle. Deleting 10% of edges reduced normalized scores from $\ge 0.8$ to $\approx 0.34$.
• Partial Recovery: Repair via evolution restored function partially (to $\approx 0.64$) but did not restore the original architecture. Repaired networks remained structurally distinct from the original, demonstrating that function can be recovered via different structural configurations.

Significance and Claims

The paper concludes that functional equivalence without structural convergence is a robust finding within the tested models. The primary significance lies in the dissociation between high functional performance and structural uniformity:

1. Degeneracy in Optimization: The functional landscape for these information-theoretic measures contains many distinct, near-optimal solutions. Independent optimization routes do not necessarily converge on a single "privileged" architecture.
2. Descriptor Sensitivity: The apparent convergence of networks depends heavily on the choice of structural descriptor and the underlying dynamics. No single descriptor provided a robust, universal signal of convergence.
3. Optimization Operators: The performance gap between DNA-encoded and continuous GAs was not statistically significant, but the superiority of ES suggests that optimization operators (rather than just encoding) drive performance gains.
4. Limits of Proxies: The study explicitly states that these results concern proxy functionals (neural complexity, Phi-like measures) and do not make claims about consciousness or biological systems. The "Phi-like" measure is not Integrated Information Theory's $\Phi$, and the models are simplified (symmetric weights, small $N$).

The author emphasizes that while the results align with biological observations of degeneracy, the study is exploratory. The high-scoring networks were brittle, suggesting that selection on a single functional without robustness pressure yields fragile solutions. The findings imply that theories linking consciousness to a specific network structure may need to account for the possibility of multiple, structurally distinct architectures achieving similar functional values.