Clustering and Degree Homogeneity Govern Reservoir Computing Performance

Technical Summary: Clustering and Degree Homogeneity Govern Reservoir Computing Performance

Problem Statement
Reservoir computing (RC) utilizes fixed, recurrent neural networks as computational substrates, where only the linear readout layer is trained. While biological brains exhibit sparse, clustered, and small-world connectivity, it remains an open question whether connectome-derived reservoirs inherently possess superior computational capabilities compared to synthetic architectures. Previous studies suggested that connectome-based networks outperform null models (scrambled connectomes), but these comparisons often failed to control for fundamental network statistics such as size, density, and degree distribution. Furthermore, the specific topological properties that govern RC performance—specifically the interplay between clustering, degree heterogeneity, and path length—have not been systematically disentangled against a controlled synthetic baseline. This study addresses whether the biological provenance of a network (the human structural connectome) confers a computational advantage over synthetic small-world and random networks when matched for size and mean degree.

Methodology
The authors benchmarked three families of reservoir architectures on two standard tasks: six-class human-activity classification and one-step prediction of the Mackey–Glass chaotic time series.

1. Network Architectures:

   • Connectome: Derived from a human structural connectome (66 cortical regions) based on diffusion spectrum imaging. To compare across varying sizes ($N=66$ to $N=400$), the authors employed an expansion and sparsification procedure (inspired by the bio2art approach) to map the coarse-grained connectome to a neuron-level network while preserving modular structure and hub dominance.
   • Small-World: Synthesized using the Watts–Strogatz model, starting from a ring lattice and rewiring edges with probability $p$.
   • Random: Generated by placing connections uniformly at random among node pairs.
   • Controls: All reservoirs were binary, symmetric, and matched for the number of nodes ($N$) and mean degree ($\langle k \rangle = 8$). This ensured performance differences reflected connection arrangement rather than connection count.

2. Experimental Protocol:

   • Tasks: Human-activity classification (UCI HAR dataset) and Mackey–Glass prediction ($\tau=17$).
   • Parameters: The recurrent weights were fixed; only the linear readout was trained via ridge regression. Spectral radius and leaking rates were fixed per task (classification: $\rho=0.9, a=0.2$; prediction: $\rho=1.2, a=0.6$).
   • Analysis: For the small-world family, the rewiring probability $p$ was swept across $\{0.01, 0.1, 0.3, 0.5, 0.7\}$. Performance was correlated with network metrics: mean clustering coefficient ($C$), degree heterogeneity (standard deviation of degree distribution), and characteristic path length ($L$).

Key Results

1. Performance Hierarchy: On both classification and prediction tasks, the Watts–Strogatz small-world reservoirs consistently outperformed both the human connectome and random networks.

   • Classification (N=400): Small-world achieved 69.24% accuracy, exceeding the connectome (67.29%) and random (66.28%).
   • Prediction (N=400): Small-world achieved a mean-squared error (MSE) of $4.96 \times 10^{-8}$, roughly 3.2-fold lower than both the connectome ($1.57 \times 10^{-7}$) and random networks ($1.64 \times 10^{-7}$).
   • Notably, the connectome performed statistically indistinguishably from the random network on the prediction task, offering no advantage over random wiring.

2. Optimal Rewiring: Performance peaked at a rewiring probability of $p = 0.1$ for both tasks. The small-world advantage persisted across all tested $p$ values, including the least favorable case ($p=0.7$), indicating the result is not an artifact of a specific parameter tuning.

3. Governing Structural Properties:

   • Clustering and Homogeneity: Performance strongly correlated with high clustering and low degree heterogeneity (Pearson correlations between 0.88 and 0.94).
   • Path Length: Performance showed no significant correlation with characteristic path length.
   • Connectome Disadvantage: The human connectome is characterized by low clustering (an order of magnitude below the small-world optimum) and high degree heterogeneity (presence of hubs). These properties place the connectome away from the performance optimum occupied by the small-world networks.

Significance and Claims
The paper concludes that for the tasks examined, the computational capability of a reservoir is governed principally by generic network statistics—specifically clustering and degree homogeneity—rather than by biological provenance.

• Mechanism: The authors suggest that degree heterogeneity (hubs) penalizes performance because hubs concentrate dynamics on a few nodes, rendering the remaining nodes redundant (scaled/delayed copies of hubs) and reducing the effective dimensionality of the reservoir states. Conversely, high clustering sustains local recurrent loops, providing the readout with a longer temporal history.
• Design Implication: For reservoir design, these findings indicate that synthetic generators can directly and cheaply control the relevant properties (clustering and homogeneity) to achieve optimal performance, without needing to replicate biological connectomes.
• Neuroscience Context: The study clarifies that while connectome topology may be advantageous against null models that preserve degree distribution (as shown in prior work), its advantage is offset when compared against degree-homogeneous architectures. The authors modestly note that connectome-derived connectivity might still be advantageous for tasks requiring specific biological features (e.g., integrating disparate input streams or multi-tasking), but it is not inherently superior for the standard memory and classification benchmarks tested here.

Limitations
The authors acknowledge that the study is limited to a single connectome parcellation (66 regions), binary networks (discarding connection strengths), and correlational analysis where clustering and heterogeneity co-vary. They also note that the spectral radius and leaking rates were fixed rather than tuned per architecture, though sensitivity checks suggested the chosen values were near-optimal.