Technical Summary: Unifying Dynamical Systems and Graph Theory to Mechanistically Understand Computation in Neural Networks

Problem Statement

A central challenge in neuroscience and machine learning is understanding how computation emerges from the connectivity of neural networks. While structural connectivity (synaptic or weighted connections) is often assumed to constrain function, empirical evidence shows that structural and functional connectivity frequently diverge. Standard approaches that analyze networks based solely on direct, single-step connections (one-hop interactions) fail to capture the temporal evolution of information in recurrent neural networks (RNNs). In RNNs, computation is mediated by multi-hop pathways that unfold over time. Consequently, traditional methods for analyzing network structure (e.g., examining weight matrices directly) or regularizing them (e.g., L1 penalties on individual weights) may not align with the actual functional mechanisms of the network. The paper posits that to truly understand and control computation, one must move beyond single-step interactions to analyze the aggregate signal propagation across multiple routes.

Methodology

Theoretical Framework

The authors model RNNs as directed graphs where weight matrices serve as adjacency matrices. They utilize graph-theoretic measures to aggregate signal propagation across different path lengths (hops):

• Graph Resolvent: To capture the cumulative effect of multi-hop pathways, the authors employ the resolvent, defined as $R = (I - \gamma A)^{-1} = \sum_{k=0}^{\infty} (\gamma A)^k$, where $A$ is the adjacency matrix and $\gamma < 1/\lambda_{max}$ ensures convergence. This corresponds to a "leaky cascade" dynamic.
• Truncated Resolvent: Since RNNs operate over a finite sequence length $L$, the authors use a truncated resolvent to avoid attributing functional influence to walks longer than the task horizon. The input-to-output influence map, $R_{io}$, is extracted to represent the signed weighted sum of all walks between input and output nodes.
• Hop-wise Decomposition: The authors decompose the resolvent into individual hop lengths ($W^k$) to analyze how information is routed temporally.

Experimental Setup

• Tasks: RNNs were trained on a family of modular temporal integration tasks requiring specific input-output routing structures:
  1. Module Averaging: Aggregating noisy features within modules.
  2. Hierarchical Subtraction: Recovering module means via successive subtraction.
  3. Hierarchical Addition: Summing module means.
  4. Multiplication: Non-linear combination of module means (defined via task Jacobian).
  5. Oscillating On-Off Signal: A task probing temporal routing where signals and noise alternate in time.
• Architecture: 1-layer RNNs with input, hidden, and output dimensions set to the number of features ($i=h=o=16$).
• Regularization Comparison:
  • L1-RNNs: Standard regularization penalizing the magnitude of individual weights ($L_{sparsity} = \sum |W|$).
  • Resolvent-RNNs (R-RNNs): A novel regularization approach penalizing the magnitude of the multi-hop input-output influence map ($L_{sparsity} = \sum |R_{io}|$). This formulation ensures that the network is penalized for the complexity of its communication pathways rather than just individual connection strengths.

Key Results

1. Recovery of Input-Output Structure via Multi-Hop Measures

The study demonstrates that raw weight matrices ($W_{hh}$) fail to recover the learned input-output structure, even when the network has sufficient capacity. In contrast, the resolvent-based measure ($R_{io}$) successfully reconstructs the expected modular routing structures for averaging, subtraction, addition, and multiplication tasks.

• Correlation: $R_{io}$ showed high correlation with the optimal analytical solutions (e.g., Pearson correlation $\approx 0.99$ for averaging and subtraction), whereas the raw hidden weights ($W_{hh}$) showed negligible correlation (e.g., $\approx 0.24$).
• Mechanism: This indicates that the functional logic of the network is encoded in the aggregate multi-hop pathways rather than the static, single-step weights.

2. Temporal Routing Revealed by Hop Decomposition

By decomposing the resolvent by hop length ($k$), the authors revealed how the network routes information temporally.

• In the oscillating on-off signal task, even-numbered hops ($k=2, 4, 6$) corresponded to signal-carrying time steps, reconstructing the modular structure.
• Odd-numbered hops corresponded to noise steps and showed no task-relevant structure.
• This confirms that hop length directly encodes the temporal delay of information propagation, a feature invisible to standard weight analysis.

3. Performance and Robustness of R-RNNs

When comparing regularization strategies:

• Sparsity Alignment: R-RNNs achieved lower test Mean Squared Error (MSE) than L1-RNNs. While L1 regularization reduces weight magnitude, it does not necessarily reduce the complexity of multi-hop pathways. R-RNNs, by penalizing $R_{io}$, induced temporally structured sparsity that matched the task requirements (e.g., suppressing redundant pathways across time).
• Robustness: Under strong regularization (high $\beta$), L1-RNNs suffered rapid performance degradation. R-RNNs maintained low MSE even at high levels of resolvent sparsity, indicating that they preserve task-relevant computation while pruning non-functional communication pathways.
• Magnitude vs. Structure: R-RNNs often exhibited larger raw weight magnitudes than L1-RNNs but significantly lower resolvent magnitudes, proving that the resolvent captures structural differences in communication that L1 penalties miss.

Significance and Claims

The paper claims to provide a unified framework linking structure, dynamics, and function in neural networks through the lens of graph theory and dynamical systems.

1. Reframing Computation: The authors argue that computation in recurrent networks is better understood as a process of multi-hop communication rather than a collection of individual connections. This perspective resolves the divergence between structural and functional connectivity by showing that function emerges from the accumulation of walks.
2. New Regularization Paradigm: The introduction of R-RNNs suggests that sparsity should be defined over functional pathways rather than individual parameters. This approach yields networks that are more robust and better aligned with task structures than those trained with standard L1 regularization.
3. Interpretability: The resolvent and hop-wise decomposition offer an input-agnostic method to characterize a network's "routing capacity." This provides a vocabulary for describing network computation based on how information moves through the system, independent of specific activation patterns or input stimuli.
4. Biological and Artificial Relevance: The findings suggest that biological networks may optimize for efficient communication pathways rather than minimal synaptic strength. The framework offers a potential bridge for understanding how structural motifs (modularity, hierarchy) translate into functional modularity in both biological and artificial systems.

The authors conclude that their results identify multi-hop communication as a key principle for linking structure to function, proposing that future analyses and regularizations of neural networks should account for these temporal and spatial pathways.