| name | rnn-structural-design-computational-ability |
| description | Paper analysis: Identifying structural design principles shaping computational abilities of recurrent neural networks. Demonstrates that local 2- and 3-cycles in connectivity strongly enhance computational ability of RNNs, and that adding sparse biologically-inspired interneurons dramatically increases capacity. Complete catalogs of network-function performance reveal most networks fail at most functions. Source: arXiv:2606.23874 (q-bio.NC, cs.NE), 2026-06-22.
Activation keywords: RNN structure-function, local cycles, computational ability, connectivity principles, Boolean functions, interneurons, network catalogs, structural statistics, recurrent neural network design, wiring principles, biological connectivity, network architecture, computational capacity, graph theory neuroscience, structure computation, network function
|
| date_added | 2026-06-29T00:00:00.000Z |
| arxiv_id | 2606.23874 |
| authors | ["Tom Talpir","Elad Schneidman"] |
Identifying Structural Design Principles Shaping the Computational Abilities of Recurrent Neural Networks
Paper Metadata
- arXiv ID: 2606.23874
- Published: 2026-06-22
- Categories: q-bio.NC, cs.NE
- Authors: Tom Talpir, Elad Schneidman
Abstract
Understanding how the architecture of neural networks shapes the computations they carry is a central challenge in neuroscience and machine learning. While specific circuit architectures have been linked to particular network computations and theoretical bounds on expressivity of broad classes of networks have been found, general principles connecting the structure of finite networks to their computational capabilities are still missing. The authors characterize the computational abilities of recurrent neural networks as a function of their connectivity by training a large collection of different networks to compute a large set of Boolean functions. For small networks, complete "catalogs" of network-function performance reveal that computational capacity varies widely across architectures and that most networks show poor performance, and most functions are hard to compute. However, having local 2- and 3-cycles in a network strongly enhances its computational ability, and networks with such cycles are often the minimal architectures that can solve particular functions. A small set of structural statistics accurately predicts networks' performance. Extending to large networks shows that typical networks fail to approximate a randomly selected function. Surprisingly, adding a small number of sparsely connected biologically-inspired interneurons dramatically increases computational capacity. Adding short cycles improved networks' capacity, outperforming acyclic or reachability-matched controls.
Methodology
Exhaustive Structure-Function Mapping
- Network generation: Large collection of RNNs with varied connectivity patterns
- Function library: Large set of Boolean functions to compute
- Complete catalogs: For small networks, all network-function pairs tested
- Training protocol: Networks trained on each function, performance recorded
Structural Analysis
- Cycle analysis: Systematic examination of local 2-cycles and 3-cycles
- Structural statistics: Small set of graph-theoretic measures
- Null models: Acyclic controls and reachability-matched controls
- Interneuron addition: Biologically-inspired sparse inhibitory neurons
Scale Extension
- Small networks: Complete catalog construction
- Large networks: Statistical analysis of capacity
- Interneuron experiments: Adding biologically-motivated inhibitory cells
Key Findings
- Local cycles are key: 2-cycles (reciprocal connections) and 3-cycles strongly enhance computational ability
- Most networks fail: The majority of random architectures perform poorly on most functions
- Most functions are hard: Boolean functions vary widely in learnability
- Minimal architectures: Networks with short cycles are often the smallest architectures capable of specific computations
- Structural predictability: A small set of structural statistics accurately predicts network performance
- Interneuron boost: Adding sparse biologically-inspired interneurons dramatically increases computational capacity
- Scale invariance: Short cycles improve capacity even in large networks, outperforming controls
- Typical networks fail: Large typical networks cannot approximate randomly selected functions
Implications
For Neuroscience
- Provides quantitative link between microscale connectivity and computational power
- Explains why biological circuits have abundant local recurrence and interneurons
- Suggests structural motifs (cycles) are evolutionary design principles
- Supports the importance of inhibitory interneurons beyond simple balance
For Machine Learning
- Architecture design should prioritize local recurrent motifs
- Random connectivity is computationally wasteful
- Interneuron-inspired architectures could improve RNN performance
- Structural statistics can predict computational capacity without training
For Network Theory
- Extends graph theory to computational function space
- Establishes that local topology (not just global) determines computation
- Provides framework for "computational topology" of networks
Critical Analysis
Strengths
- Exhaustive approach: complete catalogs, not sampling
- Bridges neuroscience and ML perspectives
- Identifies simple, interpretable structural principles
- Validates with both small and large networks
- Biologically motivated interneuron experiments
Limitations
- Boolean functions may not generalize to continuous computations
- Small networks may not capture large-scale dynamics
- Training convergence issues could confound capacity measurements
- Specific to feedforward-readout paradigm
Connections
- [[rnn-structural-design-computational-ability]] - this paper
- [[cortical-geometry-rnn-inductive-biases]] - related work on RNN inductive biases from cortical geometry
- [[cortical-microcircuit-information-flux]] - cortical microcircuit computation
- [[connectome-wiring-statistical-dynamics-separation]] - connectome wiring statistics
- [[chaos-freezing-without-plasticity]] - RNN dynamics and computation
