Title: Rare Events, Real Signals: Functional Ensembles as Units of Computation in Deep Spiking Networks

Abstract:

This study explores the emergence of internal representations within hierarchical processing systems by applying a neuroscience-inspired framework to deep spiking neural networks (SNNs), viewed through the perspective of functional connectivity. Leveraging principles from information theory and systems neuroscience, we define a first-order functionally-connected (1FC) group for each neuron. This group is constituted by neurons from the preceding layer that exhibit statistically significant pairwise correlations with the target neuron in a trained SNN architecture. We subsequently monitor the response characteristics of these groups during inference across diverse conditions.

Our findings reveal that several functional connectivity principles documented in the biological cortex are retained within spiking ResNet structures. The 1FC ensembles exhibit distinct characteristics: their collective cofiring events reliably forecast downstream neuronal activity via a robust, ReLU-like input-output mapping. The gain of this relationship scales systematically with the size of the ensemble. Notably, the faithful encoding of the target class manifests exclusively during periods of high 1FC cofiring. Since these coordinated events are infrequent, the results suggest that informative representations are concentrated within rare but highly synchronized activity patterns.

These response profiles are significantly disrupted when subjected to uniform random noise or adversarial perturbations, a vulnerability that is particularly pronounced in the early and intermediate layers of the network. This sensitivity allows for targeted, high-resolution interrogation of specific nodes and pathways. Furthermore, we demonstrate that the functional connectivity structure is a product of the learning process; this structure disintegrates when weights are permuted. Consequently, these 1FC ensembles serve as a functionally significant substrate for input encoding and information transfer, offering potential implications for the development of targeted, fine-grained diagnostics to analyze information flow.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC