Spiking ResNets' functional ensembles encode info in rare bursts

A new paper on arXiv proposes a neuroscience-inspired framework to analyze deep spiking neural networks (SNNs) through functional connectivity. The authors define first-order functionally-connected (1FC) groups for each neuron based on statistically significant pairwise correlations with neurons in the previous layer. Using spiking ResNet architectures, they find that these 1FC ensembles exhibit properties reminiscent of biological cortex: rare but highly coordinated cofiring events reliably encode the presented class, while average activity does not. The aggregate cofiring of an ensemble predicts downstream neuronal responses via a robust, ReLU-like input-output relationship, with gain scaling systematically with ensemble size.

Under uniform random noise or adversarial perturbations, these response profiles are disrupted, especially in early and intermediate layers, enabling targeted interrogation of specific nodes and pathways. The researchers also show that the functional connectivity structure is shaped by learning and breaks under weight permutation, confirming that 1FC ensembles are a functionally meaningful substrate for input encoding and information transfer. This work has potential implications for designing fine-grained diagnostics on information flow in SNNs and for more efficient neuromorphic computing by focusing on rare, informative events.