Exact N:4 pruning and recurrent dependence: a retrospective case study in an adaptive spiking network

Structured pruning in recurrent spiking neural networks reduces connectivity,but neither retained weight count nor similar classification accuracy establishes howthe recovered network uses spikes and recurrent state. We retrospectively examineexact N:4 pruning, which retains N entries per four-source group, within a fixedblock-sparse adaptive network on Spiking Heidelberg Digits. Retained per-exampleoutputs from two training seeds support equal-budget continuation comparisonsand temporal interventions. Magnitude and random 1:4 masks both remove 75%of recurrent entries, yet logical recurrent synaptic operations (SynOps) decrease by72.22% and 75.25%, respectively, relative to each seed’s matched 4:4 continuation.Neuron-spike reductions are only 1.68% and 0.63%. The random 1:4 checkpointshave 82.99% mean full-input accuracy versus 82.64% for continuation. Removingrecurrent drive costs them 0.73 percentage points versus 5.74 for continuation, withreduced sensitivity in both seed pairs. Corrected midpoint reset still costs 8.97points, and temporal shuffle costs 66.76 points. Thus this sparse classifier remainssensitive to temporal input and neuronal-state disruption while showing a muchsmaller net accuracy loss under recurrent-drive removal. We interpret this jointprofile using source-spike/fan-out accounting and absolute intervention losses. Thecheckpoints share a random ranking, and the highlighted 1:4 arm was added posthoc and selected on the reporting test set; comparisons are descriptive, not evidence of independent accuracy preservation. The case study illustrates the value ofmeasuring both spike-driven work and recurrent dependence, beyond density andrelative diagnostic tolerances. Hardware acceleration, energy savings and a generalpruning advantage are not established.This record contains a research preprint that has not undergone journal peer review. Supporting experimental data and code may be requested from Sangbum Kim at [email protected], subject to applicable component permissions.