CQP Pruning cuts SNN energy 73% with 95.6% accuracy at 90% sparsity

Deploying spiking neural networks (SNNs) on energy-constrained neuromorphic hardware requires aggressive pruning without destroying temporal computation. Existing methods either ignore neuronal criticality or rely on convex relaxations that produce fractional masks—and when binarized, those masks can cause accuracy to plummet by as much as 44 percentage points. Muhammad Hamza’s new paper introduces Criticality-Constrained Quadratic Pruning (CQP), a native PyTorch pipeline that solves these problems with an analytically exact importance metric that combines weight magnitude and surrogate-gradient criticality, eliminating rounding artifacts entirely.

The results are striking. On MNIST, CQP achieves 95.6% accuracy at a whopping 90% sparsity—2.2 percentage points higher than standard magnitude pruning. The method also uncovers a “zombie-weight” failure mode where Adam’s first-moment tensors resurrect already-pruned synapses, breaking the binary sparsity guarantee. CQP’s iterative schedule—prune, fine-tune with gradient masking, recompute criticality, repeat—solves that. A KL-divergence temporal analysis reveals a redundant simulation timestep, enabling a free 10% theoretical energy reduction. Combined weight sparsification and temporal truncation deliver a compound 73% reduction in per-inference energy at 70% sparsity, confirming the pipeline’s practical value for real neuromorphic systems.