Iterative pruning with criticality constraint for efficient SNNs

Computational efficiency has become a critical factor for deploying artificial intelligence models on resource-constrained devices. In this context, spiking neural networks (SNNs) offer a promising paradigm by emulating the biological behavior of the brain, but their implementation on neuromorphic hardware demands aggressive synaptic pruning strategies that preserve the temporal integrity of processing. Conventional methods often ignore neuronal criticality or resort to convex relaxations that, when binarizing fractional masks, degrade accuracy at medium and high sparsity levels. Faced with this challenge, an innovative approach emerges: iterative pruning with criticality constraint, which combines weight magnitude with an importance metric based on surrogate gradients, eliminating the rounding artifacts typical of convex solvers. This technique, implemented natively in PyTorch, maintains the integrity of temporal representation even when most connections are removed. A relevant finding is the identification of a failure mode called 'zombie weights', where the first-order moments of the Adam optimizer resurrect previously pruned synapses, violating the binary sparsity guarantee. The proposed solution includes an iterative scheme that alternates pruning, fine-tuning with gradient masking, and criticality recalculation, eliminating gradient stagnation at high sparsities. Furthermore, a temporal analysis based on KL divergence reveals the existence of redundant simulation steps, allowing a theoretical 10% reduction in energy consumption without modifying the weights. Experimental results on the MNIST dataset show 95.6% accuracy at 90% sparsity, surpassing magnitude pruning by 2.2 percentage points. A criticality cliff is also observed: when exceeding a critical threshold (tau=0.9), accuracy drops from 87.0% to 14.4%, demonstrating a quantitative analog of the Critical Brain Hypothesis. The combination of weight pruning and temporal truncation achieves a composite 73% reduction in energy per inference at 70% sparsity, validating its practical application in neuromorphic deployments. At Q2BSTUDIO, as a software and technology development company, we understand that optimizing artificial intelligence models is only part of the equation. That is why we offer AI solutions for businesses that integrate these advanced techniques into custom applications, ensuring optimal performance in real environments. Our AWS and Azure cloud services allow these models to scale efficiently, while our capabilities in cybersecurity protect data and infrastructure. Additionally, we combine the power of Power BI and business intelligence services to extract value from results, and we design custom AI agents that automate complex processes. Pruning with criticality constraint is an advancement that, together with a robust technological ecosystem, paves the way toward intelligent, efficient, and secure devices.