Perforated Neural Networks beat CNNs with 60% fewer parameters on Edge Impulse

A team led by Vishy Gopal and Rorry Brenner won the Best Model award at the Edge Impulse 2025 Hackathon by introducing Perforated Backpropagation to keyword spotting. Their approach adds artificial Dendrite Nodes to standard convolutional neural networks trained on Edge Impulse's keyword spotting pipeline. Across 800 hyperparameter trials, the dendritic models outperformed traditional architectures at every parameter count and accuracy threshold. The best model hit 93.3% test accuracy with just 1,500 parameters, while the baseline required approximately 4,000 parameters to reach 92.1% accuracy.

This technique addresses a core challenge in edge machine learning: strict memory budgets and compute limits that usually force trade-offs between accuracy and model size. By simultaneously improving both, Perforated Neural Networks offer edge AI engineers a powerful new tool for deploying high-quality models on constrained devices like microcontrollers. The results, published in arXiv:2605.15647, demonstrate that dendrite-inspired architectures can deliver a 60% reduction in parameters while boosting accuracy—a rare win-win in embedded ML.