DADP: Dynamic Activity-Dependent Pruning, A Reverse Hebbian-Inspired Structural Pruning Method
DADP proposes a biologically inspired method for dynamically pruning neural networks by adjusting sparsity based on activity and error sensitivity. It uses a single global threshold instead of fixed layer budgets, creating a self-regulating process that prunes connections when features stabilize, maintaining high accuracy across various architectures while preserving representation quality.