By name: attribution-neuron-utility description: "Gradient-based eXplained Deviation (GXD) neuron utility metric for plasticity restoration in deep networks. Combines attribution-guided neuron importance scoring with critical bottleneck period (CBP) detection and targeted reset. Use when: catastrophic forgetting, plasticity loss, neuron-level analysis, continual learning reset methods, attribution-guided training."
Attribution-Based Neuron Utility for Plasticity Restoration
Core Method
GXD (Gradient-based eXplained Deviation): Attribution-based neuron utility metric that identifies which neurons contribute to prediction vs. which are dead.
CBP Detection
Critical Bottleneck Periods (CBPs) are phases where gradient norms systematically decrease, indicating plasticity loss.
Reset Strategy
- Compute GXD scores per neuron
- Identify low-utility (dead) neurons during CBPs
- Selective reset of dead neurons while preserving functional ones
- Outperforms L2 regularization, dropout, and L3 methods
Paper
- Elisii et al., arXiv:2605.06834, 2026